<?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 Montro on Medium]]></title>
        <description><![CDATA[Stories by Montro on Medium]]></description>
        <link>https://medium.com/@MontroAI?source=rss-10dd02c21d1f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*jnH2wpF3As2cnKE4KzAmmQ.jpeg</url>
            <title>Stories by Montro on Medium</title>
            <link>https://medium.com/@MontroAI?source=rss-10dd02c21d1f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:30:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@MontroAI/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[The AI you never approved is inside a tool you did]]></title>
            <link>https://medium.com/@MontroAI/the-ai-you-never-approved-is-inside-a-tool-you-did-57f27239787f?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/57f27239787f</guid>
            <category><![CDATA[gdpr]]></category>
            <category><![CDATA[third-party]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[suppliers]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 17:58:30 GMT</pubDate>
            <atom:updated>2026-11-17T17:58:30.896Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Most AI risk doesn’t arrive as a new app. It arrives as an update.</em></p><figure><img alt="A software tool’s release timeline: approved at the start, then later releases, the last one adding AI with no new approval, beside the line “Same tool. New model.”" src="https://cdn-images-1.medium.com/max/1024/1*_0nuLzfJ4r7GWNT1iPXbeQ.png" /></figure><p>Most firms remember the meeting where they approved their HR platform. There was a security questionnaire, a data processing agreement, a few rounds with procurement and, eventually, a yes. Nobody remembers the meeting where that same platform started summarising CVs with a language model, because there wasn’t one.</p><p>AI gets into a firm in two ways, and they’re usually talked about as if they were one. There are AI-native tools, built around a model from the first day. And there are AI-enabled tools: software a firm already uses, which added a model somewhere along the way.</p><p>The AI-native tool is the one everyone has in mind. It’s new, so it comes in through the front door. Someone asks for it, someone reviews it, and the questions about training data and where prompts are stored get asked, because the whole product is the model.</p><p>The same goes for staff pasting text into a public chatbot. That was the first worry, and in most firms it has been handled: blocked, licensed or written into a policy. It isn’t the story any more.</p><p>The AI-enabled tool is different. It was approved years ago for a job that had nothing to do with AI, and then a release added an assistant or a summariser, sometimes switched on by default. The tool kept its approved status while what it does with the firm’s data changed underneath it.</p><p>That second kind is where most of the risk sits now. The new app gets a review. The old tool that quietly switched a model on doesn’t.</p><p>Niamh is the DPO at an insurer, and her record of processing is in good shape. The HR platform is on it, approved two years ago, with the processing described as holding employee records and running payroll.</p><p>In the spring, the vendor ships an update. Recruiters can now ask an assistant to summarise applicants, and the summaries come from a model run by a third party the vendor contracts with. The release notes go to the HR systems admin, who sees a useful feature and leaves it on.</p><p>Nothing in Niamh’s record has changed, and parts of it are no longer true. There’s a new purpose, a new party underneath the vendor, and a data processing agreement signed before the feature existed. It says nothing about models at all.</p><p>Nobody dropped the ball here. In most firms, approval is an event: it happens at onboarding, it produces a yes, and the yes is filed. Everything after that is treated as the same tool, because by name it is.</p><p>The people who see the change are admins, users and the vendor’s account manager. The people who own the record usually aren’t on that list.</p><p>Most firms reach for an AI policy. It’s worth having, but it’s aimed at a different problem. A policy tells staff what they may do with AI; it has nothing to say about what a supplier decided to do with AI on the firm’s behalf.</p><p>The other reflex is to re-review every supplier on a schedule. It catches the change eventually, long after the data has started moving somewhere new.</p><p>An AI approval that happens once isn’t approval. Re-approving every tool after every release would be absurd. What helps is treating the approval as a decision about what the tool does, so that when what it does changes, the decision comes back to the person who owns it.</p><p>There’s a contract side to this too. Opt-in AI should be a term in the agreement, not a favour from the vendor. If a supplier has to ask before a model touches the firm’s data, the change arrives as a question to someone, instead of as a line in the release notes.</p><p>Here’s one to try in an afternoon. List the tools you approved before last year. Which of them have added an AI feature since, and who decided that was fine?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=57f27239787f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A sub-processor list is not a register of information]]></title>
            <link>https://medium.com/@MontroAI/a-sub-processor-list-is-not-a-register-of-information-a826020ae9f7?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/a826020ae9f7</guid>
            <category><![CDATA[register]]></category>
            <category><![CDATA[governance]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[third-party-risk]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 14:13:51 GMT</pubDate>
            <atom:updated>2026-11-12T14:13:51.745Z</atom:updated>
            <content:encoded><![CDATA[<p><em>They start from the same supplier names. They answer different questions.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Kr71b4GhEZCMm011shVzAw.png" /></figure><p>Ask the DPO and the head of compliance at a financial firm to show you the supplier list, and there’s a fair chance you’ll be handed two different spreadsheets. Each will share most of the same names. And each team will quietly assume the other one is keeping the real record.</p><p>They’re both right, and that’s the problem.</p><p>They’re keeping two different records that happen to start from the same column of names.</p><p>One is a sub-processor list. The other is a register of information. From a distance they look like the same document; up close, they answer different questions for different people.</p><p>The sub-processor list is the one most firms already know well. It grew out of privacy work, and it answers a privacy question: who touches personal data on our behalf, where, and under what terms? It’s the DPO’s record.</p><p>A good one is genuinely hard to keep. Suppliers add and swap sub-processors, sometimes with little notice, and a list that was true at signing drifts quietly afterwards. Firms that keep it current have done real work.</p><p>The register of information is a different thing with a different starting point. Under DORA, financial entities must keep and update “a register of information in relation to all contractual arrangements” for ICT services from third parties. The regulation also asks firms to distinguish the arrangements that support critical or important functions from the ones that don’t.</p><p>Read that slowly and the difference shows. The register isn’t about data. It’s about contracts, and about what each contract keeps running. A supplier that never sees a single customer record can still sit near the top of it.</p><p>Take Ciarán. He’s head of compliance at a payments firm in Dublin, and his DPO keeps an excellent sub-processor list. Every vendor that touches customer data is on it, with the sub-processors mapped and the residency noted.</p><p>Now the request arrives for the register. For each ICT arrangement, it needs the business function the service supports and whether that function is critical or important. Ciarán opens the sub-processor list, and it can’t help him.</p><p>It knows that the CRM vendor uses an email provider in another region. It has nothing to say about the infrastructure monitoring tool, which holds no personal data at all and which the payments platform can’t run without. It doesn’t hold what the firm contracted for, which function each service supports, or what happens if one of them stops.</p><p>None of this is a failure of diligence. Both records start from the same supplier names, so it’s natural to treat one as a draft of the other. And much of what’s sold to watch suppliers was shaped around the privacy question first. It does that job well. The register question was added afterwards, if at all.</p><p>The usual answer is to add a column. A “DORA” tab on the sub-processor sheet, a criticality flag, maybe a field for the contract owner. It feels like progress, and it doesn’t reach the problem. You can’t build a register by decorating a sub-processor list.</p><p>A sub-processor list tells you who a supplier uses. A register of information tells a regulator what you contracted for, which function it supports, and what happens if it stops.</p><p>That doesn’t make the sub-processor list less important. It makes it an input. When a supplier adds a sub-processor, the DPO’s record changes, and sometimes the register should change too: the new sub-processor might sit underneath a service that supports a critical function. The two records need to talk to each other. They just can’t be the same record.</p><p>So here’s a narrow test. Pick the three functions your firm would least like to see stop. From the register alone, without asking anyone, could you name every ICT arrangement that supports each one?</p><p>Third-party records are becoming decision logs. The firms that get this right will be the ones that can show who decided, on what evidence, and when.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a826020ae9f7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Right to Your Data Is Not the Ability to Get It Out]]></title>
            <link>https://medium.com/@MontroAI/the-right-to-your-data-is-not-the-ability-to-get-it-out-ac3411471cde?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac3411471cde</guid>
            <category><![CDATA[regulatory-compliance]]></category>
            <category><![CDATA[dora]]></category>
            <category><![CDATA[google-workspace]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[third-party-risk]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 13:35:20 GMT</pubDate>
            <atom:updated>2026-10-21T13:35:20.977Z</atom:updated>
            <content:encoded><![CDATA[<p><em>An exit clause promises your data back. Someone inside the firm still has to go and get it.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ody8r5uXzLTApFepABBpfA.jpeg" /></figure><p>At some point an exit becomes a job. The contract has been terminated, the transition period has started, and someone has to pull every user’s mail, files and calendars out before the service goes away.</p><p>Exit planning usually starts in the contract. DORA’s recitals say ICT contracts “should also provide for dedicated exit strategies”, with transition periods in which the provider keeps delivering the service. Legal and procurement negotiate that clause, sign it and file it where an examiner can find it.</p><p>That clause matters. It also only covers the supplier’s side of the handover.</p><p><a href="https://montro.io/blog/free-vs-paid-ai-tools-compliance">Google Workspace</a> is a useful example, because Google publishes how its full organisational export works. Only a super administrator can run it. That account has to be at least 30 days old, which rules out creating one on the day. Organisations with more than 1,000 users are asked to contact Google Workspace support before they start.</p><p>The export typically takes 72 hours and can take up to 14 days, depending on size. It comes out as zip files. In Google’s default storage, those files are deleted 60 days after the export begins.</p><p>Whoever runs the exit needs to know who holds super administrator access right now, and whether that account is old enough. They need to check that a fortnight of export time fits inside the transition period the lawyers agreed. They also need somewhere to put the zip files before Google deletes them.</p><p>Those answers sit with the people who run the admin console. The exit plan tends to be written by the people who own risk and contracts. The two groups review their work on different cycles and keep their records in different places, so a careful, signed-off plan can leave out the export entirely without anyone having cut a corner.</p><p>Slack works differently. On the Free and Pro plans, a workspace owner can export public channels but not private channels or direct messages. On Business+, owners can apply for a self-serve tool that covers private channels and DMs, and the application needs approval from the workspace’s primary owner. On Enterprise, org owners ask Slack support for access.</p><p>The export arrives as a zip file of JSON. Slack’s own help page describes files as links, so whether the attachments themselves come across is worth checking before the day rather than on it.</p><p>The question of who can run the export gets a different answer in each system. In Slack’s case the answer can include an application that someone else has to approve before the export begins.</p><p>Testing the plan helps up to a point. A tabletop exercise shows whether people know their roles; it says little about whether the export finishes. Running a full organisational export just to prove it works is a high bar for a test, so the plan gets rehearsed while the export stays theoretical.</p><p>The fix is to write the data into the exit plan alongside the contract. For each system that matters, record who can run the export and who holds that role today, how long the export takes, what format it produces and where the files land, and any clock that starts when it does.</p><p>An examiner can check those details. So can the supplier you are moving to, which may be the more practical test.</p><p>Even a finished export only proves the data left. Zip files of an entire organisation are a long way from a working system, and how long it takes to load them somewhere new depends on the destination, the data and the people doing the work. We don’t have a good general answer to that part yet.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac3411471cde" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Spreadsheets Are Not a RoPA Strategy]]></title>
            <link>https://medium.com/@MontroAI/spreadsheets-are-not-a-ropa-strategy-1aa8f206b8e5?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/1aa8f206b8e5</guid>
            <category><![CDATA[spreadsheets]]></category>
            <category><![CDATA[ropa]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[regulation]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 17:02:18 GMT</pubDate>
            <atom:updated>2026-10-15T17:02:18.028Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="On a dark navy background, a faint spreadsheet grid is stamped diagonally with “OUT OF DATE” in cerise. Beside it, large text reads “A spreadsheet isn’t a RoPA strategy,” with “RoPA strategy” in cerise. The Montro logo sits top-left." src="https://cdn-images-1.medium.com/max/1024/1*k3N6_aYucQPjQsDnhP_34g.png" /></figure><p>Spreadsheets fail at RoPA maintenance because they have no live connection to the systems where processing actually changes, no audit trail, no automated currency check, and no workflow for multi-function completion.</p><p>A RoPA ends up as a spreadsheet because the alternatives are too expensive, too heavy, or too unproven. Those are reasonable reasons. The four ways a spreadsheet fails when the supervisor asks for the register on Monday morning are not, and I think it’s worth walking through what specifically breaks rather than repeating the generic “spreadsheets are outdated” line one more time.</p><h3>The four failure modes I see most often</h3><p>Each compounds the next.</p><p>No source-of-truth integration. The spreadsheet has no live connection to the systems where processing activities actually change, the procurement system that adds new SaaS, the SSO that authenticates new tools, the expense system that catches paid subscriptions, the contract repository that tracks DPAs. Every change in the firm’s processing posture has to be manually translated into a spreadsheet update by a human who hears about it. The information arrives late and incomplete, in my experience almost always late. The register lags reality structurally, not because of effort.</p><p>No version control or audit trail. The spreadsheet records the current state. It doesn’t record who changed which entry when, why, or against which evidence. When the supervisor asks for the register at a specific historic point in time — typically because an incident requires a backwards-looking review, the spreadsheet can’t produce it. The version history exists in the file’s revision metadata if the firm uses a collaborative platform, but the metadata doesn’t contain the reasoning. The audit trail the supervisor expects doesn’t exist.</p><p>No automated currency check. The spreadsheet doesn’t know it’s out of date. There’s no mechanism that tells the DPO “three SaaS tools have been adopted in the last quarter that are not on the register” or “two processing activities have material changes since their last review.” The currency check is a manual operation the DPO has to initiate, and in practice the prompt to initiate it is usually the next audit, not the next material change.</p><p>No role-based completion logic. Article 30 entries require contributions from multiple functions — legal confirms the lawful basis, the data architect signs off on the data categories, and security has to weigh in on the technical safeguards before the DPO can assemble anything and put their name to it. In a spreadsheet, all of that happens through email and meetings, with completion status visible only to the DPO. There’s no workflow that escalates incomplete entries, surfaces blocking inputs, or produces auditable evidence that each contribution happened, and I think the absence of that workflow is the core gap in spreadsheet-based RoPA maintenance.</p><p>Each failure mode would be tractable on its own. Together, I’ve seen them produce a register that’s consistently behind, opaque to historic inspection, manually triggered, and held together by nothing more than the DPO’s institutional memory.</p><p>The supervisor request doesn’t arrive with two weeks’ notice. It arrives on a Tuesday, with a response window measured in days. The DPO who receives it usually knows immediately whether the register is up to date, not because they’ve checked it recently, but because they know when it was last updated manually and roughly how much has changed since. The spreadsheet has the entries from three months ago. The processing activities added since then exist in a procurement approval, an expense claim, and an email thread that no one thought to loop the privacy function into.</p><p>The gap between what the register says and what the firm is actually doing isn’t a compliance failure in that moment. But it becomes one the moment the supervisor sees it.</p><h3>When I’d say a spreadsheet is genuinely enough</h3><p>The honest qualifier: spreadsheets are appropriate for some firms, and I don’t think it’s useful to pretend otherwise.</p><p>The conditions I’d check: the firm has under fifty processing activities, the rate of change is low (no more than a handful of new processing activities per year), the SaaS environment is stable, the privacy function is one DPO with a clear deputy, and the regulatory exposure profile is moderate — not financial services, not health, not high-risk AI deployment. For a firm meeting all five, I think a well-maintained spreadsheet is genuinely sufficient, and alternative tooling is over-investment.</p><p>The conditions are restrictive, though. Most European mid-market firms I come across, particularly those running any meaningful AI footprint, meet two or three at most. The four failure modes above describe what happens when a firm with the wrong conditions stays on a spreadsheet anyway.</p><h3>What the alternatives offer</h3><p>I see two architectural directions resolving these four failure modes, with different cost profiles.</p><p>Documentation platforms — OneTrust, Securiti, BigID — solve this by giving a firm the components to build the integration itself: source-of-truth connections the firm configures, version control, currency tooling, and role-based workflow, all native to the platform. That depth is real, but it assumes an implementation team with the time to configure it properly. In most of the mid-market firms I’ve worked with, that’s the part that goes unfinished, which leaves the platform running at roughly the same fidelity as the spreadsheet it replaced.</p><p>Inventory-driven automation, the architecture Montro represents, along with a small handful of newer tools, takes a different starting point. The source-of-truth integration with the systems that actually change processing activities is built in rather than configured after purchase, so the audit trail and currency checks come from the inventory layer directly instead of from a firm’s own implementation work. The DPO still owns the judgement calls — purposes, lawful bases — but doesn’t own building the plumbing underneath them. We built it this way because the operating model at most mid-market firms is one DPO and a couple of deputies, not a six-person programme.</p><p>Both close the same four gaps. Which one I’d recommend comes down to the firm’s privacy programme depth, its regulatory surface, and its budget — and, more practically, which tool is sized for the team that actually has to operate it day to day, rather than which one has the longer feature list.</p><h3>What the supervisor sees</h3><p>Article 30(4) places an obligation on the controller and processor to make the register available to the supervisory authority on request. I read the implication as operational: the time from request to defensible production is itself part of what the supervisor is assessing. A register that takes two weeks to assemble from a spreadsheet, with manual cross-checks and follow-up clarifications, is responding to the request poorly even if the eventual content turns out accurate.</p><p>The Data Protection Commission, the CNIL, and BfDI have each signalled in inspection findings and Dear DPO communications that RoPA quality and currency are inspection focus areas. The pattern I see in those findings is consistent: registers correct in the abstract but stale in operational detail; registers missing entries for processing activities the firm clearly conducts; registers where the historic state at the date of an incident couldn’t be reconstructed.</p><p>The supervisor isn’t inspecting the spreadsheet. The supervisor is inspecting the register. It fails when the underlying tooling can’t keep pace with the firm’s actual processing posture, which is why I think RoPA maintenance is increasingly a governance infrastructure question, not a documentation one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1aa8f206b8e5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Third-Party Risk in the AI Era: What Your Vendor Questionnaire Misses]]></title>
            <link>https://medium.com/@MontroAI/third-party-risk-in-the-ai-era-what-your-vendor-questionnaire-misses-55078c2ea72e?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/55078c2ea72e</guid>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[third-party-risk]]></category>
            <category><![CDATA[questionnaire]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Tue, 13 Oct 2026 13:55:31 GMT</pubDate>
            <atom:updated>2026-10-13T13:55:31.035Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TqF9CvMFBTi1vMC3UdfXpw.png" /></figure><p>A vendor sends back a SIG questionnaire, every box ticked. Nowhere does it ask where their model runs, what they retain from your prompts, who their upstream model provider is, or whether your prompts entered a fine-tuning run last quarter. The questions that matter most for an AI vendor simply aren’t in most questionnaires still in use.</p><h3>What the standard questionnaires cover well</h3><p>I don’t think the SIG and CAIQ frameworks are outdated. They’ve evolved through years of practitioner input, and for a non-AI vendor they still cover the substantial majority of the real risk surface: information security management maturity, ISO 27001 alignment, incident response, business continuity, personnel security, sub-processor identification at the contractually disclosed level. Both SIG and CAIQ have added AI-specific modules over the last two years, too. The gap I’m describing is sharpest for firms still running older questionnaire versions that predate those additions, and smaller, though still present, for firms on the current AI-aware editions.</p><h3>Eight questions I’d add for any AI vendor</h3><p>Where does the vendor’s AI capability come from? OpenAI, Anthropic, Google, an open-weight model run on their own infrastructure, a fine-tune, something proprietary, each answer carries a different risk profile, and vendors that wrap a foundation model rarely volunteer which one unprompted.</p><p>What was the model trained on? For a foundation-model wrapper, this collapses into the foundation provider’s own training-data position. For a vendor running something fine-tuned or proprietary, it’s the central provenance question, and by 2026 I think any AI vendor handling anything sensitive should be able to describe their training-data lineage in terms that are useful, not just reassuring.</p><p>What does the vendor retain from your prompts? This is the one that separates vendors who’ve thought about their AI data handling from vendors who haven’t. Ask where a model runs and most people have a rehearsed answer ready. Ask what they retain from your prompts, and the room goes quiet. Some answer specifically: retention period, purpose, opt-out mechanism. Some admit they’d need to check with engineering, which is at least honest. The group that worries me most answers confidently but describes their storage policy instead of their prompt retention policy. Those aren’t the same thing, and a vendor that conflates them hasn’t mapped what happens to a client’s data between prompt submission and output generation. That gap in their thinking shows up in your risk assessment eventually, one way or another.</p><p>Where does the inference itself run, geographically? Most vendors will confirm EU data residency at the contract level without blinking. Far fewer can tell you where inference specifically runs, which is often a narrower subset of their broader processing footprint. A vendor routing prompts to a non-EU inference region while holding an EU residency clause isn’t necessarily lying, but the commitment isn’t what most customers assumed it was.</p><p>What does the vendor retain from the outputs, not just the prompts? This gets asked even less than prompt retention, and it’s a genuinely separate question: some vendors keep outputs for support and debugging, some indefinitely, some for model improvement under an opt-out most customers never exercise.</p><p>Are your prompts excluded from the vendor’s model improvement work? The contract language is one thing. The question I’d ask is whether the technical implementation enforces that carve-out, and whether the vendor would even notice if it failed.</p><p>How far down the sub-processor chain can the vendor actually see? Standard frameworks stop at the direct sub-processor. AI vendors typically sit on top of other AI providers, who have their own sub-processor lineage above them again, and that fourth-party visibility is exactly where I see DORA enforcement heading, well ahead of where most standard frameworks currently reach.</p><p>What model version is running, and how would you find out if it changed? AI behaviour isn’t stable across model versions. A vendor switching foundation models, or upgrading a fine-tune, can change how the system behaves with no contractual amendment triggered at all, so the reasonable ask is a version-change notification cadence and some evidence of regression testing on their side.</p><h3>When the full follow-up is worth running</h3><p>Running all eight questions on every vendor is overkill. Here’s the triage I’d use:</p><p>If the vendor’s AI feature produces decisions about real people — recruiting, credit scoring, fraud and KYC, customer decisioning — all eight apply, no shortcuts. If it’s processing meaningful volumes of unstructured customer or employee data, productivity tools, support, communication AI, most of the eight matter, with prompt logging and output retention as the priority. If it’s AI embedded in engineering or product tooling, code generation, devtools, analytics, the version-change and sub-processor questions matter most, since these tools integrate deep enough into a workflow that a silent regression is genuinely dangerous. And if there’s any special category data anywhere in the picture, health data, biometrics, ID documents, all eight apply, with training data and output retention as the gating items.</p><p>Outside those triggers, I think the standard questionnaire is genuinely sufficient. The triage itself is a form of AI governance: knowing which vendors actually warrant the deeper look, and why, rather than running the full eight questions everywhere out of caution.</p><h3>Where this connects to DORA and NIS2</h3><p>For financial entities, I’d treat this eight-question follow-up as the practical, operational version of Article 28’s third-party risk obligations and the criticality assessment under Delegated Regulation (EU) 2024/1502. A supervisor’s reasonable expectation is that an AI vendor gets assessed more deeply than an equivalent non-AI ICT service. This is what that deeper assessment looks like in practice.</p><p>For NIS2 entities outside financial services, the same logic runs through Article 21(2)(d)’s supply chain risk-management measure. This follow-up is one defensible way to discharge that obligation for AI vendors specifically, and a firm that’s documented its approach here is in a materially stronger position than one that hasn’t thought about it at all.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=55078c2ea72e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What to Look for in AI Compliance Software]]></title>
            <link>https://medium.com/@MontroAI/what-to-look-for-in-ai-compliance-software-db11cc171a41?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/db11cc171a41</guid>
            <category><![CDATA[compliance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <category><![CDATA[audit]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 12:11:04 GMT</pubDate>
            <atom:updated>2026-10-07T12:11:04.049Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d_KsWGjmA0rS3kSGVxgjng.png" /></figure><p>If a compliance vendor can’t find something you never told them about, nothing else on their feature list matters.</p><p>Gartner’s AI TRiSM framework puts discovery ahead of everything else it covers, and I think that ordering is correct. Every platform in this space can show you a clean dashboard of AI systems once someone’s typed them in. That’s not a compliance feature. That’s a spreadsheet with better branding. The question that actually separates these products is whether one can surface a system nobody entered: an AI feature a SaaS vendor quietly switched on inside a tool you already approved eighteen months ago, a model sitting behind an API key nobody logged, a browser extension someone installed on their own account.</p><p>I watched this play out with a mid-market client last year. Their compliance tool had a genuinely impressive setup — clean risk scoring, a mapping to two frameworks, a reporting dashboard that would look good in a board deck. Twelve systems, fully governed. Then a routine network review turned up a thirteenth: a meeting-transcription tool a project team had connected to their calendars through a personal Google account eight months earlier. Nothing in that setup had ever flagged it, because nothing in it was built to look for what it hadn’t been told about.</p><p>Ask a vendor to find something in your own environment that they haven’t been told about, live, in the room. Some will do it without blinking. A lot will pivot to a slide about their framework coverage instead.</p><h3>Enforcement</h3><p>A policy document describing how AI should be used and a system that actually enforces that policy while something is running are two different products, and a lot of what gets sold as “compliance software” is the first one with a dashboard on top. A client of mine had a written policy requiring human review of any AI-generated content before it reached a customer. The tool that was meant to support this logged every AI interaction after the fact, in a report generated once a month. It never once stopped an unreviewed piece of content from going out, because it wasn’t built to intervene, only to record. The policy existed. The enforcement didn’t, and nobody noticed the gap until a customer complained about something that should never have shipped.</p><h3>Framework Mapping</h3><p>The EU AI Act, NIST’s AI Risk Management Framework, and ISO 42001 don’t ask the same questions or weight risk the same way. Most tools are built around one of them properly and treat the others as a relabelling exercise — the same risk assessment, the same fields, with a different cover page. This doesn’t matter until it does: a firm operating in the EU and pursuing ISO 42001 certification for a client’s procurement requirements found that its AI Act risk classifications didn’t map cleanly to what the ISO standard actually required documented, and ended up redoing the assessment from scratch for the second framework rather than extending the first one.</p><h3>Audit Trail</h3><p>A report screen is not an audit trail. An audit trail survives a specific question: for this system, on this date, who approved it, and what changed since. I’ve asked vendors this in demos and watched the answer become “we’d need to check with the account team” — which means the trail lives in someone’s inbox, not in the product. Ask for the full history on one real system before you sign anything, not a curated example built for the sales call.</p><p>Discovery still comes first in the order I’d check these, because everything else depends on knowing what you’re governing in the first place. A tool that enforces policy beautifully, maps three frameworks cleanly, and produces an airtight audit trail is still only doing that for the systems it knows exist. The twelfth system, fully governed, doesn’t make up for the thirteenth one nobody found.</p><p>This is why we built <a href="https://montro.io/">Montro around discovery</a> as the starting point rather than a feature added later. It’s the harder problem to solve, and it’s usually the one a vendor is least confident about — which is exactly why it’s worth asking about first, before any of the rest.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db11cc171a41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Classify AI Systems Under the EU AI Act]]></title>
            <link>https://medium.com/@MontroAI/how-to-classify-ai-systems-under-the-eu-ai-act-15b0a04d4cb6?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/15b0a04d4cb6</guid>
            <category><![CDATA[regulatory-compliance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[eu-ai-act-compliance]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[risk-management]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 17:09:10 GMT</pubDate>
            <atom:updated>2026-09-30T17:09:10.938Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tIMC2SwCr2IuWJ5dY24ylQ.png" /></figure><p>Most EU AI Act classification work stops at the four-tier chart — unacceptable, high, limited, minimal, and that’s exactly where the expensive mistakes start. The chart is correct. It’s also not the part of the classification process that actually determines your compliance obligations. That part is Article 6, and almost nobody outside the people who’ve had to apply it in practice reads it closely.</p><h3>The Four Tiers, Briefly</h3><p>Unacceptable-risk systems are banned outright under Article 5 — social scoring, manipulative subliminal techniques, and a handful of other prohibited practices. There’s no compliance pathway here; if a system falls into this category, the answer is don’t build it, not build it carefully.</p><p><a href="https://montro.io/blog/eu-ai-act-risk-classification">High-risk systems</a> carry the heaviest obligations: risk management, data governance, technical documentation, human oversight, conformity assessment, registration in the EU database. Limited-risk systems face transparency requirements, telling a user they’re talking to a chatbot, labelling AI-generated content. Minimal-risk systems, which is most of what gets built, carry no mandatory obligations at all.</p><p>None of this is where the real work happens. The real work is figuring out which of your systems actually lands in the high-risk category, and that’s a narrower, more specific exercise than the four-box chart suggests.</p><h3>Where Classification Actually Happens</h3><p>Article 6 sets the test. A system is high-risk if it’s a safety component in a product already covered by EU product safety law, or if it falls within one of the use-case areas listed in Annex III — biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, and a few others.</p><p>Here’s the part that gets skipped. Article 6(3) allows a provider to conclude that an Annex III system is not high-risk, if it meets a narrow set of conditions: it performs only a narrow procedural task, or improves the output of a completed human activity, or detects patterns without replacing human judgement, and it doesn’t pose a significant risk of harm. If your system genuinely fits that description, it escapes high-risk classification and the obligations that come with it.</p><p>But there’s a hard override sitting inside that same article, and it’s the part I see missed most often: if the system profiles natural persons, the exception doesn’t apply, full stop, regardless of how narrow the task looks on paper. A tool that scores loan applicants, ranks job candidates, or assigns a risk category to an individual is profiling, and profiling inside an Annex III use case is high-risk automatically. I’ve reviewed classification memos that walked carefully through the narrow-task exception, concluded the system qualified for it, and never checked whether the system was profiling anyone, which made the entire analysis irrelevant the moment someone actually read Article 6 in full.</p><h3>A System That Looked Easy and Wasn’t</h3><p>A client last year built a tool to pre-screen buy-now-pay-later applications, flagging which ones needed manual underwriting review versus which could be auto-approved. The team’s first read was that this was a narrow procedural task, it wasn’t making the final credit decision, a human always signed off on declines, and it was just triaging a queue. On paper, that reads like the Article 6(3) exception.</p><p>It isn’t. The tool assigns each applicant a risk score based on their data, which is profiling, and access to essential services — credit among them, is one of the Annex III categories. The narrow-task framing didn’t matter once profiling was in the picture. The system was high-risk, full technical documentation and conformity assessment required, and the earlier classification memo had to be redone from the ground up.</p><h3>A Separate Test for General-Purpose Models</h3><p>Everything above applies to a system built or deployed for a specific use case. General-purpose AI models run on a different track entirely, classified by their own criteria rather than by which Annex III category they might touch. A foundation model is treated as carrying systemic risk once it crosses a compute threshold set in the Act, and that classification triggers its own obligations — model evaluation, adversarial testing, incident reporting to the AI Office, independent of what any downstream application built on top of it does. If your organisation is deploying a general-purpose model rather than building a narrow application, the Annex III profiling question above isn’t the first thing to check. The GPAI provider’s own systemic-risk status is.</p><h3>Document the Decision Either Way</h3><p>If you do conclude a system escapes high-risk classification under the Article 6(3) exception, that conclusion isn’t the end of the process, Article 6(4) requires the provider to document that assessment before the system goes to market, and to register it in the EU database, and national authorities can request that documentation and challenge the conclusion. “We decided it wasn’t high-risk” is not a defensible position on its own. “We decided it wasn’t high-risk, here’s the documented reasoning, and here’s why it doesn’t profile natural persons” is.</p><p>Getting this wrong runs in both directions. Under-classifying is the one everyone worries about, miss the high-risk trigger and you’re missing conformity assessments, technical documentation, and human oversight requirements you’re legally required to have, and that exposure gets worse once enforcement actually ramps up.</p><p>The one that gets less attention is over-classifying. I’ve seen a mid-market firm run a full conformity assessment on a spell-check feature inside an HR tool, purely out of caution, because nobody wanted to be the person who called it wrong. That’s not a careful compliance posture. It’s months of a small compliance team’s time spent on a system that was never going to be high-risk, while the actual high-risk tool three departments over sits unreviewed because the team’s capacity went somewhere else. The Article 6(3) test, applied properly, is what stops both versions of this, and it’s usually the second one that quietly costs more.</p><p>If you’re working through this for your own AI inventory, the profiling question is the one I’d check first, before anything else in the exception. It’s the fastest way to find out whether the rest of the analysis is even worth doing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15b0a04d4cb6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What to Document First for DORA and NIS2]]></title>
            <link>https://medium.com/@MontroAI/what-to-document-first-for-dora-and-nis2-82d25cad96ca?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/82d25cad96ca</guid>
            <category><![CDATA[inventory-management]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[documentation]]></category>
            <category><![CDATA[dora]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 13:51:53 GMT</pubDate>
            <atom:updated>2026-09-28T13:51:53.052Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ekYEYblfVwfb5uTqRU2MNg.png" /></figure><p>Most DORA and NIS2 documentation projects start with the wrong document.</p><p>Not the wrong regulation, the wrong document within it. Teams reach for the risk management framework or the full policy suite first, because that’s what “compliance documentation” sounds like it should mean. The document that actually needs to exist before any of that is one page long, has nothing to do with risk management, and rarely gets written down at all: a signed record of which regime actually governs you.</p><p>For most mid-market financial entities, that classification question settles almost everything downstream — whether DORA applies, NIS2 applies, or, DORA displaces NIS2 for you entirely under lex specialis. Get the classification wrong, or skip documenting it, and everything built on top of it inherits the mistake.</p><p>Here’s the order I’d actually put in front of a client starting from zero.</p><h3>First: The Classification Memo</h3><p>Before anything else gets documented, write down — in one page, not a slide deck — what your organisation is under each regime. Are you one of DORA’s roughly 21 named financial entity types? If yes, NIS2’s substantive cybersecurity provisions are largely displaced for you, and DORA governs. If no, but you sit in a NIS2-covered sector, NIS2 applies to you directly and in full.</p><p>This sounds like a formality. It isn’t. I’ve watched firms build a NIS2 risk-management programme in parallel with their DORA one, duplicating months of effort on requirements that don’t stack for a <a href="https://montro.io/blog/dora-register-of-information-ireland">DORA</a>-covered entity. The classification decision determines which of the next three items you’re building against — get it wrong, and everything downstream is built against the wrong regime’s clock.</p><p>Document who made the call, on what basis, and when it was last reviewed. Supervisors ask for this more often than people expect, and “we assumed” doesn’t hold up as an answer.</p><h3>Second: The ICT Asset Inventory, Ahead of the Register of Information</h3><p>DORA’s Register of Information has become the single most-discussed compliance artefact in this space this year, and and for good reason: the Q1 2026 submissions were due at the end of March, and supervisors are now auditing them actively. But the Register isn’t a form to fill in — it’s a structured database of every ICT third-party contractual relationship an organisation has, mapped to the business functions each one supports, with subcontractor chains traced at least two tiers deep for anything supporting a critical function.</p><p>You can’t build that register accurately without first knowing what ICT services are in use. Most organisations’ procurement records fall short here for a simple reason, they show what was formally approved, not what’s running day to day. Put the honest asset inventory before the Register itself in the documentation order. A Register built on an incomplete inventory just produces a well-formatted document that’s wrong.</p><h3>Third: The Incident Classification and Reporting Playbook</h3><p>DORA’s initial incident notification window is four hours from classifying an incident as major. NIS2’s early warning window is 24 hours from becoming aware. Neither leaves room to figure out a reporting process while an incident is underway.</p><p>Document the classification criteria (what counts as “major” for your organisation), the escalation path, who has authority to submit the report, and which template goes to which authority. If you’re dual-adjacent, a DORA entity with a non-financial group entity separately in scope for NIS2 — document that the two use different channels and forms, even where the timelines have been aligned. Assuming one incident report satisfies both regimes is a mistake I’ve seen cost a group real time mid-incident, precisely when there wasn’t any to spare.</p><p>Run this once as a tabletop exercise before it’s needed for real. The gap between having a policy and having tested it with the people who’d actually execute it is usually where the four-hour clock gets missed.</p><h3>Fourth: Management Body Sign-Off</h3><p>Both regimes attach personal accountability to the management body, not just to whoever’s doing the work day to day. Document that the board or leadership team has formally approved the ICT risk management framework, reviewed the classification decision, and receives incident reporting metrics on a defined cadence. This is the piece most often skipped at mid-market firms, usually because the security or compliance lead is quietly carrying the work without ever putting it in front of leadership for sign-off, exactly the gap a supervisor is trained to look for.</p><h3>Why This Order Matters</h3><p>Every item after the <a href="https://montro.io/blog/dora-ai-risk-classification-checklist">classification</a> memo depends on getting the classification right, and every item after the asset inventory depends on that inventory being honest rather than aspirational. Building the Register of Information, or an incident playbook, before either of those two things exist just means redoing the work later against the right foundation.</p><p>A subsidiary classified as NIS2-only turns out, mid-audit, to meet DORA’s entity criteria, and suddenly the risk framework, the register format, and the incident playbook all need rebuilding against DORA’s more prescriptive requirements, months after the first version shipped. The documentation wasn’t wrong on its own terms. The classification underneath it was, and everything inherited the error.</p><p>If you’re starting this from a blank page, this is where I’d put the next two weeks of work, in this sequence. Happy to go deeper on any one of these in the comments if it’s useful.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82d25cad96ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The EU AI Act Explained for Non-Lawyers]]></title>
            <link>https://medium.com/@MontroAI/the-eu-ai-act-explained-for-non-lawyers-e955129bb929?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/e955129bb929</guid>
            <category><![CDATA[european-union]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[regulation]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 15:48:57 GMT</pubDate>
            <atom:updated>2026-09-25T15:48:57.048Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>It has 113 Articles and 13 Annexes. You need about thirty of them. Here’s the version I wish someone had handed me when I started reading it for clients.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dFo4r5kSNElCm8hliBYZ2g.png" /></figure><p>I wrote this for the DPO whose CEO just forwarded them an article with the subject line “Are we ready for this?” I’m assuming you’ve read parts of the Act already, you’ve seen the Article numbers, and you’re tired of pieces that call it the world’s first comprehensive AI regulation for the tenth time without telling you anything you can act on.</p><p>Here’s the whole thing, compressed to what matters: four risk tiers, the deployer-versus-provider distinction, three enforcement dates, eight Annex III categories, the deployer obligations, Article 50’s transparency rules, the fine bands, and the supervisor architecture. For most mid-market firms, the Act comes down to four things: knowing which of your AI systems are high-risk, documenting them correctly, assigning human oversight, and being ready before August 2026.</p><h3>The four risk tiers</h3><p>The Act sorts every AI system into one of four tiers, and the first job for any team starting cold is figuring out which tier each of your systems falls into.</p><p><strong>Prohibited (Article 5).</strong> These are practices the Act simply bans outright: subliminal manipulation, social scoring, real-time remote biometric identification in public by law enforcement, untargeted facial-image scraping, workplace emotion recognition, biometric categorisation that infers protected characteristics, and predictive policing based solely on profiling. For most mid-market firms this tier is easy to clear: you either do one of these things or you don’t. The one trap I’d check specifically: some recruiting and customer-support AI tools include sentiment or emotion analysis that, depending on configuration, can land on the prohibited side of the workplace emotion-recognition ban. Worth checking every HR and customer-support AI tool in your inventory against that specific line.</p><p><strong>High-risk (Article 6 plus Annex III).</strong> This is where the substantive bulk of the Act lives, and where I’d spend most of your time. Two paths get a system here. Article 6(1), the safety-component test, catches AI that’s a safety component of a product already covered by Union harmonisation legislation (medical devices, machinery, toys, lifts) requiring third-party conformity assessment. Article 6(2), the use-case test, catches AI used in any of the eight Annex III categories, which I’ll walk through below. The August 2026 enforcement date applies primarily here, and for most mid-market firms, Act compliance is almost entirely a high-risk-tier problem.</p><p><strong>Limited-risk transparency (Article 50).</strong> This tier covers AI that interacts with people or generates content in ways where the person might not realise AI is involved: chatbots, emotion recognition and biometric categorisation where not prohibited, deepfake-style generated media, and AI-generated text on matters of public interest. The obligation is purely informational: tell the user they’re talking to AI, label synthetic content as such. But it applies to a far wider set of tools than most teams realise: every customer chatbot, every AI-generated article, every synthetic voice agent.</p><p><strong>Minimal-risk.</strong> Everything else: spam filters, recommendation engines, productivity-tool AI that doesn’t fall into any category above. No specific Act obligations here, though GDPR and other regimes still apply as normal.</p><p>In our audits, a typical mid-market firm’s AI footprint splits roughly: 0% prohibited, 5–15% high-risk (concentrated in HR, finance, and sector-specific use cases), 25–40% limited-risk transparency, the rest minimal-risk. I’ll focus the rest of this on high-risk and Article 50, since that’s where the documentary work concentrates.</p><h3>Are you a deployer or a provider?</h3><p><a href="https://montro.io/blog/eu-ai-act-provider-vs-deployer">The Act</a> draws a sharp line here, and your obligations depend entirely on which side you sit. A <strong>provider</strong> develops an AI system, or has one developed, and places it on the market under their own name — think OpenAI, Anthropic, Google, or the smaller specialised vendors. For a high-risk system, providers carry the heaviest obligations: conformity assessment, technical documentation, a risk management system, post-market monitoring, EU database registration. A <strong>deployer</strong> uses an AI system under their own authority. Most mid-market firms are deployers. You buy Phenom; Phenom is the provider, you’re the deployer. You use ChatGPT for a workflow; OpenAI is the provider, you’re the deployer.</p><p>Three things flip that line. Substantial modification, heavy fine-tuning or system integration that materially changes the intended purpose, or putting your logo on a third-party AI in a way that obscures the original provider, makes you a provider. White-labelling a third-party AI as your own product makes you a provider. And building an AI system in-house, even purely for internal use, makes you the provider for it.</p><p>That last one is the surprise I run into almost every time I work through this determination with a new client. They arrive expecting to be a pure deployer, they buy tools, they don’t build them. Then we review the actual inventory, and somewhere in engineering there’s a workflow automation scoring support tickets, or a content tool marketing built for outreach drafts, or a model HR uses for quarterly reviews. None of it was procured. All of it was built and deployed internally under the firm’s own authority. Under the Act, the firm is the provider for every one of those, not a deployer using someone else’s system, but the provider, carrying the full set: risk management system, technical documentation, conformity assessment, post-market monitoring, registration. Compliance didn’t know the tools existed. Legal didn’t know the firm was a provider for them. Engineering had no idea any of this was relevant to them at all.</p><p>Most mid-market firms are pure deployers for the bulk of their AI footprint, with one or two provider exposures hiding in internal projects. The deployer obligations are the starting point for the rest of what follows here, since they’re what apply to nearly everything in your inventory.</p><h3>Three enforcement dates, not one</h3><p>The single biggest planning error I see is treating this as a single August 2026 deadline. It isn’t. Three dates trigger different obligations.</p><p><strong>2 February 2025 is already in force.</strong> Article 5’s prohibited practices became a regulatory offence in themselves from that date. And Article 4’s AI literacy obligation started applying too, requiring providers and deployers to ensure staff dealing with AI systems have sufficient literacy for their role. I’d call this the most overlooked obligation in the whole Act. It’s not a documentation exercise, it’s a competence requirement, and most firms have done nothing on it because there’s no checkbox to tick. A supervisor will reasonably expect, on inspection, that a firm running AI can show the people running it know what they’re doing. A short structured programme, an induction module, role-specific training for high-risk AI users, a refresh cadence, is sufficient evidence in most cases I’ve seen.</p><p><strong>2 August 2026 is the main deadline.</strong> Most of the Act’s provisions apply from here: provider and deployer obligations for high-risk systems, general-purpose AI model obligations, Article 50 transparency, and the Article 99–101 penalty regime. If you have a high-risk system running on that date, the deployer obligations need to be satisfied as of that date. Bring one into service after, and you need them satisfied at deployment.</p><p><strong>2 August 2027 is the catch-up date.</strong> Existing high-risk systems already in operation before August 2026 get an extra year to come fully into compliance. The HR system that’s been running CV screening since 2022 doesn’t enforce in 2026, it enforces in 2027. But you still need to know, before 2026, which of your existing systems are high-risk, in order to know which date applies to them. Discovery and classification have to happen regardless of which deadline governs the paperwork behind it.</p><h3>Annex III, walked through</h3><p>Annex III sets out the eight <a href="https://montro.io/blog/eu-ai-act-risk-tiers-explained">high-risk</a> use-case categories under Article 6(2). Below is each, with the kind of mid-market tool I’d expect to find in it.</p><p><strong>Biometrics</strong>: facial recognition for office access, sentiment analysis routing customer complaints. <strong>Critical infrastructure</strong>: AI in utilities or transport operational technology, low exposure outside that sector. <strong>Education and vocational training</strong>: AI assessing student performance or admissions. <strong>Employment</strong>: recruiting platforms like Phenom, Eightfold, HireVue, and most modern AI-enabled hiring tools, plus AI-driven performance management and scheduling that affects pay. <strong>Essential services</strong>: credit scoring, insurance underwriting and risk pricing, benefits eligibility. <strong>Law enforcement</strong>, <strong>migration and border control</strong>, and <strong>administration of justice</strong>: all specific and rare for general mid-market deployment.</p><p>For most European mid-market firms, real exposure concentrates in employment, essential services (especially financial services and insurance), and biometrics where you run access control or customer-facing biometric systems. The other five categories matter enormously for the sectors they govern, but rarely touch general mid-market deployment. The mapping work itself, taking your inventory and assigning each system to “out of scope,” a specific Annex III category, or the Article 6(1) safety-component test, is the actual classification task everything else depends on.</p><h3>What deployer compliance requires</h3><p>Once you’ve identified a high-risk system, Chapter III’s deployer obligations apply, and they’re operational, not just documentary.</p><p>Use the system per the provider’s published instructions, which sounds trivial but isn’t; I’ve seen plenty of firms deploy AI in workflows the provider’s documentation never anticipated. Assign a specific named person to monitor the system’s operation, with real competence and authority to act on what they find, not just a name on a form. Where you control the input data, make sure it’s relevant and representative. For HR AI that means assessing the candidate data, job descriptions, and historical hiring decisions feeding it. Keep records of operation, monitor for unexpected behaviour, and report serious incidents to the provider and the relevant national authority. Inform workers before a high-risk system goes live in their workplace. Where you fall into specific categories, mainly public-law bodies and essential service operators, conduct a fundamental rights impact assessment under Article 27. Cooperate with the provider’s post-market monitoring by sharing operational and incident data. Register in the EU database if you’re a public authority. And keep the logs the system automatically generates, for the period national law requires, or six months by default.</p><p>The pattern across all of this is that it’s continuous, not point-in-time.</p><p><em>None of this gets satisfied by a documentation sprint the month before the deadline. It needs an operating model that runs from go-live onward, not a project that ends the day the register gets signed off.</em></p><h3>Article 50: the obligation that catches almost everyone</h3><p>This one sits in an odd spot: lighter than the high-risk regime, but far broader in what it touches. Every AI system that interacts directly with people has to make that fact known — every customer chatbot, every AI email autoresponder, every AI receptionist or scheduling assistant. Every AI generating synthetic audio, image, video, or text has to mark it as such: marketing’s AI image generators, sales’ AI-personalised outreach, legal’s AI-drafted documents. Emotion recognition and biometric categorisation systems have to inform the people exposed to them. And anything generating deepfake-style content has to label it as artificially generated.</p><p>For most mid-market firms this is mostly implementation work: configuring the chatbot to announce itself, updating email templates to disclose AI involvement, making sure AI-generated marketing content carries the right metadata. It’s less burdensome than the high-risk regime, but it touches more tools, and in my experience it’s the compliance gap most teams find last, because it sits entirely outside the high-risk classification process.</p><h3>The penalties, and how enforcement really works</h3><p>Article 99 sets three fine bands: up to €35 million or 7% of global turnover for Article 5 prohibited-practice violations; up to €15 million or 3% for most other obligations (provider, deployer, conformity assessment, transparency); and up to €7.5 million or 1.5% for supplying incorrect or misleading information to authorities. These are sized to be felt, and the architecture will look familiar to anyone who’s lived through GDPR enforcement, though the upper band here runs materially higher. One thing I always tell an executive committee that hasn’t been through a GDPR cycle: European regulatory fines land where the supervisor decides, not on a fixed schedule. Most firms won’t be fined. The ones that are will be fined heavily. The regime exists to set examples and shape industry practice, not to penalise every infraction uniformly.</p><p>Enforcement runs through three layers, and I’ve found most firms only expect to deal with one of them. The AI Office, inside the European Commission’s DG CNECT, coordinates at Union level, general-purpose AI oversight, the EU database, harmonised standards. National competent authorities, designated separately by each member state, handle most operational supervision. Some are the existing DPA with an extended remit; some are new dedicated AI authorities; some are sectoral regulators. And sectoral regulators keep enforcing what they already enforce while folding in AI risk explicitly; BaFin and ACPR have signalled DORA enforcement will incorporate AI risk in financial services. For a European mid-market firm, that often means engaging with all three layers at once: the DPA for GDPR and possibly AI Act exposure, the sector regulator for sector-specific AI risk, and occasionally the AI Office directly for general-purpose AI or cross-border questions.</p><h3>If you want to read the Act yourself</h3><p>The sections worth reading in full: Article 3 (definitions), Article 5 (prohibited practices), Article 6 (high-risk classification), Article 26 (deployer obligations), Article 27 (fundamental rights impact assessment), Article 50 (transparency), Article 99 (penalties), and Annex III itself (short and operationally critical). Skim the provider chapters (Articles 9–24) unless you’re a provider, the general-purpose AI provisions (51–55) unless you build on or distribute GPAI models, and the supervisory architecture (60–98), useful context, not much you’ll act on directly. I’d trust the consolidated text on EUR-Lex, the Commission’s own AI Act summary, and your member state’s national competent authority guidance over anything else you’ll find. It’s more readable than its reputation suggests; a focused two hours on the sections above gets you a working knowledge.</p><p>A few things worth being explicit about, since they trip people up consistently.</p><p>The Act reaches further than most people assume. It applies to providers placing systems on the EU market from anywhere, to deployers established in the EU, and to providers and deployers outside the EU whenever their AI’s output gets used in the EU. A US firm whose AI output reaches EU customers is in scope for at least some of this.</p><p>Open-source AI is mostly carved out at the model level, but the carve-out narrows fast: fine-tuning an open model for an internal use case still attracts deployer obligations if the resulting system turns out high-risk.</p><p>General-purpose models like GPT-5 or Claude are treated as a separate category entirely. The provider obligations there, training data transparency, systemic risk assessment for the largest models, sit with OpenAI, Anthropic, and the other foundation labs, not with you as a deployer calling their API. Your obligations sit at the system you build with it, not the model underneath.</p><p>And don’t wait for the harmonised standards to be finalised before starting the work. The obligations enforce on the dates above regardless of whether the standards exist yet. When they do land, they’ll offer a presumption of conformity for specific obligations, not create the underlying obligation itself.</p><p>If there’s one thing I’d tell a CEO who just forwarded this piece to their DPO: get an AI inventory started, today, if you don’t already have one. Nothing above can happen without it, no classification, no prioritisation, nothing. The discovery work itself takes about thirty days at most mid-market firms. Everything else is downstream of that one step.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e955129bb929" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Employees Bypass AI Governance Policies — and What It Costs]]></title>
            <link>https://medium.com/@MontroAI/why-employees-bypass-ai-governance-policies-and-what-it-costs-9f0d6c81755c?source=rss-10dd02c21d1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f0d6c81755c</guid>
            <category><![CDATA[regulatory-compliance]]></category>
            <category><![CDATA[regulation]]></category>
            <category><![CDATA[eu-ai-act]]></category>
            <category><![CDATA[gdpr]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <dc:creator><![CDATA[Montro]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 14:19:45 GMT</pubDate>
            <atom:updated>2026-09-23T14:19:45.997Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Shadow AI isn’t a carelessness problem. It’s what happens when caution becomes more expensive than avoidance, and employees do the maths correctly.</strong></p><figure><img alt="Montro cover graphic on a dark navy background. A balance scale tips downward to the right. The higher left pan, in teal, is labelled “the official path — too costly to engage: submit a request, wait for review, risk a flat no.” The lower right pan, in cerise, is labelled “the workaround — wins the decision: sign up quietly, use it today, no visible consequence.” Headline reads “Employees aren’t careless. They’re doing the maths." src="https://cdn-images-1.medium.com/max/1024/1*Cv5OrsVjNHWq2o02qKkMzw.png" /></figure><p>I’ve come to think shadow AI gets misdiagnosed constantly. It reads as a technology problem when it’s really a behaviour problem, and the behaviour, once you look at it closely, is rational.</p><p>When employees adopt AI tools outside the official governance process, they’re usually not acting carelessly. They’ve assessed the situation and reached a conclusion: engaging with governance carries real costs, time, uncertainty, the risk of a flat refusal, and those costs outweigh the risk of quietly using the tool instead. Avoidance is the logical output of a system that’s made caution feel expensive. This is the part of the conversation I think gets skipped over. Governance programmes spend most of their energy on what employees <em>should</em> do. They spend almost none on what the system is telling employees through how it works, and what behaviour that produces in return.</p><h3>The calculation employees are making</h3><p>Picture the choice facing someone who wants to use an AI tool for a legitimate task. They can go through the official process: submit a request, wait for IT review, risk a no, explain to their manager why the tool is delayed. Or they can sign up quietly, use it, and keep the exposure low.</p><p>The second path is faster, less uncertain, and carries no immediate visible consequence. The governance risk is abstract and sits in the future. The productivity benefit is concrete and sits right now. Every time the official process runs slow, stays opaque, or ends in a refusal with no clear reasoning attached, the calculation tips a little further toward avoidance.</p><p>A Fishbowl survey of over 11,700 professionals found that 70% of those using generative AI at work hadn’t disclosed that usage to their employer. I don’t read that figure as a measure of dishonesty. I read it as a measure of how many people worked out, correctly, that non-disclosure was the lower-friction path. The process was available to them. They chose not to use it.</p><p>That’s the uncomfortable part for CISOs and IT Directors to sit with: shadow AI usually isn’t evidence that employees don’t care about governance. More often, it’s evidence that the governance system they encountered taught them avoidance was the smarter move.</p><h3>How caution becomes the whole team’s default</h3><p>This pattern doesn’t stay with one person. Once a few team members find that the official process is slow or opaque, the knowledge spreads informally. New hires get shown the official process, and then quietly shown how things really get done. Nobody documents the workaround, but it travels anyway, through onboarding conversations, Slack messages, and the simple observation that nobody who took the unofficial route ever faced a consequence.</p><p>Over time, the team’s real default shifts. The governance process is still technically there. It’s just not the path anyone takes unless someone specifically asks them to demonstrate compliance. AI tool decisions start happening informally, in channels the governance system can’t see, at a speed the official process was never built to match.</p><p>This is where the real enterprise risk shows up structurally. The programme exists on paper, while the actual adoption decisions happen somewhere else entirely. What the register shows and what teams are running have quietly diverged, and that gap <em>is</em> the unpriced risk. It widens every week, usually with no obvious signal that anything’s wrong.</p><h3>What the governance system is communicating</h3><p>Every governance programme communicates something well beyond its written rules — what the organisation genuinely values, and what it’s willing to tolerate. A slow system tells people that process matters more than speed. A system that keeps refusing tools without explanation tells people the default answer is always no. A system that’s hard to find or navigate tells people the organisation never really expected them to use it.</p><p>Employees read those signals accurately, every time. When the process signals that caution and avoidance are the smart defaults, that’s precisely what it gets. I’d argue the resulting shadow AI has little to do with employee judgement failing. The system is producing exactly the behaviour it was built to produce, even where nobody intended that outcome. This is the hardest thing for a security or governance team to sit with: a programme generating widespread avoidance isn’t being ignored. It’s working exactly as designed, just pointed in a direction nobody meant.</p><h3>What avoidance costs</h3><p>The immediate consequence of routing around governance is a shadow AI estate growing outside the security team’s visibility. But the downstream cost is more specific than that, and it’s the part I think most CISOs underweight.</p><p><a href="https://montro.io/blog/gdpr-documentation-for-ai-beyond-ropa">GDPR</a> Article 30 requires a record of every processing activity involving personal data. The tools employees use informally, the ChatGPT sessions drafting customer replies, the AI writing tools processing internal documents, the productivity platforms taking personal data as input, are all processing activities that belong in that record. Nobody concealed them. The process that would have surfaced them simply never got engaged in the first place, so they were never declared at all. The RoPA’s gaps trace back to that missing pipeline, not to anything the DPO failed to do.</p><p><a href="https://montro.io/blog/eu-ai-act-risk-tiers-explained">EU AI Act</a> Article 26 requires deployer oversight of AI systems in use, human oversight, logging, use consistent with the intended purpose. None of that can happen for tools the organisation doesn’t know it’s running. Where avoidance is widespread, the deployer obligations that apply from August 2026 go unmet, and the reason isn’t that the organisation chose to ignore them. It’s that the process meant to surface those tools cost employees more than it was worth to engage with.</p><p>Here’s the part that should worry a compliance team specifically, and the one I raise first whenever I’m in this conversation with a DPO: a supervisory authority reviewing an incomplete Article 30 record doesn’t distinguish between a tool that was deliberately hidden and one that was simply never declared because the process discouraged declaration. The gap in the record looks identical either way.</p><h3>What changes when governance stops making avoidance the smart choice</h3><p>A programme that produces genuine compliance, rather than avoidance, makes engaging with it easier than avoiding it. That sounds obvious written down. In my experience it’s harder to build than it sounds, because most governance programmes were designed around the compliance requirement, not around the employee’s actual friction calculation, which is a different starting point entirely.</p><p>An AI tool request that gets a response inside 48 hours, even a conditional approval with specific data handling conditions attached, is a process people will use again. One that takes three weeks and ends in an unexplained no, they won’t touch twice. The speed and clarity of the answer matters nearly as much as the answer itself, because it’s what tells an employee whether the official path is worth trying next time.</p><p>Reasoning matters just as much as speed. Boundaries that come with a visible explanation get respected far more than rules that don’t. When someone understands <em>why</em> a tool got declined, no adequate DPA, no AI Act conformity declaration, data routed through servers outside the EEA, they can apply that same reasoning to the next tool themselves. And proportionality matters more than most programmes admit: demanding heavy documentation for a genuinely low-risk tool signals that the process isn’t calibrated to reality at all, and that erodes trust faster than slowness ever does. Automation that embeds these checks directly into the adoption workflow removes the friction right at the point of decision, instead of leaving it sitting in front of an employee who might quietly decide it isn’t worth the effort.</p><p>IBM’s 2024 Cost of a Data Breach Report found shadow data in unmanaged sources was a factor in 35% of breaches. Shadow AI sits squarely in that category. The reason isn’t that any organisation decided to leave it ungoverned; the process meant to bring it into view was simply harder to use than the workaround. Lowering that cost of engagement is more than a soft cultural goal. It’s what decides whether a programme reflects the actual AI estate, or only the slice of it that found the official process navigable.</p><p>The honest way to check how wide that gap really is, and the method I trust over anything self-reported: don’t ask employees to declare it, since that just reproduces the same undercounting the governance process already suffers from. Look instead at signals that don’t depend on anyone volunteering anything — SSO logs showing tools connected without IT ever reviewing them, expense data surfacing SaaS subscriptions nowhere on the sanctioned list, email metadata revealing tools talking to work accounts regardless of how they got adopted. The distance between what those signals show and what the official register contains is a direct measurement of how much avoidance is happening.</p><p><em>The question was never whether teams are using AI tools outside the process. They already are. The real question is whether the programme is built to surface that fact, or built in a way that guarantees it stays hidden.</em></p><p>I’d put it plainly: a programme producing avoidance has already decided what the AI estate looks like. It just hasn’t looked at the result yet.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f0d6c81755c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>