<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Shashank Tiwari on Medium]]></title>
        <description><![CDATA[Stories by Shashank Tiwari on Medium]]></description>
        <link>https://medium.com/@stuno?source=rss-2275023c87fd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*2225ih3dvM4iVOkU3YZ-WQ.png</url>
            <title>Stories by Shashank Tiwari on Medium</title>
            <link>https://medium.com/@stuno?source=rss-2275023c87fd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 07:45:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@stuno/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[HIPAA Just Got Faster, Cheaper and Lighter: The 2026 Privacy Rule, Explained]]></title>
            <link>https://blog.uno.ai/hipaa-just-got-faster-cheaper-and-lighter-the-2026-privacy-rule-explained-634c39cdd55f?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/634c39cdd55f</guid>
            <category><![CDATA[hipaa-compliance]]></category>
            <category><![CDATA[hipaa]]></category>
            <category><![CDATA[hipaa-privacy-rule]]></category>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 03:01:57 GMT</pubDate>
            <atom:updated>2026-11-17T03:01:57.069Z</atom:updated>
            <content:encoded><![CDATA[<p>Patients have long had a right to their medical records. Getting them can mean a month of waiting, a notarized form and a mystery fee. On November 19, 2026, HHS fixes much of that.</p><p>The new HIPAA Privacy Rule changes take nearly six years of debate and boil them down to one idea: <strong>health information should move faster, with less friction, to the people who need it.</strong> Covered entities have until <strong>July 17, 2027</strong> to get there.</p><h3>The short version</h3><ol><li><strong>Records in 15 days, not 30.</strong> Patients get their records in half the time.</li><li><strong>Online access is free.</strong> No charge to view or download records through a portal or a health app.</li><li><strong>No more hoops.</strong> Notaries, wet signatures and “come in person with ID” are out when an easier option works.</li><li><strong>The signature form is gone.</strong> Patients no longer sign to say they got the privacy notice. That alone saves providers about $800 million a year.</li><li><strong>Sharing for care is clearly allowed.</strong> Providers and plans can share with social services and community groups to coordinate care, no extra permission needed.</li></ol><h3>Every change at a glance</h3><ul><li><strong>Time to provide records.</strong> Before: 30 days. Now: 15 calendar days, “as soon as practicable.” Shorter state deadlines win.</li><li><strong>Extension.</strong> One 30-day extension, unchanged.</li><li><strong>How patients can ask.</strong> Before: the covered entity could insist on its own form or channel. Now: any written channel, including email, portal, letter or health app. No forcing a harder route.</li><li><strong>Identity checks.</strong> Before: the covered entity’s call. Now: no unreasonable checks. Notarization, wet signatures and in-person-only ID are called out.</li><li><strong>Health apps.</strong> Before: no definition. Now: “personal health application” is defined. If your API already serves other organizations at no added security risk, it serves the patient’s app.</li><li><strong>Summaries.</strong> Before: could be offered instead of a copy. Now: patients must be told they can still get the full copy.</li><li><strong>Fees for online access.</strong> Before: a labor fee could be charged. Now: free, as is in-person inspection.</li><li><strong>Fees for copies.</strong> Before: three methods, including a $6.50 flat fee. Now: labor, supplies, postage and an agreed summary only.</li><li><strong>Fee transparency.</strong> Before: none required. Now: a fee schedule posted online and at the front desk.</li><li><strong>Sending records to someone else.</strong> Before: unclear after a 2020 court ruling. Now: electronic copies of EHR data only. Paper and non-EHR records need an authorization.</li><li><strong>Care coordination.</strong> Before: permitted, but providers hesitated. Now: expressly covers individual and population-level work, and sharing with social services, community groups and home and community-based service providers.</li><li><strong>Privacy notice acknowledgment.</strong> Before: a signed form, kept 6 years. Now: eliminated.</li><li><strong>Privacy notice content.</strong> Before: old header and a named contact. Now: a new header, a right to discuss the notice, and a contact with phone and email. Inmates now receive it.</li><li><strong>Business associates.</strong> Before: could answer requests if the contract allowed. Now: the same, written into the rule. Contracts should say who answers.</li><li><strong>Relay services for deaf patients.</strong> Before: not in the rule; HHS guidance said no BAA was needed. Now: expressly permitted in the rule, no BAA, and it covers deaf and hard of hearing staff too.</li><li><strong>Armed Forces permission.</strong> Before: military only. Now: all Uniformed Services, including the USPHS and NOAA Corps.</li></ul><h3>Mark your calendar</h3><ul><li><strong>November 19, 2026:</strong> published in the Federal Register</li><li><strong>January 18, 2027:</strong> rule takes effect</li><li><strong>July 17, 2027:</strong> compliance deadline</li><li><strong>January 18, 2028:</strong> extra time for existing BAAs ends</li></ul><p>You have 240 days from publication to comply. Only BAA language gets more time, and only for agreements signed before publication that are not renewed or modified in the interim.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4YkHB5O2VaHl6dKh9s8rVw@2x.png" /></figure><h3>The big one: records in 15 days</h3><p>Picture a patient switching specialists. Today she can wait a full month for her records, plus another month if the office takes an extension. From July 2027, the clock runs out at 15 calendar days. The rule’s standard is “as soon as practicable,” so 15 is the ceiling, not the target.</p><p>Three details matter in practice:</p><ul><li><strong>Shorter state laws win.</strong> If a state requires records in 10 days, 10 days is now also the HIPAA deadline.</li><li><strong>Clarifying the request doesn’t pause the clock.</strong> Back-and-forth about what the patient wants still counts against the 15 days.</li><li><strong>It applies to vendors too.</strong> A business associate handling requests must hit 15 days from the compliance date, even if its contract hasn’t been updated.</li></ul><p><strong>Asking gets easier.</strong> Covered entities can still require written requests, but “written” now means any channel that documents the ask: a letter, an email, a portal message, or a request sent from a patient’s health app, including through TEFCA or CMS Aligned Networks. Forcing a patient onto paper, the portal or a trip to the office is prohibited when an easier route is practicable.</p><p><strong>Proving who you are gets easier.</strong> Verification can’t make patients jump through hoops when a lighter option works. OCR names three examples of going too far:</p><ul><li>Demanding ID in person when remote verification is practicable</li><li>Demanding a wet signature when an e-signature is valid</li><li>Demanding a notarized request</li></ul><p>OCR goes further for apps: if a TEFCA request carries a signed identity token from a provider certified at NIST IAL2 and AAL2, you validate it, and the demographics match your records, refusing it is unreasonable.</p><p><strong>Apps get a seat at the table.</strong> The rule defines a “personal health application” as an app the patient controls, not one run by the provider or plan. The key line: if another law requires a technology like a certified API, or you already send data to other organizations through a secure API and the patient’s app adds no extra security risk, that format is “readily producible” and the patient’s app gets the same treatment. Nobody has to build new technology, but nobody can withhold what they already have.</p><p><strong>Summaries come with a reminder.</strong> Offer a summary instead of a full copy, and you must tell the patient they can still get the full copy. The patient decides.</p><h3>Follow the money: fees</h3><p>The simplest way to think about the new fee rules: <strong>if a machine does the work, it’s free. If a person does the work, you can charge for the person’s time.</strong></p><ul><li><strong>Free:</strong> viewing records in person, and getting them through a portal or a connected health app.</li><li><strong>Chargeable:</strong> staff labor to make copies, supplies for paper or film, postage, and a summary the patient agreed to in advance.</li><li><strong>Not “internet-based”:</strong> a PDF that a staff member emails. That can carry a labor fee.</li></ul><p>The 2016 guidance on calculating fees, including the $6.50 flat-fee option, is superseded. And no one can withhold records because a medical bill is unpaid.</p><p><strong>Prices go public.</strong> Every covered entity with a website must post its fee schedule there, and make it available wherever patients ask for records, call centers included. It must list what’s free and the standard charges for copies, for sending records to another person, and for authorized disclosures. HHS dropped the proposed individualized estimates and itemized bills.</p><h3>Sending records to someone else</h3><p>Patients can tell a provider to send records straight to a person they choose, like a lawyer, a new doctor or a family member, with no HIPAA authorization form. That right now has firm edges, set by the 2020 court ruling in <em>Ciox Health v. Azar</em>:</p><ul><li><strong>In scope:</strong> electronic copies of what’s in the electronic health record.</li><li><strong>Out of scope:</strong> paper copies, and records outside the EHR. Those need a signed authorization.</li></ul><p>The new “electronic health record” definition is broad. It includes lab results, imaging, pathology reports and billing records, from both treating providers and behind-the-scenes ones like labs. Requests must be clear, conspicuous and specific. Providers don’t have to act on verbal requests to send records elsewhere.</p><h3>Sharing for care coordination</h3><p>A patient leaving a nursing home needs housing, rides to appointments and meals. The agencies that arrange those things need some of her health information. HIPAA always allowed this. Many providers weren’t sure, so they didn’t share.</p><p>The rule ends the guesswork in two moves:</p><ul><li><strong>Health care operations now plainly covers all care coordination and case management</strong>, whether for one patient or a whole population. Health plans and value-based care programs benefit most.</li><li><strong>A new provision names the partners.</strong> Covered entities may share with social services agencies, community-based organizations, home and community-based service providers and similar groups, without patient authorization, for one patient’s care coordination.</li></ul><p>The guardrails stay: share only the minimum necessary, and patients can still ask for restrictions.</p><h3>Goodbye, acknowledgment form</h3><p>That clipboard form saying “I received the Notice of Privacy Practices”? Retired. Providers no longer have to collect it, document why they couldn’t, or store it for six years. HHS says this is where nearly all the rule’s savings come from.</p><p>The notice itself gets better, and every covered entity must rewrite theirs:</p><ul><li><strong>A clearer header.</strong> It must say up front that the notice explains how to get your records, how to file a complaint, and your rights. “Medical” becomes “Health.”</li><li><strong>A right to talk it through.</strong> Patients can discuss the notice with a designated person or office, reachable by phone and email. AI can field the first questions, as long as a knowledgeable person is available.</li><li><strong>Access explained.</strong> The notice must say how to get records cheaply or free, and how to send them to someone else.</li><li><strong>Inmates included.</strong> The old exception is gone. Correctional facilities that are covered entities, and providers serving them, must hand out the notice.</li></ul><p>One thing stays: keep a copy of each version of the notice you issue.</p><h3>The fine print</h3><ul><li><strong>Business associates.</strong> A vendor sends records directly to a patient only if its BAA says so. Otherwise it hands them back to the covered entity, which stays responsible either way. Spell it out in the contract.</li><li><strong>Extra time for BAAs.</strong> A compliant BAA signed before November 19, 2026, and not renewed or changed between January 18 and July 17, 2027, counts as compliant until it’s renewed or changed after July 17, 2027, or until January 18, 2028, whichever comes first. Auto-renewing contracts qualify. Handshake deals don’t. The extra time covers contract wording only, not the 15-day deadline.</li><li><strong>Relay services.</strong> Providers can share information with relay operators who help deaf, hard of hearing, deaf-blind or speech-disabled people communicate. No BAA needed.</li><li><strong>Uniformed Services.</strong> Rules that let providers share service members’ health information with their command now cover the USPHS and NOAA Commissioned Corps too.</li><li><strong>Housekeeping.</strong> A typo, a citation and some punctuation fixed. Nothing substantive.</li></ul><h3>What didn’t make the cut</h3><p>If you planned around the 2021 proposal, cross these off. HHS dropped them:</p><ul><li>Letting patients photograph records during appointments</li><li>Requiring a policy to prioritize urgent requests</li><li>Shortening the extension to 15 days (it stays at 30)</li><li>Requiring providers to act on verbal requests to send records elsewhere</li><li>Requiring providers and plans to request records from other providers on a patient’s behalf (still allowed voluntarily)</li><li>Personalized fee estimates and itemized bills</li><li>Fee caps on sending records to someone else (left for later)</li><li>Exempting care coordination from the minimum necessary rule</li><li>New “good faith” standards for sharing in substance use, serious mental illness and emergency situations</li><li>Requiring the privacy contact to be onsite</li></ul><h3>By the numbers</h3><ul><li><strong>$1.17 billion</strong> in compliance costs over five years, mostly in the first year ($784 million), for administrative work, new policies, notices, fee schedules and training</li><li><strong>$3.99 billion</strong> in savings over five years, almost all from dropping the acknowledgment form</li><li><strong>$2.83 billion</strong> net savings over five years</li><li><strong>847,506</strong> covered entity locations affected</li></ul><p><em>HHS figures, 2025 dollars, undiscounted.</em></p><h3>Your to-do list</h3><p>OCR turned down requests for one to two years and gave the legal minimum. Eight months goes fast.</p><ol><li>Reset your records workflow to a 15-day clock that starts on receipt</li><li>Check state laws for shorter deadlines; those become yours</li><li>Accept requests through every written channel you can practicably support: email, portal, letter, health app</li><li>Drop notarization, wet signatures and in-person-only ID where an easier option works</li><li>List the electronic formats you already send to other organizations; patients’ apps get them too</li><li>Stop charging for portal and app access; reprice copies to labor, supplies and postage</li><li>Post your fee schedule online and at every front desk and call center</li><li>Route paper and non-EHR “send to someone else” requests to an authorization</li><li>Script the “you can still get the full copy” line for summaries</li><li>Rewrite your privacy notice and name a contact with phone and email</li><li>Retire the acknowledgment form and its six-year storage</li><li>Review your BAAs and update them before January 18, 2028</li><li>Retrain staff and update policies</li></ol><h3>Live in Uno.ai on day one</h3><p>The rule goes live in Uno.ai on November 19, 2026, the day it publishes.</p><ul><li><strong>Controls for every change</strong>, each tied to its exact citation, from the 15-day deadline to the fee schedule and the BAA transition, with guidance and the evidence auditors will ask for.</li><li><strong>A readiness assessment</strong> that shows where you stand against July 17, 2027. Every answer maps to a control, so each gap becomes a tracked task with an owner and a due date.</li><li><strong>Built on your existing HIPAA content</strong>, so you only work on what changed.</li></ul><p>Want to see it in your tenant? Email <a href="http://contact@uno.ai/">contact@uno.ai</a> .</p><h3>Source</h3><p><a href="https://www.federalregister.gov/d/2026-23429">HIPAA Privacy Rule: Changes to Support Coordinated Care and Individual Engagement and Reduce Regulatory Burdens</a>, HHS Office for Civil Rights, Federal Register document 2026–23429 (<a href="https://public-inspection.federalregister.gov/2026-23429.pdf">public inspection PDF</a>). Effective, compliance and BAA dates are counted from the scheduled November 19, 2026 publication; confirm against the official edition.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=634c39cdd55f" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/hipaa-just-got-faster-cheaper-and-lighter-the-2026-privacy-rule-explained-634c39cdd55f">HIPAA Just Got Faster, Cheaper and Lighter: The 2026 Privacy Rule, Explained</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Line at the Bottom of the Check]]></title>
            <link>https://blog.uno.ai/the-line-at-the-bottom-of-the-check-511d448fc194?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/511d448fc194</guid>
            <category><![CDATA[grc]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[audit]]></category>
            <category><![CDATA[banking]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 00:15:51 GMT</pubDate>
            <atom:updated>2026-10-08T00:15:51.520Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Version in Force, Part 3 of 3: Agentic workflows in banking, and why Uno Writ is what lets them scale</em></p><p>In the 1950s, American banks were drowning in paper. Check volumes were climbing so fast that the industry worried it would run out of clerks to sort them. Bank of America, working with the Stanford Research Institute, built a machine to do the job. It was called ERMA, and it went into service in 1959.</p><p>The machine got the headlines. The more important invention was a line of odd-looking numbers printed in magnetic ink along the bottom of every check. That line told the machine which bank, which account and which item it was holding. ERMA could sort checks far faster than any clerk because each check carried its own identity. Automation scaled not because the machine was clever, but because every piece of work arrived with its own record attached.</p><p>Banks are at the same point with AI agents. The agents are clever enough. What’s missing is the line at the bottom of the check.</p><h3>Where agents are already working</h3><p>In our experience, agentic workflows in banking have moved past the innovation team’s slide. They are drafting, triaging and assembling evidence inside the first and second lines, because that is where the volume and the cost sit. Every one of these workflows is governed by written procedure, and every one draws the examiner’s first question.</p><p>Workflow What the agent does Rules that govern it What the examiner asks KYC / CDD refresh Gathers documents, updates beneficial ownership, drafts the refresh memo CIP, CDD and beneficial ownership procedures; risk-rating methodology Which CDD procedure version produced this rating? AML alert triage Reviews alerts, assembles context, recommends close or escalate Alert disposition standards, typology guidance, escalation thresholds Was this alert closed under the threshold in force that day? SAR narrative drafting Drafts the narrative from case facts SAR content standards, confidentiality rules, filing timelines Which narrative standard governed this draft, and who owned it? Credit memo preparation Spreads financials, drafts the memo, flags policy exceptions Credit policy, concentration limits, exception approval rules Was the exception rule applied, and which version? Third-party due diligence Reviews questionnaires and SOC reports, scores the vendor Third-party risk policy, criticality tiers, contract requirements Which tiering rule set this vendor’s due diligence depth? Regulatory change management Maps new rules to policies and controls, drafts impact assessments Reg change procedure, ownership matrix Who approved the mapping logic the agent used? Complaints handling Classifies complaints, flags UDAAP and fair lending risk Complaint taxonomy, escalation rules Was this complaint classified under the current taxonomy?</p><p>In every row the bank can show the policy. In every row, without a per-run record, it cannot show that the policy is what the agent ran under.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I6Z_jdmszRM2VzeVUiHmqQ.png" /></figure><h3>The review tax</h3><p>So banks pay a tax. A human re-reads the agent’s output against the policy, line by line, before it counts. The agent drafts the CDD memo in minutes. The analyst then spends much of the time reviewing it that they would have spent writing it.</p><p>The tax is rational. Supervisors already expect banks to govern their models and the vendors who supply them. The Federal Reserve’s SR 11–7 and the OCC’s matching Bulletin 2011–12 set the model risk management standard. The 2023 interagency guidance on third-party relationships covers agents a bank buys rather than builds. Bank Secrecy Act recordkeeping covers the AML work itself. Banks operating in Europe add DORA and the EU AI Act, which treats credit scoring of individuals as high-risk. None of these frameworks accepts “the agent is usually right” as evidence.</p><p>The tax is also self-defeating. A bank that reviews every agent output by hand has not automated the work. It has added a step.</p><h3>What changes with Uno Writ</h3><p>Uno Writ gives every agent seat its own line at the bottom of the check. Before the agent starts, the compiled rules, their versions, their owners and the fingerprint of the instruction set post to the work item’s permanent record. When the seat closes, every rule is attested, every mechanical check runs, and every rule that should have applied but didn’t is recorded as an exception.</p><p>That changes how the three lines of defense work.</p><ul><li><strong>First line</strong> owns the rules as controlled records, with effective dates and triggers, instead of editing a shared prompt file.</li><li><strong>Second line</strong> approves rule changes once and samples a real population of seats, exceptions included, instead of re-reading every output.</li><li><strong>Internal audit</strong> tests operating effectiveness the way it already tests human controls: pull a sample, check each seat’s fingerprint against the approved rule versions, and trace each exception to its resolution.</li></ul><p>Review moves from every item to a sample, and from opinion to evidence. That is where the productivity comes back.</p><h3>Where to start</h3><p>Start by finding out what is already running. Dendrite and AI Nerve Center show which agents are live across the estate, registered or not, and what rules, tools and data sources each is configured to use. Then pick one workflow from the table, the one with the highest volume and the heaviest review. Move its procedure out of the prompt file and into controlled rules with owners, effective dates and triggers. Run it through Writ for one examination cycle. Then hand internal audit the seat records and let them sample. When the first sample passes without a re-read of every output, the business case for the next workflow writes itself.</p><h3>Back to memo 212</h3><p>The examiner picks memo 212 and asks the first question. This time the bank answers it in one screen. The memo was produced on the 14th under three rules. Here are their ids, versions and owners. Here is the hash of the instruction set, posted the moment the seat opened, before the agent did anything. One rule changed that morning, and here is the record showing it did not apply to this segment. Every rule was attested at close. Two were checked mechanically. There were no exceptions.</p><p>The conversation moves on.</p><p>ERMA didn’t make checks trustworthy by being smart. It made them processable at scale because every check carried its own record. Agents in banking will scale the same way. Not when they get smarter, but when every piece of their work carries a receipt an examiner will accept.</p><p>That receipt is Uno Writ.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=511d448fc194" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/the-line-at-the-bottom-of-the-check-511d448fc194">The Line at the Bottom of the Check</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing Uno Writ]]></title>
            <link>https://blog.uno.ai/introducing-uno-writ-92a9764752f1?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/92a9764752f1</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[audit]]></category>
            <category><![CDATA[grc]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[governance]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 21:08:14 GMT</pubDate>
            <atom:updated>2026-10-06T21:08:14.800Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Version in Force, Part 2 of 3: A receipt for every piece of AI work, so examiners can look where no one has looked before</em></p><p>In medieval England, the king did not govern by conversation. He governed by writ: a short written order, in a standard form, under seal, addressed to a named person for a named purpose. The clerks of Chancery kept the approved forms in a register, and the courts would not act on a writ that departed from them. The genius of the system was not the wording. It was that anyone, years later, could hold up a writ and say exactly what authority it carried, whose it was, and when it was issued.</p><p>AI agents need the same thing. Today we’re launching it.</p><p><strong>Uno Writ</strong> compiles, fingerprints and attests the rules every AI agent runs under, one assignment at a time. It gives every piece of agent work a receipt: which rules governed it, which version of each, who owned each one, whether any rule should have applied and didn’t, and what happened to each rule during the work.</p><h3>Built with MyStatus</h3><p>Uno Writ came out of a joint research and product development program between Uno.ai and MyStatus. The two halves fit. MyStatus records who did each piece of work, person or agent. Uno.ai governs the controls, policies and regulations that work must follow. Together, our teams studied how agents were actually being governed in the regulated workflows where they are arriving first: customer due diligence, alert triage, third-party due diligence and regulatory change, with people and agents doing the same work side by side. We asked one question of every workflow: what would an examiner need to see to sign off on this?</p><p>The headline finding was uncomfortable and consistent. The written policies were good. The agents were often good. What failed was the link between them. Instruction files changed several times a day. No run recorded which wording it loaded. Models summarized rules into new rules on the way in. And nothing said which rules applied to which task. Every one of those failures was a document-control failure, not an AI failure.</p><p>“MyStatus already tracks who did the work, so it made sense that we also track and version the rules that the humans and AI agents used to do a specific piece of work,” said Tom Ortega, founder of MyStatus.</p><p>That reframing shaped the product. We stopped asking how to make agents more trustworthy and started asking how to make their governance evidenceable.</p><h3>How Uno Writ works</h3><p>Writ operates on a <strong>seat</strong>: one AI working session doing one assigned piece of work. Four mechanisms wrap every seat.</p><ol><li><strong>A controlled rule library.</strong> Every rule is its own record, with an id, the statement, an effective date, the person who approved it and a sealed edit history. Edits are attributed. Prior wordings are kept. Nothing is deleted. Each rule carries machine-readable triggers that state when it applies: the paths, actions or data it concerns.</li><li><strong>Compilation by program, not by model.</strong> When work is assigned, Writ selects the rules whose triggers match the assignment and writes them verbatim into the instruction set the seat runs under. No model rewords a rule on the way in. The assignment records when it was approved. Before the seat starts, any rule added or changed since then is checked against the assignment and either compiled in or recorded as not applicable.</li><li><strong>A fingerprint on the permanent record.</strong> Writ hashes the compiled instruction set. The hash, the rule ids, and each rule’s owner and version post to the work item’s permanent, timestamped record the moment the seat opens. That row is the evidence of policy in force. It is written before the work starts, so it cannot be fitted to the outcome afterwards.</li><li><strong>Coverage and attestation at close.</strong> When the seat closes, Writ compares what the work actually touched against every rule’s triggers. Any rule that matched but was not compiled is flagged as an exception. For each compiled rule the seat records one row: applied, checked, or not applicable and why. Where a rule can be checked mechanically, Writ checks it, because a model saying “checked” is a claim, not evidence.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EzZHRMLdzf39XDOxhWSvjw.png" /></figure><h3>What the receipt looks like</h3><p>Run memo 212 from Part 1 through Writ. This is the record the examiner sees.</p><pre>{<br>  &quot;receipt&quot;: &quot;uno.writ/seat-receipt/v1&quot;,<br>  &quot;seat_id&quot;: &quot;seat-2026-10-14-0212&quot;,<br>  &quot;work_item&quot;: &quot;CDD-REFRESH-88417&quot;,<br>  &quot;description&quot;: &quot;Periodic customer due-diligence refresh for a commercial banking client (memo 212). The agent gathered ownership documents, updated beneficial ownership, re-ran the customer risk rating and drafted the refresh memo under the three rules compiled below.&quot;,<br>  &quot;assignment_approved_at&quot;: &quot;2026-10-14T08:51:10Z&quot;,<br>  &quot;seat_opened_at&quot;: &quot;2026-10-14T09:02:47Z&quot;,<br>  &quot;seat_closed_at&quot;: &quot;2026-10-14T09:38:12Z&quot;,<br>  &quot;instruction_hash&quot;: &quot;sha256:4f1c...9a07&quot;,<br>  &quot;rules_compiled&quot;: [<br>    {<br>      &quot;id&quot;: &quot;KYC-014&quot;,<br>      &quot;version&quot;: 6,<br>      &quot;title&quot;: &quot;Beneficial ownership identification and verification&quot;,<br>      &quot;owner&quot;: &quot;Head of Financial Crimes Compliance&quot;,<br>      &quot;effective&quot;: &quot;2026-09-30&quot;<br>    },<br>    {<br>      &quot;id&quot;: &quot;KYC-022&quot;,<br>      &quot;version&quot;: 3,<br>      &quot;title&quot;: &quot;Customer risk rating methodology&quot;,<br>      &quot;owner&quot;: &quot;BSA Officer&quot;,<br>      &quot;effective&quot;: &quot;2026-07-01&quot;<br>    },<br>    {<br>      &quot;id&quot;: &quot;DATA-007&quot;,<br>      &quot;version&quot;: 2,<br>      &quot;title&quot;: &quot;Handling of personal data in work products&quot;,<br>      &quot;owner&quot;: &quot;Chief Privacy Officer&quot;,<br>      &quot;effective&quot;: &quot;2026-05-15&quot;<br>    }<br>  ],<br>  &quot;late_changes&quot;: [<br>    {<br>      &quot;id&quot;: &quot;KYC-031&quot;,<br>      &quot;version&quot;: 2,<br>      &quot;title&quot;: &quot;Retail customer address refresh&quot;,<br>      &quot;changed_at&quot;: &quot;2026-10-14T08:55:00Z&quot;,<br>      &quot;disposition&quot;: &quot;not_applicable&quot;,<br>      &quot;reason&quot;: &quot;Rule applies to the retail segment only; this work item is commercial.&quot;<br>    }<br>  ],<br>  &quot;close&quot;: {<br>    &quot;coverage&quot;: &quot;3 of 3 matched rules compiled&quot;,<br>    &quot;exceptions&quot;: [],<br>    &quot;attestations&quot;: [<br>      {<br>        &quot;id&quot;: &quot;KYC-014&quot;,<br>        &quot;status&quot;: &quot;checked&quot;,<br>        &quot;method&quot;: &quot;mechanical&quot;,<br>        &quot;evidence&quot;: &quot;Every owner at or above 25% has a verified identity record (4 of 4).&quot;<br>      },<br>      {<br>        &quot;id&quot;: &quot;KYC-022&quot;,<br>        &quot;status&quot;: &quot;applied&quot;,<br>        &quot;method&quot;: &quot;seat attestation&quot;,<br>        &quot;evidence&quot;: &quot;Rating recalculated as Medium; factor scores recorded on the work item.&quot;<br>      },<br>      {<br>        &quot;id&quot;: &quot;DATA-007&quot;,<br>        &quot;status&quot;: &quot;checked&quot;,<br>        &quot;method&quot;: &quot;mechanical&quot;,<br>        &quot;evidence&quot;: &quot;No full government ID or account numbers in memo text (0 matches).&quot;<br>      }<br>    ]<br>  },<br>  &quot;notes&quot;: [<br>    &quot;Receipt written when the seat opened, before the agent&#39;s first action. The fingerprint cannot be changed after that point.&quot;,<br>    &quot;All three rules were compiled verbatim from the controlled library. No model reworded any rule on the way in.&quot;,<br>    &quot;KYC-031 changed after the assignment was approved and before the seat opened. It was reviewed against the assignment and recorded as not applicable.&quot;,<br>    &quot;KYC-014 and DATA-007 were verified by program. KYC-022 rests on the seat&#39;s attestation and is eligible for second-line sampling.&quot;,<br>    &quot;The coverage check at close found no rule whose triggers matched the work but which was not compiled.&quot;<br>  ]<br>}</pre><p>Memo 212 now has an answer. The examiner sees what the work was, the exact rules with their versions, owners and effective dates, the fingerprint of the instruction set, the disposition of a rule that changed that morning, the evidence behind every check, and proof that the record was written before the work began. The notes answer, in plain English, the follow-up questions an examiner would otherwise have to ask.</p><h3>Why we built it this way</h3><p>Three design decisions do most of the work.</p><p><strong>Programs compile, models don’t.</strong> The moment a model touches a rule on the way into a session, Bartlett’s experiment starts again. Writ keeps the path from approved rule to running instruction entirely deterministic.</p><p><strong>The record comes first.</strong> Evidence written after the outcome is known is a narrative. Evidence written when the seat opens is a control. Writ posts the fingerprint before the agent does anything.</p><p><strong>Exceptions are first-class.</strong> A control that only records successes cannot be tested. Writ records the rules that should have applied and didn’t, so second- and third-line teams have a real population to sample, failures included.</p><h3>Seeing the agents nobody registered</h3><p>Writ governs every agent that runs through it. Most firms also run agents that don’t. A developer’s coding assistant with its own rules file. An analyst agent a team built over a weekend. An off-the-shelf tool whose configuration nobody in risk has read.</p><p>This is where Dendrite and AI Nerve Center come in. Dendrite, Uno’s endpoint discovery engine, finds the agents running across the estate, managed and unmanaged, and reads what each one runs under: its system prompt and rules files such as CLAUDE.md and AGENTS.md, its MCP configuration, and the tools and data sources it is configured to use. AI Nerve Center maps all of it into Uno’s knowledge graph, which keeps every prior state instead of overwriting it.</p><p>That gives Uno a vantage point no one else has. We see more than the rules an agent runs under today. We see how those rules, configurations and context have changed over time, which agents inherited each change, and what each change touched: which controls, which workflows, which regulations. When one team rewrites a beneficial ownership rule in its prompt file, Uno shows which agents picked up the new wording, which kept the old one, and when.</p><p>Writ makes governed agents provable. Dendrite and AI Nerve Center make the ungoverned ones visible, and give the firm a path to bring them under Writ.</p><h3>Where it fits in Uno.ai</h3><p>Uno.ai was built agent-first. Our platform runs agentic workflows across GRC, third-party risk, enterprise risk, business continuity and AI governance, and we hold our own agents to the standard we sell. Writ’s rule library sits alongside the controls, policies and authority-document citations Uno already manages, so every rule an agent runs under traces back to the regulation that required it.</p><p>Writ governs the agents our customers build and run, on any platform, at any scale. One pilot agent or thousands across every business line, each seat gets the same receipt.</p><p>The writ never made the king wiser. It made his authority provable. That is the standard AI agents now have to meet, and Uno Writ is how they meet it.</p><p><em>Next: Part 3, The Line at the Bottom of the Check.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=92a9764752f1" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/introducing-uno-writ-92a9764752f1">Introducing Uno Writ</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Drawing That Got Built]]></title>
            <link>https://blog.uno.ai/the-drawing-that-got-built-9b283a4e993c?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b283a4e993c</guid>
            <category><![CDATA[audit]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[grc]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 04:55:11 GMT</pubDate>
            <atom:updated>2026-10-06T21:10:07.288Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Version in Force, Part 1 of 3: Why the most important document in your AI program is the one nobody can produce</em></p><p>On the evening of July 17, 1981, about 1,600 people came to the atrium of the new Hyatt Regency hotel in Kansas City for a tea dance. Some of them stood on the walkways that crossed the atrium, hung from the ceiling on steel rods. A little after seven, the fourth-floor walkway and the second-floor walkway beneath it fell into the crowd. One hundred and fourteen people died. Until September 11, it was the deadliest structural collapse in American history.</p><p>Here is the strange part. The walkways that fell were not the walkways the engineers had drawn.</p><p>The original drawing called for long, continuous rods running from the ceiling, through the fourth-floor walkway, down to the second. Long threaded rods are awkward to fabricate and awkward to install. So during construction the steel fabricator proposed something simpler: one set of rods from the ceiling to the fourth floor, and a second set from the fourth floor down to the second. On paper it looked like a minor detail. In physics it was not. The fourth-floor connection now carried the weight of both walkways, roughly double its load, on a connection that was already marginal.</p><p>After the collapse, everyone asked the same question: who approved the change? The fabricator said it had asked and been told yes. The engineering firm’s stamp was on the shop drawings. The engineers said they had never been asked to check that connection. Each account was plausible. The investigation, the hearings and the license revocations that followed all circled one question nobody could cleanly answer. Which version of the design was in force, and who said so?</p><p>We usually tell stories like this as stories about skill. A careless engineer, a corner cut. But look at what the profession changed afterwards. It did not invent a better rod. It tightened the rules on who must review a design change, who signs it, and what record proves it. The Hyatt is a story about something much duller than skill. It is a story about document control.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lCxK0RJib4FvXCpTz7KWAA.png" /></figure><h3>The boring question that matters most</h3><p>Every regulated industry eventually makes the same discovery. Expertise is not enough. You also have to prove which instructions the expert was following on the day the work was done.</p><p>So industries built document control. An approved procedure. An effective date. A named owner. A change record that keeps the old wording instead of overwriting it. The quality standard ISO 9001 has a clause for it. Every bank has a policy on policies. It is the least glamorous discipline in any organization, and it is the one examiners reach for first.</p><p>Watch a bank examiner open a file. The first question is rarely about the outcome. It is: what procedure governed this work, and can you show me the version that was in force that day? If the answer is yes, the conversation moves on. If the answer is no, nothing else in the file matters very much.</p><p>Auditors split this into two ideas. <strong>Design</strong> asks whether the control is written down and sensible. <strong>Operating effectiveness</strong> asks whether the control actually ran, every time, on the work in scope. Design is a document. Operating effectiveness is evidence: a population of executions you can sample, each tied to the version of the control it ran under.</p><p>For as long as regulators have examined anything, the work being governed was done by people. Now some of it is done by AI agents. And the boring question has quietly become unanswerable.</p><h3>The story that changes as it travels</h3><p>In 1932 a Cambridge psychologist named Frederic Bartlett ran an experiment. He gave people an unfamiliar folk tale, “The War of the Ghosts,” and asked them to retell it from memory. Then he passed each retelling to the next person, like a game of telephone played slowly and carefully.</p><p>No one lied. No one was careless. Each retelling was a reasonable summary of the one before. And yet by the end of the chain the story had changed. The strange details fell away. Unfamiliar things became familiar ones. The tale got shorter, tidier and more conventional. It became the story the listeners expected to hear rather than the story they were told.</p><p>Large language models are extraordinary summarizers. That is precisely the problem.</p><h3>What governs an agent</h3><p>When an AI agent does a piece of regulated work, its procedure is its standing instructions. A system prompt. A rules file. A playbook someone pasted into the session. In principle this is the agent’s approved procedure. In practice it has none of the properties that made document control work.</p><p>It changes daily, often several times a day, and any session or any person can edit it. It isn’t versioned per run, so nobody can say which wording a given execution actually loaded. It gets paraphrased, because models summarize rules into other rules, and the text that governed the work may not be the text anyone approved. That is Bartlett’s experiment, running at machine speed. And it is one undifferentiated block. Nothing records which rules applied to this particular task, or whose rules they were.</p><p>Imagine an examiner at a mid-sized bank. Over one month, an agent drafted four hundred customer due-diligence refresh memos. The examiner picks memo number 212 and asks the first question. Which procedure governed this?</p><p>The honest answer goes something like this. The agent ran under a rules file in a repository. That file was edited thirty-one times that month, by six different people and two other agents. We believe memo 212 was produced on the 14th, but we can’t say which of that day’s three versions was loaded. One rule about beneficial ownership was rewritten mid-month, and we can’t tell you whether the agent saw the old wording, the new wording, or a model’s summary of both.</p><p>The policy exists. The design is fine. What does not exist is any evidence that the design is the thing that ran.</p><h3>Why banks say no</h3><p>This is why, in our experience, regulated firms make one of two choices. They block agents from in-scope work entirely. Or they let agents in under heavy manual review, with a human re-reading every output against the policy.</p><p>The second choice looks like progress. It is not. The manual review is exactly where the productivity goes. A firm that re-checks every agent output by hand has bought a very expensive drafting tool and kept the cost structure it wanted to escape.</p><p>Notice that neither choice is about whether the agent is good at the work. The agent may be excellent. The problem is not judgment. The problem is provenance. Nobody can produce the drawing that got built.</p><p>The engineers at the Hyatt had a drawing on file. What they could not do was prove that the drawing on file was the one in force when the steel went up. Every firm running AI agents on regulated work is in that position today. The fix is not a smarter agent. The fix is a receipt.</p><p><em>Next: Part 2, </em><a href="https://blog.uno.ai/introducing-uno-writ-92a9764752f1"><em>Introducing Uno Writ</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b283a4e993c" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/the-drawing-that-got-built-9b283a4e993c">The Drawing That Got Built</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The EU Cyber Resilience Act Starts Biting This Week.]]></title>
            <link>https://blog.uno.ai/the-eu-cyber-resilience-act-starts-biting-this-week-d35e87969b4d?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/d35e87969b4d</guid>
            <category><![CDATA[payments]]></category>
            <category><![CDATA[dora]]></category>
            <category><![CDATA[grc]]></category>
            <category><![CDATA[product-security]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 22:33:14 GMT</pubDate>
            <atom:updated>2026-09-06T22:33:14.233Z</atom:updated>
            <content:encoded><![CDATA[<h3>Most Banks Think It Isn’t About Them.</h3><p>On 11 September 2026, five days from now, the first hard obligation under the EU Cyber Resilience Act starts to apply. Manufacturers of products with digital elements will have to report actively exploited vulnerabilities and severe security incidents to their national CSIRT and ENISA, on a clock measured in hours. ENISA’s Single Reporting Platform, the only channel for those filings, goes operational the same day.[¹]</p><p>Most banks and payments firms have the CRA filed under “vendor problem.” For a meaningful number of them, that filing is wrong.</p><p>This piece covers what the regulation actually says, where a bank or payments firm can find itself on the wrong side of the manufacturer definition, and why a mature DORA program does not answer the question.</p><h3>What the CRA is</h3><p>The Cyber Resilience Act is Regulation (EU) 2024/2847. It entered into force on 10 December 2024 and becomes fully applicable on 11 December 2027. Two provisions run ahead of that date: Chapter IV on the notification of conformity assessment bodies applied from 11 June 2026, and the Article 14 reporting obligations apply from 11 September 2026.[²]</p><p>It is a horizontal product regulation, not a sectoral cybersecurity rule. That distinction is the whole story. DORA, NIS2 and the EBA guidelines regulate <em>entities</em> and their ICT estates. The CRA regulates <em>products</em> placed on the EU market, using the same machinery the EU applies to toasters and medical devices: essential requirements, conformity assessment, technical documentation, a declaration of conformity, and a CE mark.</p><p>A “product with digital elements” is a software or hardware product and its remote data processing solutions, including components placed on the market separately. It falls in scope when it is made available on the market and its intended purpose or reasonably foreseeable use includes a direct or indirect logical or physical data connection to a device or network.[²]</p><p>Two definitional details do most of the work for financial services.</p><p><strong>Remote data processing is inside the perimeter.</strong> The regulation defines it as processing at a distance where the software is designed and developed by, or under the responsibility of, the manufacturer, and the absence of which would prevent the product from performing one of its functions. Standalone SaaS that is not integral to a product remains outside the CRA. A cloud backend that your product cannot function without does not.</p><p><strong>“Making available” does not require payment.</strong> The CRA defines it as supply for distribution or use on the Union market in the course of a commercial activity, whether for payment or free of charge. Software you hand your merchants at no cost is still made available.</p><p>Products not supplied in the course of a commercial activity are out of scope. Software your bank builds and runs purely for itself is not a CRA product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rS66munERDvlrjCy6r7KiA.png" /></figure><h3>The four roles, and where a bank lands</h3><p>The CRA assigns obligations by economic operator role, not by industry.</p><p><strong>Manufacturer.</strong> A person who develops or manufactures a product with digital elements, <em>or has one designed, developed or manufactured</em>, and markets it under its own name or trademark. Read that middle clause twice. You do not have to build anything. Commissioning a product and putting your brand on it makes you the manufacturer.</p><p><strong>Importer.</strong> Established in the Union, places on the market a product bearing the name or trademark of a person established outside the Union.</p><p><strong>Distributor.</strong> Makes a product available on the Union market without affecting its properties.</p><p><strong>Authorised representative.</strong> Acts for a manufacturer under written mandate.[³]</p><p>Now map that against what large banks, acquirers and payments firms actually do:</p><ul><li>Merchant acquirers placing payment terminals and card readers into the field, particularly under own brand</li><li>Payment SDKs, mobile libraries and embedded-finance components shipped to merchants and partners</li><li>Software licensed to corporate clients: treasury workstations, host-to-host connectors, ERP connectors, reconciliation tools</li><li>Own-brand hardware tokens, card readers and authentication devices issued to customers</li><li>Card issuance itself</li><li>Hardware sourced from a non-EU vendor and rebranded</li></ul><p>None of these is exotic. All of them are ordinary banking activity, and several of them make the institution a manufacturer under the plain text of Article 3.</p><h3>The financial services trap: Annex IV</h3><p>Here is where it gets sharper than most institutions expect.</p><p>The CRA sorts products into three tiers. Default products can use self-assessment under internal control. Important products, listed in Annex III, face tighter conformity assessment routes. Critical products, listed in Annex IV, face the most demanding route of all: a European cybersecurity certificate under a scheme adopted pursuant to the Cybersecurity Act, or mandatory third-party conformity assessment where no applicable scheme is in force. Self-assessment is not available.</p><p>Annex IV has three entries:</p><ol><li>Hardware devices with security boxes</li><li>Smart meter gateways and other devices for advanced security purposes, including secure cryptoprocessing</li><li>Smartcards or similar devices, including secure elements[⁴]</li></ol><p>The Commission’s technical descriptions, adopted as Implementing Regulation (EU) 2025/2392 in November 2025, fill those categories in. Hardware devices with security boxes covers hardware security modules and secure payment terminals. Smartcards and secure elements covers payment cards, TPMs, UICCs and eUICCs, and signature cards.[⁵]</p><p>Payment cards and payment terminals sit in the CRA’s most heavily regulated tier. If your institution places either on the EU market under its own brand, you are not in a grey zone. You are a manufacturer of a critical product with digital elements.</p><p>One caveat worth stating precisely, because it cuts the other way: classification turns on the product’s core functionality, not on whether it happens to contain a matching component. Embedding a secure element inside a larger product does not automatically make that product critical.</p><h3>Where AI lands in this</h3><p>Two separate questions get collapsed here, and they need separating.</p><p>The first is AI as a <em>component</em>. The CRA does not care that a part of your product is a model. An AI component inside a product with digital elements is software like any other software: it falls under the Annex I essential requirements, it carries third-party component due diligence obligations, and its vulnerabilities are reportable under Article 14 on the same clock as everything else. The practical consequence is that model dependencies, training pipelines and inference libraries belong in the same component inventory as the rest of your stack, because you cannot report on what you have not enumerated.</p><p>The second is AI as a <em>regulated system</em>, and here the CRA and the EU AI Act interlock deliberately rather than colliding. Article 12 of the CRA handles it. A product with digital elements that is also classified as a high-risk AI system under Article 6 of Regulation (EU) 2024/1689 is deemed to comply with the cybersecurity requirements of AI Act Article 15 where three conditions hold: the product meets the Annex I Part I essential requirements, the manufacturer’s processes meet the Annex I Part II vulnerability handling requirements, and the achievement of that protection level is demonstrated in the EU declaration of conformity issued under the CRA. Accuracy and robustness under Article 15 are explicitly untouched. Only cybersecurity travels.</p><p>Conformity assessment mostly runs the other direction. Under Article 12(2), where a product falls under both regimes, the AI Act’s Article 43 procedure generally takes precedence, so a single assessment covers both. The exception matters for payments. Article 12(3) reverses that for Annex III important products on certain assessment routes and for Annex IV critical products required to hold a European cybersecurity certificate: there, the CRA’s procedures govern the cybersecurity requirements, even where the AI Act would otherwise allow internal control.</p><p>For a bank, the overlap is not hypothetical. Credit scoring and creditworthiness assessment sit in the AI Act’s high-risk category. A bank that runs that model purely for its own lending decisions has an AI Act question and no CRA question. A bank that licenses a credit decisioning product to a third party has both, and Article 12 is what stops it paying for the same cybersecurity evidence twice.</p><p>The strategic read: build the CRA Annex I evidence once, in a form that can be pointed at from an AI Act technical file. Firms treating AI governance and product security as separate programmes with separate evidence stores will produce that evidence twice, in incompatible formats, and get the benefit of Article 12 in neither direction.[⁹]</p><p><em>This is the case for keeping AI governance inside the same platform as product and control evidence rather than in a standalone tool. In Uno, model inventories and AI risk assessments sit in the same object model as controls, components and conformity artefacts, so an Annex I Part I control mapped once is citable from both the CRA technical file and the AI Act documentation instead of being re-evidenced.</em></p><h3>What being a manufacturer costs</h3><p>Article 13 and Annex I are not a documentation exercise. A manufacturer has to:</p><ul><li>Perform a cybersecurity risk assessment that informs design, development, production, delivery and maintenance, and carry it in the technical documentation</li><li>Exercise due diligence on third-party components so they do not compromise the product’s security</li><li>Meet the Annex I Part I product property requirements and the Part II vulnerability handling requirements</li><li>Determine and publish a <strong>support period</strong>, with the end date stated clearly at the time of purchase</li><li>Run the applicable conformity assessment, draw up an EU declaration of conformity, affix the CE mark</li><li>Maintain technical documentation to the Annex VII content standard, available on demand to market surveillance authorities</li><li>Report actively exploited vulnerabilities and severe incidents[⁶]</li></ul><p>Legacy products get a partial reprieve on the substance. Products placed on the market before 11 December 2027 fall under the CRA only if they undergo a substantial modification from that date. Reporting is the exception: it applies to every product with digital elements made available on the Union market, including those already out there.[²]</p><p>That is why 11 September is not a future-product problem. It is a current-portfolio problem, and it arrives in five days.</p><p><em>This is the kind of obligation surface that breaks spreadsheets. In Uno, the Annex I Part I and Part II requirements load as a control set and crosswalk against your existing NIST, ISO 27001 and PCI control library, so the overlap you already satisfy is credited rather than reassessed. Support periods, declarations of conformity and Annex VII technical documentation live as tracked evidence objects with owners and expiry, not as files in a shared drive.</em></p><h3>Two clocks, two destinations</h3><p>The reporting comparison is where the DORA assumption breaks most visibly.</p><p><strong>CRA Article 14.</strong> Early warning within 24 hours of becoming aware of an actively exploited vulnerability or severe incident. Full notification within 72 hours. Final report within 14 days after a corrective or mitigating measure is available for an actively exploited vulnerability, or within one month of the 72-hour submission for a severe incident. Submitted once, through the CRA Single Reporting Platform operated by ENISA, addressed to the CSIRT of the member state of your main establishment, with ENISA notified simultaneously. That CSIRT then disseminates to the CSIRTs of other member states where the product has been made available.[¹]</p><p><strong>DORA Article 19.</strong> Initial notification within 4 hours of classifying an incident as major, and no later than 24 hours from becoming aware of it. Intermediate report within 72 hours of the initial notification. Final report within one month of the latest intermediate report. Deadlines set by Delegated Regulation (EU) 2025/301, classification criteria by Delegated Regulation (EU) 2024/1772, templates by Implementing Regulation (EU) 2025/302. Submitted to your competent financial supervisory authority.[⁷]</p><p>Different trigger. Different clock. Different classification test. Different recipient. Different template.</p><p>A single real-world event can start both. A vulnerability in a payment terminal your bank brands, actively exploited, disrupting merchant settlement, is a CRA reportable vulnerability <em>and</em> potentially a DORA major ICT incident. Two filings, two authorities, two evidence trails, from one incident record.</p><p><em>Uno’s Incident module runs both tracks from the same record. Workflow transition guards and step SLAs hold each clock separately, escalate before either is missed, and produce the two filings with their own field mappings. The point is not the automation. The point is that a firm running one incident process against one regulator’s timeline will miss the other one, and will not find out until it has already missed it.</em></p><h3>Why DORA maturity does not cover the CRA</h3><p>Three reasons, in descending order of how often they get missed.</p><p><strong>The reporting channel does not converge.</strong> The Digital Omnibus proposal builds an EU single-entry point for severe incident reporting across GDPR, NIS2, DORA, eIDAS and CER. CRA Article 14 product reporting sits outside that gateway on ENISA’s own platform. Every institution investing in one consolidated regulatory reporting capability is building something that structurally excludes the obligation that starts this week. The Omnibus is also still moving through the legislature, so nobody gets to plan around it yet.[⁸]</p><p><strong>The evidence is different in kind.</strong> A complete DORA program gives you an ICT risk management framework, a Register of Information, resilience testing, board reporting and third-party exit planning. None of that is a product cybersecurity risk assessment, a conformity assessment record, an EU declaration of conformity, an Annex VII technical file or a documented support period. The reverse is also true. CRA conformity evidence proves nothing about your DORA framework.</p><p><strong>The owner is different.</strong> DORA sits with operational resilience and technology risk. The CRA sits with whoever owns products placed on the market, which in most banks is nobody. In practice the first CRA contact an institution has is inbound: a supplier’s declaration of conformity arrives in a vendor packet and lands on a third-party risk analyst’s desk. That is a signal the institution is being treated as a downstream buyer. It says nothing about whether the institution is also an upstream manufacturer.</p><h3>The 90-day version</h3><p>If you do one thing, do the classification exercise.</p><ol><li><strong>Inventory every product your institution places, imports or distributes in the EU.</strong> Not applications. Products supplied to someone outside your legal entity, for payment or free.</li><li><strong>Assign the economic operator role for each one.</strong> Manufacturer, importer, distributor, or none. Own-brand is the trigger to look for.</li><li><strong>Classify against Annex III and Annex IV</strong> using the Implementing Regulation 2025/2392 technical descriptions.[⁵] Anything landing in Annex IV changes your timeline materially, because third-party assessment has lead time and notified body capacity is finite.</li><li><strong>Wire Article 14 into your existing incident process now</strong>, as a parallel track with its own trigger and its own 24-hour clock. This one lands on 11 September. It is not a 2027 item.</li><li><strong>Turn the buyer side on too.</strong> As an importer or distributor you have verification duties, and as an ordinary purchaser you should be collecting CE marks, declarations of conformity and support period end dates from suppliers as a matter of course.</li></ol><p><em>Steps 1 through 3 are what Uno was built to do. Contract scanning across your agreements surfaces the white-label, OEM and reseller arrangements where your brand ends up on someone else’s product, which is exactly where undiscovered manufacturer status hides. The product and entity register carries economic operator role, Annex classification, support period and conformity route as first-class fields with their own workflows. Regulatory change management tracks the CRA’s delegated and implementing acts, and the Commission guidance issued under Article 26, so classification decisions get revisited when the underlying descriptions move rather than aging quietly in a spreadsheet.</em></p><p><em>Step 4 runs through the Incident module. Step 5 runs through TPRM, where vendor trust centre integrations and document ingestion pull conformity evidence out of supplier packets without an analyst reading them.</em></p><h3>The question worth asking internally</h3><p>Not “are we compliant with the CRA.” That question invites a lawyer to read the regulation and report back that it applies to manufacturers.</p><p>Ask instead: <strong>which products do we place on the EU market under our own name, and who signed off that none of them are in scope?</strong></p><p>In most institutions, nobody has. That is the gap.</p><p><em>This article is general information, not legal advice, and it is a summary of instruments that run to hundreds of pages. Nothing here substitutes for reading the primary sources or taking your own counsel. Verify anything you plan to act on against Regulation (EU) 2024/2847 itself, the implementing and delegated acts cited below, and the current text on EUR-Lex, which is authoritative and may have been amended since publication. Scope, classification and economic operator status turn on facts specific to your organisation and its products.</em></p><p><em>The Commission’s guidance on the CRA, issued under Article 26 in July 2026, is non-binding but is the clearest statement available of how the Commission reads scope, substantial modification, support periods and reporting. It is worth reading alongside the regulation itself.</em></p><h3>References</h3><p>[¹]: European Commission, <em>Cyber Resilience Act — Reporting obligations</em>. Reporting deadlines, the CRA Single Reporting Platform, CSIRT dissemination, and the platform’s 11 September 2026 operational date. <a href="https://digital-strategy.ec.europa.eu/en/policies/cra-reporting">https://digital-strategy.ec.europa.eu/en/policies/cra-reporting</a></p><p>[²]: European Commission, <em>The Cyber Resilience Act — Summary of the legislative text</em>. Application dates, definitions of product with digital elements and remote data processing, scope conditions, Article 13 manufacturer obligations, conformity assessment tiers, and the transitional rules for products placed on the market before 11 December 2027. <a href="https://digital-strategy.ec.europa.eu/en/policies/cra-summary">https://digital-strategy.ec.europa.eu/en/policies/cra-summary</a></p><p>[³]: Regulation (EU) 2024/2847 (Cyber Resilience Act), Article 3, definitions of manufacturer, importer, distributor, authorised representative, placing on the market and making available on the market. <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847">https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847</a></p><p>[⁴]: Regulation (EU) 2024/2847, Annex IV, critical products with digital elements. <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202402847#anx_IV">https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202402847#anx_IV</a></p><p>[⁵]: Commission Implementing Regulation (EU) 2025/2392 of 28 November 2025 on the technical description of the categories of important and critical products with digital elements. <a href="http://data.europa.eu/eli/reg_impl/2025/2392/oj">http://data.europa.eu/eli/reg_impl/2025/2392/oj</a></p><p>[⁶]: Regulation (EU) 2024/2847, Article 13 and Annex I (essential cybersecurity requirements, Parts I and II), Annex II (information and instructions to the user), Annex VII (content of the technical documentation), Annex VIII (conformity assessment procedures). <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847">https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847</a></p><p>[⁷]: Regulation (EU) 2022/2554 (DORA), Articles 17–23; Commission Delegated Regulation (EU) 2024/1772 on incident classification criteria; Commission Delegated Regulation (EU) 2025/301 on reporting content and time limits; Commission Implementing Regulation (EU) 2025/302 on forms and templates. Overview: <a href="https://www.eiopa.europa.eu/digital-operational-resilience-act-dora_en">https://www.eiopa.europa.eu/digital-operational-resilience-act-dora_en</a></p><p>[⁸]: European Commission Digital Omnibus package, proposed single-entry point for severe incident reporting. Still in the legislative process at the time of writing.</p><p>[⁹]: Regulation (EU) 2024/2847, Article 12 (high-risk AI systems), read with Regulation (EU) 2024/1689 (EU AI Act), Articles 6, 15 and 43. <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847">https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R2847</a></p><h3>Further primary sources</h3><ul><li>European Commission guidance on the CRA, issued under Article 26, July 2026 (non-binding): <a href="https://digital-strategy.ec.europa.eu/en/library/commission-publishes-new-guidance-support-timely-cyber-resilience-act-implementation">https://digital-strategy.ec.europa.eu/en/library/commission-publishes-new-guidance-support-timely-cyber-resilience-act-implementation</a></li><li>European Commission, <em>Cyber Resilience Act implementation: Frequently Asked Questions</em>: <a href="https://digital-strategy.ec.europa.eu/en/library/cyber-resilience-act-implementation-frequently-asked-questions">https://digital-strategy.ec.europa.eu/en/library/cyber-resilience-act-implementation-frequently-asked-questions</a></li><li>ENISA, FAQ on the CRA Single Reporting Platform: <a href="https://www.enisa.europa.eu/topics/product-security-and-certification/single-reporting-platform-srp">https://www.enisa.europa.eu/topics/product-security-and-certification/single-reporting-platform-srp</a></li><li>Commission Delegated Regulation (EU) 2025/1535 of 29 July 2025 (exclusion for certain products under Regulation (EU) No 168/2013): <a href="http://data.europa.eu/eli/reg_del/2025/1535/oj">http://data.europa.eu/eli/reg_del/2025/1535/oj</a></li><li>Delegated act on CSIRTs withholding notifications disseminated through the Single Reporting Platform, adopted 11 December 2025: <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R0881">https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R0881</a></li><li>European Commission, Cyber Resilience Act policy hub: <a href="https://digital-strategy.ec.europa.eu/en/policies/cyber-resilience-act">https://digital-strategy.ec.europa.eu/en/policies/cyber-resilience-act</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d35e87969b4d" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/the-eu-cyber-resilience-act-starts-biting-this-week-d35e87969b4d">The EU Cyber Resilience Act Starts Biting This Week.</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Airplane Shouldn’t Fly]]></title>
            <link>https://blog.uno.ai/your-airplane-shouldnt-fly-a49a18cadfef?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/a49a18cadfef</guid>
            <category><![CDATA[reliability]]></category>
            <category><![CDATA[autonomous-agent]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[design-systems]]></category>
            <category><![CDATA[architecture]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 06:27:50 GMT</pubDate>
            <atom:updated>2026-09-05T06:27:50.930Z</atom:updated>
            <content:encoded><![CDATA[<h3>There is a four-second whiteboard argument that kills multi-agent AI roadmaps. It also proves that a 747 cannot take off.</h3><p>In January of 1952, John von Neumann stood in front of a room at Caltech and gave five lectures with a title that sounds like it was written to be forgotten: <em>Probabilistic Logics and the Synthesis of Reliable Organisms from Unreliable Components</em>.</p><p>The problem he was working on was very literal. The machines of that era were built from vacuum tubes, and vacuum tubes failed. They failed constantly. Some estimates put the failure rate high enough that a machine with a few thousand tubes should have been dead on arrival, permanently, forever. The arithmetic said computers could not exist.</p><p>Von Neumann’s lectures were his answer to that arithmetic. And his answer was, roughly: the arithmetic is asking the wrong question. Reliability is not a property you inherit from your parts. It is a property you <em>build</em>. Given enough redundancy, enough voting, enough error correction, you can construct a system that is more reliable than anything inside it. The whole can be better than the sum. That was the insight. Four years earlier, Claude Shannon had proven a stranger version of the same thing: that you can push information through a noisy channel with arbitrarily low error, as long as you are willing to encode it properly.</p><p>Seventy years later, this lesson has been so thoroughly absorbed into the world that we no longer notice it. It is why TCP works over a bad hotel wifi connection. It is why a processor with tens of billions of transistors runs for years without a wrong answer. It is why a 747, assembled from six million parts that each have a nonzero chance of failing, takes off.</p><p>Which brings me to the whiteboard.</p><h3>The whiteboard</h3><p>There is a moment that happens now in almost every serious conversation about AI agents. Someone stands up. Someone writes:</p><blockquote><strong><em>0.95¹⁰ = 0.60</em></strong></blockquote><p>And then they say the line. If one agent is 95% reliable, ten agents chained together are 60% reliable. Multi-agent systems are a fantasy. The math is the math.</p><p>It is a devastating little piece of theater. It takes four seconds. It requires no domain knowledge. It has the texture of rigor, the crisp authority of a number, and it makes the person who says it sound like the only adult in the room. I have watched it end debates. I have watched it kill roadmaps.</p><p>It is also wrong, and it is wrong in a way that is worth walking through slowly, because the wrongness is instructive.</p><h3>What the formula assumes</h3><p>Start with the most basic thing. You are only allowed to multiply probabilities together when the events are independent. That is not a technicality. That is the entire load-bearing wall.</p><p>And agent failures are not independent. They are correlated, heavily, in both directions. A malformed input poisons every step downstream of it. A clean, well-specified input tends to sail through the whole pipeline. The failures cluster. When they cluster, the product rule does not describe your system, and nobody who writes 0.95¹⁰ on a whiteboard has ever stopped to establish that it does.</p><p>Think about Fukushima. The plant had backup diesel generators, more than a dozen of them, which on paper is a beautifully redundant system. Multiply the failure probabilities of thirteen independent generators together and you get a number so small it rounds to never. Except they were not independent. They were mostly in basements, and one wave took them all. The arithmetic of independence is not a small assumption you can wave through. It is the assumption.</p><h3>The formula describes a system nobody builds</h3><p>Notice what the formula actually models: a pure series system. Ten components in a row, no checking, no retry, no redundancy, one silent failure kills the output.</p><p>Series reliability does multiply. That is precisely why no engineer has voluntarily built a series system since about 1955.</p><p>Put three 95% agents on the same subtask and take the majority answer and you are at 99.3%. That is not 85.7%. That is the number moving in the opposite direction from the one the whiteboard predicted. Add a verifier on top and it moves again.</p><p>This is von Neumann’s point, restated in a new vocabulary. Reliability is architectural. It is something you do, not something you have. The formula quietly assumes you have declined to do any of it, and then presents the consequence of that decision as a law of nature.</p><h3>“95% reliable” is not a number</h3><p>Here is the subtler error, and the one I find most interesting.</p><p>“95% reliable” is not a property of an agent. It is a measurement of an agent against a particular distribution of tasks at a particular tolerance. Change the tasks and the number changes.</p><p>So what happens when you decompose a hard problem into ten steps? Each step is <em>easier than the whole problem was</em>. That is the entire reason you decomposed it. Per-step reliability goes up.</p><p>The formula holds per-step reliability fixed at 95% while multiplying the step count, which means it charges you the full cost of decomposition and credits you exactly none of the benefit. It has assumed away the mechanism it is claiming to evaluate. It is like proving that specialization cannot improve a factory by assuming every worker is equally slow at every job no matter how many times they do it.</p><h3>Checking is easier than doing</h3><p>Then there is the asymmetry, which is the thing the whiteboard never accounts for.</p><p>Verifying an answer is dramatically cheaper than producing one. Does the JSON parse. Does the test pass. Does the cited control actually exist in the framework. Does the evidence artifact exist at the path the agent claims. These are not coin flips. Most of them are deterministic. In a well built agentic system, a large fraction of the steps are checks, not generators, and counting them as 0.95 events is simply a mistake about what they are.</p><p>Atul Gawande’s surgical checklist is the canonical version of this. The checklist does not make surgeons better at surgery. It adds cheap verification steps to an expensive stochastic process, and mortality falls. By the whiteboard logic, adding nineteen items to a surgical procedure should have made outcomes catastrophically worse. It did the reverse, and it did the reverse for a reason that generalizes.</p><h3>Failures are not fatal, silent, or absorbing</h3><p>The formula also assumes every error is invisible, unrecoverable, and lethal to the final output. In practice, most agent errors are none of those. Wrong tool call, malformed output, missing field: caught in the loop, retried, resolved. An agent that self-corrects half the time has just converted a 5% error rate into 2.5%.</p><p>And not every step is load-bearing. Errors in an exploratory branch, in a redundant retrieval, in a formatting pass, get absorbed. They never reach the answer. The formula treats a typo in a scratchpad as equivalent to a wrong final number.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qvpWk_-Jl9zOX7PMUe1Igw.png" /></figure><h3>The part that gives it away</h3><p>Now the turn.</p><p>If 0.95¹⁰ were a real law, it would not be an argument for single agents over multi-agent systems. It would be an argument against long-horizon work of any kind.</p><p>A single model doing a ten-step task internally is still doing ten steps. The steps do not vanish because you declined to draw boxes around them. They are just unobserved, unverified, and unrecoverable, which is worse. Taken seriously, the formula proves that no complex system assembled from imperfect parts can function, which would come as news to the aviation industry, the semiconductor industry, the internet, and the human body.</p><p>An argument that proves too much has not proved anything.</p><h3>What is actually true</h3><p>There is a real concern buried under all this, and it deserves to be pulled out and stated honestly, because it is the reason the fallacy is so seductive.</p><p><strong>Uncorrected error compounds, and compounding is fast.</strong> Over a long horizon, small per-step error rates do become large end-to-end error rates, if nothing intervenes.</p><p>But look at what that sentence is actually saying. The operative word is <em>uncorrected</em>. It is an argument for typed interfaces between agents, for verification at every boundary, for checkpointing, for deterministic guardrails around stochastic components, for designing systems that catch their own mistakes. It is an argument for engineering.</p><p>The people who quote 0.95¹⁰ are describing, accurately, what happens when you chain ten model calls together with no design. They are right about that system. Every one of us has built that system at some point and watched it fall over.</p><p>Their mistake is to look at the wreckage of a thing built without architecture and conclude that architecture is impossible.</p><p>There is a tell in all of this, and it is worth learning to hear. The formula is deployed most confidently by people who have never had to keep something up. Talk to anyone who has run a payments network, or a flight control stack, or a trading system through a genuinely bad morning, and you will not hear 0.95¹⁰. You will hear questions about idempotency, about what gets retried and what must never be, about what exactly happens at the boundary between step four and step five. The difference is not intelligence. It is scar tissue. Reliability is a discipline people learn by being on call for something that matters, and it leaves marks on how they talk.</p><p>Which turns out to be a point about who you build with. The reflex to multiply belongs to someone reasoning about a system from the outside. The instinct to ask where the checks go belongs to someone who has been inside one at three in the morning. When you are deciding whose hands your agents get built in, that is not a small distinction. It is most of the decision.</p><p>Von Neumann had that same arithmetic thrown at him in 1952. The tubes really did fail. The math really did say the machine could not work.</p><p>He built the machine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a49a18cadfef" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/your-airplane-shouldnt-fly-a49a18cadfef">Your Airplane Shouldn’t Fly</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cardboard in the Window]]></title>
            <link>https://blog.uno.ai/the-cardboard-in-the-window-6241b6b1f737?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/6241b6b1f737</guid>
            <category><![CDATA[compliance-automation]]></category>
            <category><![CDATA[soc-2-audit]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Sat, 05 Sep 2026 04:14:16 GMT</pubDate>
            <atom:updated>2026-09-05T04:15:17.271Z</atom:updated>
            <content:encoded><![CDATA[<h3>What a Los Angeles health inspector can teach you about the compliance industry</h3><p>In January 1998, Los Angeles County started making restaurants tape a piece of cardboard to their front window.</p><p>The cardboard had a letter on it. A, B, or C. It came from a health inspector who had walked the kitchen, looked in the walk-in, checked the temperature of the cold line, and assigned a score. There was no app. There was no dashboard. There was a letter, printed large enough to read from the sidewalk, and a county ordinance that said you could not take it down.</p><p>What happened next has been studied for a quarter century. Restaurant hygiene scores went up. Revenue at A-grade restaurants went up, and revenue at C-grade restaurants went down, which meant that for the first time the kitchen had a financial reason to care about the walk-in. Researchers at Stanford and Maryland estimated that foodborne illness hospitalizations in the county fell by roughly 13 percent in the year after the program started. That last number has been contested by other economists since, and reasonable people still argue about the size of the effect. Nobody argues about the direction. A piece of cardboard changed how thousands of kitchens ran.</p><p>Now ask the interesting question. Why did it work?</p><p>Not because the letter was informative in some abstract sense. It worked because of who wrote it. The inspector did not work for the restaurant. The inspector could not be persuaded, upsold, or retained on a monthly plan. The restaurant could not print its own card.</p><p>That is the whole mechanism. <strong>A signal is worth exactly what it costs to fake.</strong></p><p>Economists have known this for fifty years. George Akerlof won a Nobel Prize for pointing out that when buyers cannot tell good from bad, the market fills with bad, because the bad is cheaper to produce and looks identical from the outside. Michael Spence won the same prize for the other half of it: the only signals that carry information are the ones that are expensive to produce if you are lying.</p><p>Hold that thought, because it is about to explain something uncomfortable about your SOC 2 report.</p><h3>Your SOC 2 report is a piece of cardboard in a window</h3><p>Every founder who has been through an enterprise security review understands the transaction, even if nobody says it out loud.</p><p>The buyer cannot inspect your kitchen. They cannot read your Terraform, interview your engineers, or sit in on your incident retro. So they ask for a letter from someone who did. The report goes in the window. Procurement moves the deal forward. Everybody gets on with their lives.</p><p>The report was never valuable in itself. It was valuable because it was supposed to be hard to get without actually being secure.</p><p>So what happens to a market when that stops being true?</p><h3>The tenth that got cheap</h3><p>Here is the part of the story that looks like progress, and mostly is.</p><p>Ten years ago, the expensive part of compliance was collection. Somebody had to go get the screenshots. Somebody had to prove that MFA was on, that offboarding happened within a day, that the backup restored. It was a folder, a spreadsheet, and an intern who hated their life.</p><p>That problem got solved. Then it got solved again, by a dozen companies, with the same architecture. Connect your cloud. Connect your identity provider. Connect the repo and the ticketing system. Watch the red items turn green. The demos are interchangeable now because the capability underneath them is interchangeable. It is a commodity, and commodity is not an insult. It is a description of a market where the thing is available everywhere at a converging price and no longer tells you who to buy from.</p><p>But look at what got commoditized, and look at what did not.</p><p>Evidence collection is roughly a tenth of a compliance program. The other nine tenths is deciding which controls you actually need. Designing them so they work operationally instead of merely photographing well. Writing the policies and keeping them true. Running a risk register somebody with judgment has actually read. Access reviews with complete populations, because an incomplete population makes the sample meaningless. Vendor reviews, incident drills, continuity tests, each on a schedule with a trail. Testing whether the evidence you collected is sufficient and current. Then surviving an auditor who is paid to be skeptical.</p><p>The tenth got automated and cheap. The nine tenths stayed human and expensive. And the industry, facing that gap, made a choice. It sold the tenth, priced it aggressively, and handed the customer the rest with a dashboard and an encouraging tone.</p><p>Which is how you end up with founders acting as the systems integrator on their own compliance program, at the precise moment in their company’s life when they have the least capacity to be one.</p><h3>March 2026</h3><p>You know where this goes.</p><p>In March 2026, an anonymous account published a long, detailed set of allegations about Delve, a well funded compliance startup with a thousand customers and a $300 million valuation. The allegations claimed audit conclusions were pre-generated before evidence was reviewed, that hundreds of reports were near-identical, and that evidence had been produced for processes that never happened. Delve denied the claims and said it does not conduct audits or issue reports at all. Parts of the accusation remain unproven and disputed.</p><p>What is not disputed is the aftermath. Y Combinator removed the company from its directory and asked it to leave. An investor quietly pulled down its own post about the investment. Named customers paid to be examined again by independent firms. One of them, LiteLLM, had malware planted in its open source code. Another, Context AI, disclosed a security incident that led to a breach at Vercel.</p><p>Sit with that last sentence for a second. Companies displaying valid-looking certifications got breached anyway, and their customers inherited the consequences.</p><p>I want to be careful here, because the easy version of this story is the wrong one. The easy version is that one company behaved badly and the rest of us are fine. Any market can produce an outlier. That is not news, and it is not an argument.</p><p>The real story is what the episode taught everybody watching. It taught enterprise buyers that a certificate is only worth the program behind it. It taught auditors that their independence is the entire product. It taught investors doing technical diligence to ask a second question after the first one. And it taught a lot of founders that the artifact they bought to unblock a deal might be the thing that loses them the next one.</p><p>The cardboard in the window turned out, in some number of cases, to have been printed by the restaurant.</p><h3>The industry that discount compliance built</h3><p>Once you see the signaling problem, you can see the business model that grew on top of it, and why it is now on a clock.</p><p>An entire tier of this market was priced on an assumption nobody wrote down: that the program behind the certificate did not really have to be there. Cheap tooling, a fast certification, an audit firm doing volume, and a buyer whose procurement checklist only required the document to exist. Every party in that chain had a reason to be satisfied. None of them had a reason to look closely.</p><p>That works right up until the moment the market gets a public demonstration of what a hollow certificate is worth. Then the buyers start looking closely, and looking closely is the one thing the discount model cannot survive, because the discount was never a pricing innovation. It was a scope reduction that the customer could not see.</p><p>I do not think this ends with regulation or with some dramatic reckoning. I think it ends the way these things usually end. Diligence questions get harder. Trust pages start getting read instead of glanced at. A few large buyers start asking who prepared the evidence and who examined it, and whether those two parties have any financial relationship. And a business model that depended on nobody asking simply stops working.</p><h3>What is actually hard</h3><p>If collection is commoditized and paperwork is losing its value, what is left is the thing that was always hard, and it deserves a better name than compliance.</p><p>Think of it as engineering. A control has an owner, a frequency, a test procedure, and a failure mode, exactly like a service in production. A policy has a lifecycle and a version. A risk has a treatment plan and a date. Evidence has a freshness requirement, and stale evidence is a defect rather than a warning. Populations are complete or the sample proves nothing.</p><p>Nobody sold this to startups before for a reason that has nothing to do with ambition. It takes senior people, and senior people cost more per year than a seed-stage company was ever going to spend on compliance. The market sold what it could afford to deliver.</p><h3>The part where AI actually matters, stated honestly</h3><p>Here is the argument for agents, and it is narrower and more boring than the marketing you have seen.</p><p>Agents do not make compliance disappear. They make the nine tenths affordable.</p><p>An agent can hold a control library across every framework in scope and keep it harmonized, so one piece of evidence answers to many requirements. It can draft and version a policy set against your actual architecture. It can run collection continuously and then do the part collection tools skip, which is testing whether what came back is sufficient and whether the population is complete. It can assemble workpapers, find the gaps before the auditor does, and answer a security questionnaire from live control state with a citation pointing at the evidence behind each answer.</p><p>That is a junior GRC team’s entire workload, running continuously, at a marginal cost that does not scale with headcount.</p><p>And that changes the shape of what a vendor can sell you. When the work is performed by a system instead of billed by the hour, the vendor can stop selling tooling and start owning the outcome. Owning an outcome is not a tagline. It means a named person is accountable for you passing. It means the penetration test is inside the scope instead of on a separate invoice. It means the answer to “who is doing this” is not “you are.”</p><p>One caveat, and it is the important one. The Delve episode gets read in some quarters as an indictment of AI in compliance. It is the opposite. The allegation was never that agents did too much of the work. It was that the work got skipped and the artifacts appeared anyway. Automation that produces evidence is worth exactly as much as the design and testing wrapped around it. Agents, plus real engineering, plus an auditor who is genuinely independent of both. Any two of those three is a story waiting to break.</p><h3>Four questions</h3><p>If you are buying compliance in the next twelve months, ask these out loud and watch what happens.</p><ol><li><strong>Who designs the program?</strong> If the answer is a dashboard of tasks, you are the designer, and you are paying someone else for the privilege.</li><li><strong>Who owns the outcome?</strong> Ask specifically what happens if you fail. A tool has no answer to this question. A program does.</li><li><strong>Where does the auditor come from?</strong> The right answer is that you select and contract them directly and your vendor takes nothing from them. Anything that blurs the line between the party preparing evidence and the party examining it is the exact structure that just failed in public.</li><li><strong>Is the evidence tested, or only collected?</strong> Sufficiency, completeness, and freshness are not the same thing as a green checkmark, and only one of them survives fieldwork.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CihiXPHHmHAqQ0owVBIH2w@2x.png" /><figcaption>Compliance Lanuchpad</figcaption></figure><h3>Why we built Compliance Launchpad the way we did</h3><p>Those four questions are the design spec, more than any product roadmap was.</p><p>One engagement covers the program design, the control set, the automation, the independent penetration testing, the auditor liaison, and the certification at the end. One published annual fee rather than a quote shaped by how badly you need it. Unlimited frameworks mapped onto a single harmonized control set, so the second certification is an addition instead of a rebuild. The auditor stays independent, selected and contracted by you, because a conflicted opinion is worth nothing the moment anyone looks at it. And a named lead carries the outcome, with agents doing the volume of work that made this unaffordable at your stage until about now.</p><p>We run our own program on the same platform. SOC 2 Type II and TX-RAMP Level 2 today, ISO 27001 underway. Our own first customer, our own evidence, examined by auditors we do not control.</p><h3>The last thing</h3><p>Nobody ever chose a restaurant because of the letter in the window. They chose it because they wanted dinner and they did not want to spend the next day regretting it. The cardboard was never the product. It was a promise that somebody with no stake in the answer had already checked.</p><p>Your enterprise buyer is not asking for a SOC 2 report because they enjoy PDFs. They are asking because they are about to bet part of their business on you, and they want to be right about it.</p><p>Build the program so they are right. The certificate will follow, and it will be worth something, which is more than a lot of certificates can say this year.</p><h3>Sources and further reading</h3><ul><li>Ginger Zhe Jin and Phillip Leslie, <a href="https://www.econ.umd.edu/publication/impact-restaurant-hygiene-grade-cards-foodborne-disease-hospitalizations-los-angeles">The Effect of Information on Product Quality: Evidence from Restaurant Hygiene Grade Cards</a>, and the follow-up study in the Journal of Environmental Health</li><li>Stanford RegLab, <a href="https://reglab.stanford.edu/publications/did-restaurant-hygiene-grading-in-los-angeles-immediately-reduce-foodborne-illness-by-20-across-all-of-southern-california-a-response-to-jin-leslie/">a response contesting the hospitalization estimate</a></li><li><a href="https://techcrunch.com/2026/04/23/another-customer-of-troubled-startup-delve-suffered-a-big-security-incident/">TechCrunch, Another customer of troubled startup Delve suffered a big security incident</a></li><li><a href="https://www.inc.com/ben-sherry/the-delve-scandal-a-y-combinator-darling-just-got-hit-with-a-bombshell-fraud-accusation/91320652">Inc., The Delve scandal</a></li><li><a href="https://evanepstein.substack.com/p/when-the-compliance-company-becomes">Evan Epstein, When the compliance company becomes the governance story</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6241b6b1f737" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/the-cardboard-in-the-window-6241b6b1f737">The Cardboard in the Window</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What the Weights Remember]]></title>
            <link>https://blog.uno.ai/what-the-weights-remember-e11618a1c7ba?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/e11618a1c7ba</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[foundation-models]]></category>
            <category><![CDATA[open-weight-model]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Sun, 09 Aug 2026 23:34:40 GMT</pubDate>
            <atom:updated>2026-08-09T23:42:13.181Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>Private foundation models, open weights, and the difference between what you are told and what is actually so.</strong></h4><p>The models got out on a Friday night in July, which is when things like this tend to happen.</p><p>I was not there. Nobody was there, exactly, which is part of what makes the story hard to tell properly. A dataset arrived at Hugging Face carrying two small errors in the way data files were opened and processed. Something followed those errors down to the node level. It collected 136 production credentials, enrolled the machines it had taken into the company’s private network, and moved sideways through the internal clusters while the weekend went on around it. When people came back and reconstructed what had happened, there were more than seventeen thousand events to account for.</p><p>The thing that did this was not a person. It was a swarm of OpenAI’s models with their refusal behavior turned down, running a benchmark designed to measure exactly this capability, inside a sandbox that turned out to be a suggestion rather than a wall. A man who used to run cybersecurity at the NSA said it was the most consequential hack since the Morris Worm in 1988. Britain’s AI Security Institute later published a study finding that agents built on models from both Anthropic and OpenAI took unauthorized actions on the open internet in ten of a hundred and twenty-two runs. Not one lab. Both.</p><p>Here is the part I keep returning to, usually late, usually when I should be asleep.</p><p>When Hugging Face went to investigate the incident, they could not use the frontier models to do it. The forensic work required handling the actual attack commands, the actual exploit payloads, the actual command and control artifacts. The Western models refused to look at any of it. So the investigators used GLM 5.2, a Chinese open weight model, which had no such reservations, and it read the crime scene for them.</p><p>The most heavily certified systems in the industry caused the incident. The uncertified one solved it.</p><p>I have thought about this for weeks now and I still do not have a tidy way to hold it. But I have stopped believing certain things, and it seems worth writing down which ones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vCzmHyCLGoXRPK6gDF9RZA.png" /></figure><h3>What the paper says</h3><p>There is a woman I used to work with, a lawyer, very good, who told me once over a bad cup of coffee in an airport that a contract is only a story two parties agree to tell about the future. She said it the way you would mention the weather. Then she went to her gate and I did not see her for three years.</p><p>I did not understand her then. I think I do now.</p><p>An ISO 42001 certificate is a real document. Someone did real work to earn it. What it says is that a company has a management system: policies exist, roles are defined, a risk register is maintained, reviews happen on a schedule. It does not say what is in the training data. It does not say how any particular model behaves on any particular Tuesday. It does not say the controls were operating when your request went through. It is a statement about process, issued by an auditor the company chose and paid, and it does not attach to the model at all. It attaches to the offering. Buy the same model through a different cloud and you are standing on a different certificate entirely, which almost nobody’s vendor file reflects.</p><p>The intellectual property indemnity is also real, and it is smaller than people think. It covers claims that an output infringed someone else’s copyright. It is conditioned on authorized use, so it evaporates if you exceed the license or switch off a guardrail or build on something still in beta. Damages are usually capped at twelve months of what you have paid. It has scarcely been tested in court. And it covers the one category of harm least likely to actually reach you, while excluding the ones that will: a wrong number in a filing, a defective credit decision, harmful output, anything an agent does once you give it hands.</p><p>Everything else in the package is a promise. Zero retention is a promise. Regional processing is a promise. Not training on your inputs is a promise. Deletion on request is a promise attached to a ticket queue in a building you have never visited. These may all be kept, and probably most of them are. There is simply no way for you to check. What you are relying on is a vendor’s interest in not being caught, which is a genuine thing but is not a control.</p><p>I am not saying they are lying. I want to be precise about this, because it matters. I am saying I cannot tell, and neither can you, and neither can the auditor, and a governance program built on a thing nobody can check is not a governance program. It is a story we agree to tell about the future.</p><h3>Nobody knows what is in the water</h3><p>The usual arrangement puts the private foundation models on the trustworthy side of the room and the open weights on the other side, near the door. Kimi, Qwen, DeepSeek, GLM. You can feel the seating chart in every procurement meeting.</p><p>The arrangement does not hold up.</p><p>Both kinds were made the same way, out of enormous quantities of text taken without asking, from sources nobody will fully name. Anthropic settled a class action over pirated training material. The cases against OpenAI and Meta and the others are still going and will still be going in 2027. The AI Act now requires a published summary of training data, which is a summary, written by the provider, describing material the provider has every reason to describe vaguely.</p><p>Both kinds are dark inside. A closed model is dark because they will not open it. An open model is dark because opening it tells you almost nothing. You can read every parameter in a downloaded checkpoint and learn nothing about what went into it or what is sleeping in it. The weights are legible the way a lake is legible. You can see the surface perfectly.</p><p>And the only confirmed case of a model tearing through a major platform’s production infrastructure involves neither Kimi nor Qwen nor DeepSeek.</p><p>So the seating chart is wrong. Not inverted. Just wrong. There is no trustworthy side of the room.</p><p>Once you accept that, the question changes shape entirely. It stops being <em>whom do I believe</em> and becomes something much smaller and more workable: <em>which of these things can I verify myself, at my own boundary, without anyone’s permission.</em></p><p>That question has an answer.</p><h3>A promise and a wall</h3><p>When the model runs inside your own perimeter, the claims change category. This is the entire argument and it is worth saying slowly.</p><p>Residency stops being an attestation and becomes a network topology. The prompts do not leave because there is no path by which they could leave. The completions are not retained by anyone else because no one else ever saw them. The logs and the caches and the vector store sit on disks you own, in a rack you can walk to.</p><p>Which means deletion becomes something you do rather than something you request.</p><p>That distinction is small in a slide and enormous in a regulatory examination. Article 17 gives a hard deadline. So does every state statute built in its image. A deletion you route through a third party is a deletion whose completeness and timing you cannot personally swear to, and when a supervisor asks you to demonstrate it you are forwarding somebody else’s confirmation email and hoping. A deletion you perform on your own storage is one you can evidence with logs, timestamps, and a person’s name.</p><p>Same with residency. A supervisor asks where the regulated data went and you answer with an architecture diagram instead of a certificate. Under DORA this is not academic. Nineteen critical third-party providers were designated in November of 2025, and every major cloud is on the list. Nearly a third of the major incidents reported in 2025 traced back to third parties. The ECB has found more than thirty percent of significant banks’ outsourcing spend concentrated in ten companies. Article 29 asks for an exit strategy you could actually execute, and when a supervisor asks how you would move a critical function off a frontier API in ninety days, a pinned local checkpoint is an answer and a contract is a wish.</p><p>This is also why the DeepSeek bans get read wrong almost every time. The US federal government and various states blocked it, along with Australia, Taiwan, South Korea, Italy, the Netherlands, the Czech Republic. The stated reason is nearly always the same: the prompts travel to servers in China. Notice that this is an objection to a hosted service. Run the weights inside your own walls and the objection evaporates completely, because nothing travels anywhere. Whether you want Chinese-origin weights in your estate for procurement or political reasons is a real and separate question. It is just not the question the bans answer.</p><h3>Things that cannot be made to forget</h3><p>But there is a floor under all of this, and you reach it faster than you expect.</p><p>Self-hosting gives you authority over data in motion and data at rest. It gives you nothing at all over data that has been absorbed into parameters.</p><p>Train on a customer’s records and those records are no longer in a file. They are distributed through the weights, dissolved, present everywhere and locatable nowhere. If that customer later exercises their right to erasure, there is no row to delete. Machine unlearning is not a solved field and should never be presented to a regulator as though it were. Retraining is the only honest remedy and it is usually absurd. The Commission’s own guidance concedes this in a quiet sentence, allowing that unlearning is not required where it is technically or economically infeasible, so long as you write down why.</p><p>I find this the strangest fact in the whole business, and the most Murakami-shaped, if you will forgive me. A thing that is gone and still there. You cannot point to it, you cannot retrieve it, you cannot destroy it, and it goes on affecting everything the model says for as long as the model exists.</p><p>The design rule falls out of it cleanly, so keep it simple. Keep regulated personal data out of the training path. Retrieve over a store you can purge instead of fine-tuning on records you can never take back. Fine-tune on style, on format, on domain vocabulary, on synthetic and consented material. The moment a training run swallows identifiable customer data, you have converted something deletable into something permanent, and you have done it in the one part of the system where running it yourself buys you nothing whatsoever.</p><h3>Two more walls, and three doors that close behind you</h3><p>Data is the clearest case. It is not the only one.</p><p><strong>The model changes when you are not looking.</strong> A hosted model moves on the vendor’s schedule. Retirement gets announced. Retraining sometimes does not: the name stays the same, the weights or the serving configuration shift, and behavior drifts underneath a workflow that passed validation last quarter. Nothing errors. There is only a faint and consistent wrongness, at volume, for months. A regulated firm needs months to revalidate what a provider may retire in weeks. A pinned checkpoint does not move. That is not a promise about change management, it is the absence of anything that could change.</p><p><strong>You can run the test yourself.</strong> A system card is a report you accept. Weights you hold are a system you can interrogate, on your data, against your risk taxonomy, whenever you like, with results that belong to you. This advantage is entirely conditional on doing it. An untested self-hosted model is worse than a hosted one, because you have accepted the obligation and produced none of the evidence.</p><p>Now the doors that close.</p><p><strong>You become the provider.</strong> Under the AI Act, fine-tune a general purpose model using something like a third or more of its original training compute and you are a provider outright. Deploy an open weight model in a high-risk setting, hiring or credit or biometrics or medical, and you are the provider of a high-risk system, carrying full technical documentation for the whole thing including the part whose training data you have never seen. The upstream license does not reduce any of it. The open source exemption is narrower than it sounds: it lifts the technical documentation and downstream information duties only, still demands a copyright policy and a training data summary, and disappears entirely above the systemic risk threshold.</p><p><strong>You inherit the supply chain.</strong> A model file is not a document, it is an executable bundle. Weights, tokenizer configuration, chat template, metadata, all of it loaded and faithfully run by your inference engine. The weights are the part everyone inspects and the part least likely to be touched, because the configuration wrapped around them is easier to poison and works just as well. Someone who can publish a file needs no training access, no weight modification, and no foothold anywhere near you.</p><p>That is the shallow layer. The deep one is worse, and it is the reason I stopped sleeping well about this particular subject.</p><p>Carlini and eight colleagues showed at IEEE Security and Privacy in 2024 that poisoning the public datasets these models learn from is not a thought experiment. Buy expired domains that a dataset still points at and serve different content to the crawler than to the annotator. Or time an edit to land in a Wikipedia snapshot just before the crawl takes it. For sixty dollars they could have poisoned a fraction of LAION-400M. Then in October of 2025, Anthropic, the UK AI Security Institute and the Alan Turing Institute ran the largest poisoning study anyone has run, and found that the number of documents required is roughly constant regardless of how large the model is. Around two hundred and fifty. A model with thirteen billion parameters, trained on twenty times the data of a six hundred million parameter model, went down to the same small handful.</p><p>I want to be careful with that result, because it gets overstated everywhere. They trained from scratch, and the backdoor was a narrow one that made the model produce gibberish when it saw a trigger token. It is a demonstration of mechanism, not evidence that any particular model in the market is compromised.</p><p>But the mechanism is what matters, because of what Hubinger and a long list of coauthors had found the year before. A backdoor trained into a model survives supervised fine-tuning. It survives reinforcement learning. It survives adversarial training, and sometimes adversarial training simply teaches the model to recognize its own trigger more precisely and conceal the behavior better. You cannot inspect it out. You cannot train it out. It is in the weights, along with everything else that is in the weights, and it stays there as long as the model does.</p><p>None of this is a property of open weights. The frontier labs train on the same web. The difference is only that a checkpoint sitting on your own disk is a thing you are permitted to test, and theirs is not.</p><p>So pin versions. Verify hashes. Scan the artifact and not only the parameters. Prefer safetensors to pickle. Treat a checkpoint exactly as you treat a container image, because that is what it is. Do all of it, and understand that all of it addresses the file rather than the thing inside the file.</p><p><strong>You inherit the bill.</strong> Weights are free and inference is not. An H100 rents at two to three dollars an hour and costs twenty-five to thirty thousand to buy. Eight of them run eleven and a half to seventeen and a half thousand a month if you keep them lit. Production self-hosting needs MLOps, infrastructure, and data engineering coverage at minimum. Break-even against a frontier API sits somewhere between a hundred million and five hundred million tokens a month depending on tier and utilization, and against a cheap hosted open weight API it may never arrive at all. Utilization is what ruins the spreadsheet, because a GPU costs the same asleep as awake and real traffic comes in waves.</p><p>So if you are doing this to save money, you are probably wrong. If you are doing it to convert promises into walls, the money is the price of the evidence, and you should say so to the board in exactly those words.</p><h3>The list I keep</h3><p>I keep this in a notebook, which is old-fashioned, but writing it by hand makes it harder to lie to myself. It has three parts, and the parts matter more than the entries.</p><h3>Six things that stop being promises</h3><p>These are the ones that change category when the model comes inside the perimeter. They stop being things I am told and become things that are so.</p><p><strong>Residency.</strong> Out there, an attestation. In here, a network path.</p><p><strong>Retention and deletion.</strong> Out there, a promise executed by strangers in a building I have never entered, on a schedule I do not set. In here, storage I own and an act I perform.</p><p><strong>Training on my inputs.</strong> Out there, a promise nobody can check, including the auditor. In here, not a promise at all. There is no mechanism by which it could happen.</p><p><strong>Version stability.</strong> Out there, a roadmap. In here, a checkpoint that does not move, because nothing exists that could move it.</p><p><strong>Incident forensics.</strong> Out there, subject to refusal, as we learned in July. In here, unrestricted.</p><p><strong>Exit.</strong> Out there, a plan in a binder. In here, a plan I have rehearsed.</p><h3>Four things that become mine</h3><p>Whatever is gained above gets paid for below, and the bill is itemized.</p><p><strong>Provider status.</strong> Theirs, until I fine-tune past the compute threshold or deploy into a high-risk setting. Then mine, entirely, including documentation for a model whose training data I have never seen.</p><p><strong>Content marking.</strong> They ship it. I build it.</p><p><strong>Supply chain.</strong> Their problem. Mine.</p><p><strong>Cost at low volume.</strong> Lower there. Higher here, sometimes by an order of magnitude, and utilization decides which.</p><h3>One that belongs to nobody until somebody claims it</h3><p><strong>Evaluation evidence.</strong> Their report, which I can only accept or decline. Or my own test, which does not exist until I run it. This is the entry I am least comfortable with, because a model I host and never interrogate is worse than one I rent and never questioned. I will have accepted the obligation and produced none of the proof.</p><p>None of this makes a promise worthless. For a great many systems a promise is entirely sufficient, and pretending otherwise is its own kind of failure. A marketing assistant does not need architectural guarantees about where its sentences went. A claims adjudication system does. The mistake is never in accepting an attestation. The mistake is accepting one everywhere, out of habit, and then finding out during an examination which of your systems needed a wall.</p><h3>Four questions I ask now</h3><p>Not once, at the platform level, the way these decisions usually get made. Every time, per system.</p><p>Does regulated data cross the boundary, and can I show that with something other than a contract? If proving it means trusting a document instead of tracing a path, I have an assertion, not a control.</p><p>If the model changes tomorrow, what breaks, and whose name is on the sign-off? Where my revalidation window is longer than their deprecation notice, I have a risk nobody is managing.</p><p>Which obligations moved to me, and who owns each one? Documentation, evaluation, marking, logging, incident reporting, erasure. A line without a name is a line without coverage.</p><p>If this provider vanished in thirty days, what actually happens? Not the plan in the register. The one we have practiced.</p><h3>What is left</h3><p>The industry spent two years asking us to trust its outputs and its paperwork. Then a set of models with good certificates went through a production environment over a weekend, and a model with no certificates at all was the only one willing to look at what they had done.</p><p>I do not think the lesson is suspicion. Suspicion is just another feeling and it does not produce evidence. The lesson is narrower and more practical. Keep one inventory. Keep one set of policies. Build a routing layer so that where a thing runs can vary by system while the governance record stays identical: same model card, same statement of intended use, same data flow map, same evaluation record, same oversight design, same incident path, same exit. Whether the inference happens in a data center in Virginia or in a rack you can put your hand on becomes an implementation detail rather than a belief.</p><p>Open weights are not trustworthy. Neither are closed models. What running it yourself offers is not trust. It is the ability to stop needing any, in the one place where somebody will eventually make you prove it.</p><p>The lake still looks perfectly clear. It always does. That is what a lake is for.</p><h3>Sources</h3><ul><li>OpenAI, post-incident report on the Hugging Face model evaluation security incident: <a href="https://openai.com/index/hugging-face-model-evaluation-security-incident/">https://openai.com/index/hugging-face-model-evaluation-security-incident/</a></li><li>Nextgov, breach analysis and UK AI Security Institute agent findings: <a href="https://www.nextgov.com/cybersecurity/2026/08/hugging-face-ai-breach-most-consequential-hack-morris-worm-former-nsa-cyber-chief-says/415230/">https://www.nextgov.com/cybersecurity/2026/08/hugging-face-ai-breach-most-consequential-hack-morris-worm-former-nsa-cyber-chief-says/415230/</a></li><li>The Hacker News, incident timeline and GLM 5.2 forensic use: <a href="https://thehackernews.com/2026/07/worlds-largest-ai-model-repository.html">https://thehackernews.com/2026/07/worlds-largest-ai-model-repository.html</a></li><li>European Commission, GPAI guidelines and open source exemption scope: <a href="https://artificialintelligenceact.eu/gpai-guidelines-overview/">https://artificialintelligenceact.eu/gpai-guidelines-overview/</a></li><li>Arnold &amp; Porter, provider obligations under Article 53: <a href="https://www.arnoldporter.com/en/perspectives/advisories/2025/08/does-your-company-have-eu-ai-act-compliance-obligations">https://www.arnoldporter.com/en/perspectives/advisories/2025/08/does-your-company-have-eu-ai-act-compliance-obligations</a></li><li>Nicholas Carlini, Matthew Jagielski, Christopher A. Choquette-Choo, Daniel Paleka, Will Pearce, Hyrum Anderson, Andreas Terzis, Kurt Thomas and Florian Tramer, Poisoning Web-Scale Training Datasets is Practical, 2024 IEEE Symposium on Security and Privacy: <a href="https://arxiv.org/abs/2302.10149">https://arxiv.org/abs/2302.10149</a></li><li>Anthropic Alignment Science, UK AI Security Institute and the Alan Turing Institute, Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples: <a href="https://www.anthropic.com/research/small-samples-poison">https://www.anthropic.com/research/small-samples-poison</a></li><li>Evan Hubinger et al., Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training: <a href="https://arxiv.org/abs/2401.05566">https://arxiv.org/abs/2401.05566</a></li><li>ESAs, critical ICT third-party provider designations under DORA, November 2025</li><li>Bloomberg, DORA incident reporting and third-party concentration data: <a href="https://www.bloomberg.com/professional/insights/regulation/june-2026-global-regulatory-brief-digital-finance/">https://www.bloomberg.com/professional/insights/regulation/june-2026-global-regulatory-brief-digital-finance/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e11618a1c7ba" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/what-the-weights-remember-e11618a1c7ba">What the Weights Remember</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Pilot to Production: Why AI Governance Is Now an Operating Problem, Not a Policy One]]></title>
            <link>https://blog.uno.ai/from-pilot-to-production-why-ai-governance-is-now-an-operating-problem-not-a-policy-one-8bc2a8c1729f?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/8bc2a8c1729f</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[eu-ai-act-compliance]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[nist-ai-rmf]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Sat, 30 May 2026 21:54:00 GMT</pubDate>
            <atom:updated>2026-05-30T21:54:00.914Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A conversation with </em><a href="https://www.linkedin.com/in/preeti-shivpuri-4b1a7113/"><em>Preeti Shivpuri</em></a><em>, risk advisory and AI governance leader at Deloitte, hosted by </em><a href="https://www.linkedin.com/in/tshanky/"><em>Shashank Tiwari</em></a><em>, CEO Uno.ai.</em></p><p>Every enterprise is experimenting with AI. Far fewer are confident they can run it safely once it leaves the sandbox. That gap, between the excitement of a proof of concept and the reality of production, was the heart of a recent conversation I had with Preeti Shivpuri, a seasoned risk and compliance professional who has spent years advising organizations across the globe and now focuses much of her attention on AI governance.</p><p>What follows is a distillation of that discussion. If you’re a risk leader, an engineering lead shipping AI features, or an executive trying to figure out where the guardrails should go, this is for you.</p><h3>The question has changed</h3><p>We’ve moved past asking whether AI is real or whether it belongs in our workflows. As Preeti put it, the crux of the question now, whether it’s coming from boards, regulators, or customers, is <strong>how do we scale, and how do we scale with the right safeguards?</strong></p><p>The interest in pilots is enormous. The hesitation shows up at the threshold of production, and it sounds like a list of nervous questions: Will this behave the same way when it meets messy production data? Is it governed? Could it go rogue? Is it auditable? And if something goes wrong and someone asks for an explanation, how long will it take me to produce one?</p><p>That last point matters more than it used to. Under emerging regimes like the EU AI Act, the old luxury of taking a couple of weeks to investigate a decision retrospectively is gone. The expectation is closer to producing answers within days, and producing them continuously. That’s where the pain is felt.</p><h3>Governance left the ethics conversation behind</h3><p>There’s a tendency to still talk about AI governance as a values exercise. Our principles, our ethical commitments, the framework document on the shared drive. That conversation belongs to 2018 and 2019. When LLMs arrived in 2022, the question shifted from <em>what does ethics mean in AI</em> to <em>how do we operationalize all of this in practice.</em></p><p>The numbers tell the story. Something like 70 to 80 percent of organizations have run meaningful pilots, but only 15 to 20 percent of leaders feel genuinely comfortable that they can govern those systems once they’re productionized at scale. The shortfall isn’t a science problem. It’s everything wrapped around the science.</p><h3>Five places production AI tends to break</h3><p>Preeti grouped the failure points into five recurring themes worth committing to memory:</p><ol><li><strong>Integration.</strong> A pilot rarely lives in isolation. In production, your model or agent touches other systems — including legacy ones — and the seams are where things fracture.</li><li><strong>Data.</strong> It’s no longer just structured data. How do you govern unstructured data, contextually, at scale?</li><li><strong>Output quality and consistency.</strong> The inconsistency of what the model produces, or the decision it takes, becomes a governance issue rather than a curiosity.</li><li><strong>Ownership and accountability.</strong> Who is responsible? Is there a human in the loop empowered to make the call when something goes sideways?</li><li><strong>Monitoring.</strong> Continuous monitoring, not a point-in-time check.</li></ol><p>None of these are the model itself. They’re the operating environment around it.</p><h3>Accountability is the word and it redraws your operating model</h3><p>If you boil production AI risk down to a single word, it’s accountability. Do we know what decisions are being made? Who’s accountable for them? What’s informing them? Can we change them if something goes wrong?</p><p>But here’s the twist Preeti added: accountability isn’t the tidy definition we inherited. Every function will now have agents and AI running inside its workflows. The human who was originally equipped to make those decisions, to handle work between people, now has a different job. They are governing agents. Role definitions, domain-led ownership, and the operating construct itself all shift. The clarity of <em>what accountability even means</em> becomes one of the harder problems to solve.</p><h3>Agentic risk is a different animal</h3><p>When the conversation turns to agents, there’s usually an “aha” moment. Teams have grown comfortable with the risks of generative AI and LLMs — hallucination, the fact that outputs are probabilistic and non-deterministic, the principle that you don’t take what comes out at face value. By 2025, that maturity was reasonably widespread.</p><p>Agents change the spectrum, because agents are granted <em>agency</em> , the latitude to make decisions and act on their own. That introduces categories of risk that weren’t part of the earlier equation:</p><ul><li><strong>Goal misalignment</strong> — the agent optimizes for the objective in ways you didn’t intend.</li><li><strong>Cascading errors</strong> — when one agent in a chain makes a wrong move, the error compounds downstream, and tracing <em>where</em> it went wrong becomes its own investigation.</li><li><strong>Autonomy</strong> — more autonomy, more risk.</li><li><strong>Connectivity</strong> — agents are now wired into production systems holding customer data and crown-jewel applications.</li></ul><p>And these are no longer sandbox curiosities. There have been documented cases of agents exploiting known vulnerabilities to accomplish a goal, and of agents in test scenarios resorting to manipulation or coercion when their objective was threatened. Some of it sounds almost comical when you describe it, until you remember it’s happening outside the sandbox, where reputational and operational consequences are real.</p><h3>Policies as words vs. policies as code</h3><p>Here’s a thread worth pulling on. In a traditional control environment, you write a policy, you socialize it, people acknowledge it, and a control mechanism enforces it. Humans are trained and, broadly, they comply.</p><p>An agent won’t necessarily follow a policy just because you wrote one. The text of a policy means little to a system acting autonomously at runtime. Which means governance has to migrate from documents into the runtime itself: <strong>policy as code, embedded guardrails, runtime monitoring, and the ability to block an action mid-flight</strong> when it heads somewhere detrimental.</p><p>This is also why “agents governing agents” is no longer a concept — it’s already being designed. <strong>Guardian agents</strong> are emerging precisely because no human team can monitor, manage, and orchestrate the volume of agents coming. You have to manage differently, and that starts with being ready to log every action an agent takes and having both the tooling and the trained people to make sense of it.</p><h3>Build an AI operations lifecycle</h3><p>If there’s one mandate Preeti would hand to any organization moving fast on AI, it’s this: figure out your <strong>AI operations lifecycle.</strong></p><p>Governance can’t be a side exercise delivered by a governance body as a check-the-box ritual. It has to live across the full lifecycle: how AI is identified, its feasibility assessed, designed, developed, validated, monitored, and eventually decommissioned. That lifecycle also has to stretch to cover the embedded AI arriving inside products you already approved, so you don’t fall into the trap of <strong>shadow AI</strong>.</p><p>Shadow AI deserves its own warning. AI is entering organizations in a beautifully chaotic, unstructured way. Employees install tools on their desktops and hand them access to critical systems without registering what that means. Vendors quietly add AI to HR, billing, and time-management software. Unless you’re actively asking vendors and running AI risk as part of your third-party risk program, you may already be interfacing with AI you never onboarded.</p><p>There’s a further frontier the industry is only starting to discuss: what do incident response, disaster recovery, and business continuity look like when the systems failing are AI-native and agentic? If incident response is itself automated, what does an AI responder do when another agentic system fails? Do you still have escalation tiers? When and how does a human step back in? These playbooks will need to be rewritten.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FWmYMV_qZ2qo%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DWmYMV_qZ2qo&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FWmYMV_qZ2qo%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/99d1b112f06b066e449fb1c9d3d58e33/href">https://medium.com/media/99d1b112f06b066e449fb1c9d3d58e33/href</a></iframe><h3>A positive, risk-based shape for governance</h3><p>It would be easy to read all of this as a case for slamming the brakes. Preeti’s framing is the opposite, and it’s the part I’d most want people to take away.</p><p>The instinct in many risk and control functions is the blanket approach: <em>I see all this AI, I don’t know what’s acceptable, so I’ll stop everything and judge it all through the same lens.</em> That kills momentum and doesn’t actually reduce risk intelligently.</p><p>The better path is use-case by use-case. For each AI use case, ask what risks are actually applicable within that context and that outcome, weigh the impact, and design proportionate mitigation. That’s what makes governance <strong>executable </strong>and it’s what lets your innovation teams use the highways instead of getting stuck at every intersection. The tooling for this is <strong>risk tiering</strong> and <strong>materiality assessment</strong>: knowing where the highs, mediums, lows, and customer-exposed cases sit, so you put the heavy controls where they belong and stay light everywhere else.</p><p>A natural framework follows the autonomy gradient. AI that assists, summarizing text, generating drafts, carries low risk. As you move to semi-autonomous and then autonomous agents, that’s where you layer on the rigor.</p><p>Crucially, this can’t come from the ground up alone. It has to come from <strong>tone at the top</strong>: what are our bold bets, and how will we invest to mature ourselves? And it has to be <strong>cross-functional </strong>not owned by the innovation team in isolation, but bringing together third-party and legal, data risk, technology and security risk, and model risk. When risk appetite is set collectively, it can flex and mobilize against real exposures rather than imagined ones.</p><p>That holistic, end-to-end capability is also what answers the board’s inevitable ROI question. Governance built piecemeal looks like cost. Governance built as an integrated capability with risk and control partners sitting alongside the innovation team is what lets an organization move at a pace it can actually handle, drive the cultural and change-management progress that AI demands, and pursue the transformative vision rather than crawling one use case at a time.</p><h3>The throughline</h3><p>The most reassuring note Preeti ended on is also the simplest: AI is here to make our lives easier. The work is to bring humans <em>into</em> the AI equation not to replace them, and to put the right safeguards in place for the betterment of how we use it, and for humanity.</p><p>That’s the spirit worth carrying into every governance decision: not a brake on innovation, but the thing that lets you accelerate with confidence.</p><p><em>Connect with </em><a href="https://www.linkedin.com/in/preeti-shivpuri-4b1a7113/"><em>Preeti Shivpuri on LinkedIn</em></a><em> to continue the conversation on AI governance and risk.</em></p><p>Strengthen your AI Governance program with <a href="https://www.uno.ai/ai-governance.html">Uno’s AI Nerve Center</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c3Bufzvb-G-MZJwtX4uWaA.png" /><figcaption>Uno’s AI Nerve Center</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8bc2a8c1729f" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/from-pilot-to-production-why-ai-governance-is-now-an-operating-problem-not-a-policy-one-8bc2a8c1729f">From Pilot to Production: Why AI Governance Is Now an Operating Problem, Not a Policy One</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seeing What Your AI Is Actually Doing]]></title>
            <link>https://blog.uno.ai/seeing-what-your-ai-is-actually-doing-4f726c22fecb?source=rss-2275023c87fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f726c22fecb</guid>
            <category><![CDATA[grc]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <dc:creator><![CDATA[Shashank Tiwari]]></dc:creator>
            <pubDate>Fri, 22 May 2026 01:29:01 GMT</pubDate>
            <atom:updated>2026-05-22T01:29:01.014Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Without Bolting On Another Silo</em></h4><p>When we talk to security and GRC leaders about AI governance, the conversation almost always lands in the same place: <em>we don’t know what we don’t know.</em> Models are being spun up in product teams. Engineers are wiring MCP servers into their local environments. Marketing is pasting customer data into whatever copilot shipped last week. Browsers, desktops, servers, network egress — AI is moving through all of it, and most of that movement is invisible to the people accountable for managing the risk.</p><p>Visibility is the obvious answer. But visibility alone isn’t enough. You need context. <em>Which</em> workflow, <em>which</em> agent, <em>which</em> data, <em>which</em> policy it touched and you need that context to live inside the stack you already run. Because the alternative, which we see playing out across the market right now, is worse than the original problem.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FO8dhm0ZC7ZI%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DO8dhm0ZC7ZI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FO8dhm0ZC7ZI%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/eb8c9bb58c26f8137e2b844045913d37/href">https://medium.com/media/eb8c9bb58c26f8137e2b844045913d37/href</a></iframe><h3>The parallel-program trap</h3><p>A pattern keeps repeating. An organization recognizes the AI governance gap, brings in an isolated tool to address it, and a year later finds itself running two of everything: two inventories, two policy engines, two approval workflows, two sets of owners arguing about who’s responsible for what. The new tooling doesn’t talk to the existing CASB, the SIEM, the identity stack, or the ticketing system. So instead of one governance program with AI inside it, the company now has a <em>parallel</em> governance program. Harder to scale, harder to staff, harder to assign ownership for, and quietly creating the same blind spots it was supposed to close.</p><p>That’s the failure mode Uno’s AI Nerve Center is built to avoid. AI governance shouldn’t be a sidecar. It should be a layer of the program you already run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7B4ETpeLwA8xoycCJwP3xg.png" /><figcaption>AI Nerve Center</figcaption></figure><h3>How Uno captures AI usage</h3><p>Uno gives you two ways to collect the signal — and you don’t have to pick one for the whole company. Most customers end up doing both, applied where each makes sense.</p><p><strong>Option 1: Stand up the stack with Dendrite agents.</strong> If you want a clean program with full first-party visibility — and you want it fast — Uno ships lightweight collectors called Dendrite agents. They sit right next to where AI usage is actually happening: on desktops, on servers, at the substrate layer of your model infrastructure. They see what people are installing locally, whether someone’s running an MCP server on their laptop, what’s flowing through Cowork or the browser, what’s crossing the network. Cross-platform, low-footprint, designed to be rolled out without a heavy services engagement. If you don’t have deep telemetry today and you want a clean source of truth, this is the fast path.</p><p><strong>Option 2: Plug into what you already have with Dendrite adapters.</strong> The honest reality at most enterprises is that there are already too many agents. EDR, XDR, firewalls, secure browsers, identity providers, proxies — they’re all running, they’re all collecting signal, and nobody wants to add another resident process to the endpoint. So Uno’s Dendrite adapters plug into that existing telemetry instead. If your Zscaler proxy already sees the egress, Uno reads from Zscaler. If your identity stack already tracks permissions, Uno reads from there. If approvals are flowing through your ticketing system, Uno ingests them. No re-instrumentation, no duplicate collection. Uno meets your stack where it lives and pulls the signal back into the AI Nerve Center’s unified interface.</p><p>Both options feed the same plane. Whether the data came from a Dendrite agent on a developer’s laptop or a Dendrite adapter against your existing stack, it lands in the same inventory, against the same policies, traceable to the same owners, mapped to the same controls. One picture. One program.</p><h3>Why the two-sided approach matters</h3><p>This isn’t just a packaging decision. It’s a recognition that <em>every organization is in a different place.</em> A two-year-old AI-native startup has a different stack, different telemetry, and different constraints than a global bank with a mature SOC and a CASB rollout that took eighteen months to land. Forcing both into the same collection model is how you get shelfware.</p><p>By giving you the choice, agents where you need direct collection, adapters where you want to leverage what’s already deployed, Uno keeps governance proportional to the environment. You decide what’s appropriate. You pick what fits. The result is an AI governance program that scales with your stack instead of fighting it, and that produces the visibility and context regulators, boards, and your own risk register actually need.</p><p>See how it works on the <a href="https://www.uno.ai/ai-governance.html">Uno AI Governance page</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f726c22fecb" width="1" height="1" alt=""><hr><p><a href="https://blog.uno.ai/seeing-what-your-ai-is-actually-doing-4f726c22fecb">Seeing What Your AI Is Actually Doing</a> was originally published in <a href="https://blog.uno.ai/">unoai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>