<?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 Revanth on Medium]]></title>
        <description><![CDATA[Stories by Revanth on Medium]]></description>
        <link>https://medium.com/@jprevanth?source=rss-149486c4464------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*723ZpS5B2EeZkPwiqbcncQ.png</url>
            <title>Stories by Revanth on Medium</title>
            <link>https://medium.com/@jprevanth?source=rss-149486c4464------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 07:48:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jprevanth/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 Deployment Paradox: Why Solving AI Requires More Humans, Not Fewer]]></title>
            <link>https://medium.com/@jprevanth/the-ai-deployment-paradox-why-solving-ai-requires-more-humans-not-fewer-78a377dab00b?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/78a377dab00b</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[data-integration]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Wed, 12 Aug 2026 14:50:29 GMT</pubDate>
            <atom:updated>2026-08-12T14:50:29.565Z</atom:updated>
            <content:encoded><![CDATA[<p>A $20M startup just launched to fix something nobody wants to admit is broken.</p><p>The company is called June. It was founded by Efrat Rapoport, a former Salesforce executive. It’s backed by Marc Benioff, Michael Dell, and Aaron Levie. And its entire premise is this: getting AI to actually work inside large enterprises is so hard that you need a dedicated team of specialists just to make it happen.</p><p>They call them “forward-deployed engineers.” People who fly in, set up your AI systems, fix the gaps, and move on.</p><p>This is where we are in 2026.</p><p><strong>We built the most powerful technology in history. And our best deployment strategy is… more consultants.</strong></p><p>There’s no judgment in that. It’s just an interesting mirror to hold up.</p><h3><strong>Why is enterprise AI so hard to deploy?</strong></h3><p>Ask ten engineers and you’ll get ten different answers. Context windows. Prompt engineering. Model selection. Governance and compliance. Change management.</p><p>But spend enough time in the weeds and a pattern emerges. Almost every AI deployment problem, eventually, traces back to data.</p><p>The model is fine. The infrastructure is fine. The use case is valid.</p><p>But the data feeding the system? Scattered across a dozen SaaS tools. Inconsistently structured. Stale by 24 hours. Governed by three different teams who don’t fully agree on definitions.</p><p>So the agent hallucinates. Or returns the wrong number. Or works perfectly in the demo and fails in production.</p><p>And the answer, apparently, is to fly someone in to fix it manually.</p><h3><strong>The same week June launched, a few other things happened</strong></h3><p>IBM announced vector search built directly into Netezza — their data warehouse platform. Not a separate tool. Not a bolt-on. In-database. The argument being that if AI needs vector retrieval, the place to do that retrieval should be where the data already lives.</p><p>Boomi got ranked #1 in the ISG Buyers Guide for application integration. A category that, let’s be honest, exists because enterprise data is still too siloed to work with as-is.</p><p>And an analysis from Streamline Feed flagged traditional ETL pipelines as “losing strategic value” — not because ETL is dead, but because the discipline of manually building and maintaining pipelines can’t keep up with the pace of AI adoption.</p><p>Three separate stories. One thread: the data layer is the bottleneck.</p><h3><strong>The uncomfortable truth about AI ROI</strong></h3><p>There’s a stat from earlier this year that stuck with me. Nearly half of enterprises spend 76–100% of their entire AI budget on inference alone. And the most cited reason agents stall in production? Cost.</p><p>But inference cost is a symptom, not a cause. Agents are expensive partly because they’re compensating for a messy data layer. They query the same fragmented sources repeatedly. They resolve inconsistencies that should have been resolved upstream. They do work that a clean, connected data foundation would have done already.</p><p>Forward-deployed engineers are doing the same thing at the human layer. They’re compensating for what the data layer didn’t do.</p><h3><strong>What “data readiness” actually means</strong></h3><p>It’s not a new term. But I think it’s finally getting the weight it deserves.</p><p>Data readiness doesn’t mean your data is perfect. It means your data is connected, governed, and trusted enough that the system working with it — whether that’s a human analyst, a BI tool, or an AI agent — doesn’t have to start every interaction by questioning whether what it’s seeing is real.</p><p>That’s it. That’s the bar.</p><p>And most enterprise stacks aren’t there yet. Not because the people building them aren’t smart. But because data infrastructure has historically been treated as a cost center, not a strategic asset. Something you maintain, not something you invest in.</p><p>AI is changing that calculus fast.</p><h3><strong>The companies that win won’t need the cavalry</strong></h3><p>Here’s where I land on all of this.</p><p>The forward-deployed engineer model will work. June will probably do well. Enterprises need help, and there’s real value in hands-on implementation expertise.</p><p>But it’s a workaround, not a solution.</p><p>The companies that pull ahead in the next two years won’t be the ones who hire the most specialists. They’ll be the ones who built a data foundation robust enough that their AI systems can actually use it without hand-holding.</p><p>Data readiness isn’t a project you complete. It’s an operating principle you commit to.</p><p>The ones who get that early are the ones who won’t need to call for backup.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78a377dab00b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI control gap is not a governance problem. It is a data-access problem.]]></title>
            <link>https://medium.com/@jprevanth/the-ai-control-gap-is-not-a-governance-problem-it-is-a-data-access-problem-b9de6eac71c5?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9de6eac71c5</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Thu, 16 Jul 2026 15:40:35 GMT</pubDate>
            <atom:updated>2026-07-16T15:40:35.492Z</atom:updated>
            <content:encoded><![CDATA[<p>Everyone read the IBM study as a call for more policy. The two-thirds of CIOs accountable for systems they cannot control are not short on policy. They are short on a map of where their data actually goes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*0s11PLXbsxfMcLwTV00Iuw@2x.jpeg" /></figure><p>In June, IBM’s Institute for Business Value surveyed 2,000 CIOs and CTOs and found that two-thirds are held accountable for AI systems they do not fully control. Seventy percent said business teams are deploying technology faster than IT can track. Only 11% felt ready for the volume of AI agents coming in the next year, even though 80% are under a CEO mandate to scale faster. Foundry’s 2026 AI Priorities study put the same tension differently: 97% of IT decision-makers hit obstacles deploying AI, and infrastructure and systems integration is the single biggest one. And Integrate.io’s 2026 data found that 95% of IT leaders name integration as their primary AI adoption barrier, with only 28% of enterprise applications actually connected.</p><p>The consensus reading arrived within a day. This is a governance gap. Write the policy, stand up the review board, mandate the guardrails, and control follows. IBM’s own framing leaned this way: redesign how you govern and invest, embed control from the start.</p><p>That reading is comfortable and wrong. You cannot govern a system you cannot see, and IT cannot see these systems because the data those systems reach was never routed through anything IT built. The control gap is downstream of an access gap. Fix the access problem and governance becomes possible. Skip it and every policy you write is a memo about a room you have no key to.</p><p>Why the governance framing fails the AI control gap</p><p>Governance assumes a chokepoint. A policy works because requests flow through a place where the policy can be enforced – a gateway, a warehouse, a permission layer. Traditional shadow IT had one. As IDC’s analysts put it in CIO, “traditional shadow IT introduced unmanaged software. Shadow AI introduces unmanaged judgment.” The difference matters more than it sounds. Unmanaged software sat somewhere you could eventually find. Unmanaged judgment runs live, inside a SaaS copilot or an agent framework, reaching whatever data the vendor wired it to reach.</p><p>IDC found only about 16% of organizations have a unified AI governance model. The other 84% are not lazy. They are trying to govern access paths that terminate outside their control – an embedded assistant in a CRM, an agent invoking an API, a copilot reading a folder no one inventoried. A policy cannot sit in front of a path IT never built. That is the mechanism the governance framing skips.</p><p>The real mechanism behind the AI control gap</p><p>Picture the actual failure. A finance team turns on an AI assistant inside their billing tool. It answers questions well, so someone points it at the data warehouse through a connector the vendor ships. Now an agent is reading governed financial data through a path that shows up in no one’s architecture diagram. It works, so it spreads. Six months later the CIO is accountable for an agent making judgment calls on revenue data, and there is no single place to see what it touched or why.</p><p>This is not hypothetical drift. IBM found organizations averaged 54 AI agent incidents last year that needed human correction. Of the high-severity ones, 37% caused data exposure or a security breach and 33% triggered cascading system failures. Those are not policy failures. They are access failures – agents reaching data through routes that had no governance layer because no one designed the route in the first place. Greyhound Research’s chief analyst named the structural version of this: enterprises have decentralised experimentation far faster than they have decentralised accountability.</p><p>Why more governance policy will not close it</p><p>The obvious fix is to buy a governance platform and write the framework. It will not work, for the same reason a speed limit does not work on roads you have not mapped. Every governance tool needs a surface to enforce against. If your agents reach data through fifty vendor-specific connectors, each with its own auth model and its own idea of what “read access” means, your governance layer has fifty surfaces to chase and no authority over any of them. You end up with a policy document and a growing list of paths it does not cover.</p><p>The Foundry and Integrate.io numbers already tell you this. When 95% of leaders call integration the primary barrier and only 28% of applications are connected, the missing thing is not a rulebook. It is a single, governed path between the data and the things that consume it. Governance is what you do once that path exists. It is not a substitute for building it.</p><p>What actually closes the AI control gap</p><p>Control comes back when access is consolidated, not when policy is multiplied. The organizations that scale AI without losing the plot route agent and analytics access through one governed layer – a place where the data lands, where permissions are enforced once, and where every read is logged regardless of which agent or tool made it. When agents reach data through that layer instead of through scattered point-to-point connectors, the CIO gets the one thing the IBM study says they lack: visibility. You cannot be accountable for what you cannot see, and you can only see what runs through a place you own.</p><p>This is the problem I spend my days on at Peliqan. What I have watched happen, repeatedly, is that teams treat AI access as a governance project when it is first an integration project. Once the data sits in one governed warehouse and agents read from that instead of from a dozen live source systems, the governance question stops being a chase and starts being a setting. The order matters. Foundation first, then visibility, then policy – not the other way around.</p><p>The IBM study will be quoted for the rest of the year as proof that enterprises need better AI governance. Read it again. Two-thirds of technology leaders are accountable for systems they cannot see, and you cannot govern your way to sight. You have to build the path first. The control gap closes the day IT owns the road again – not the day it writes another rule for a road someone else paved.</p><p>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along – I publish a few times a week.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9de6eac71c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The HTTP QUERY method is here. It matters more for AI agents than for your REST API]]></title>
            <link>https://medium.com/@jprevanth/the-http-query-method-is-here-it-matters-more-for-ai-agents-than-for-your-rest-api-da4fd6b0f6f2?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/da4fd6b0f6f2</guid>
            <category><![CDATA[api-design]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[http-query-method]]></category>
            <category><![CDATA[https]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Sat, 04 Jul 2026 03:38:32 GMT</pubDate>
            <atom:updated>2026-07-04T03:38:32.979Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Everyone is reading RFC 10008 as a long-overdue fix for a 16-year-old annoyance. They are missing where the demand is actually going to come from.</em></p><p>In June 2026, the IETF quietly published <a href="https://www.rfc-editor.org/rfc/rfc10008.html">RFC 10008</a>, “The HTTP QUERY Method.” It is the first new method added to core HTTP since <a href="https://www.rfc-editor.org/rfc/rfc5789">PATCH landed in 2010</a>. The authors are not fringe hobbyists: Julian Reschke of greenbytes, James Snell of Cloudflare, and Mike Bishop of Akamai, which means two of the world’s largest CDNs helped write it. For a protocol that most engineers treat as frozen, 16 years of silence being broken is a genuine event. The trade press has covered it the way it covers most standards news, as plumbing. That framing is not wrong. It is just aimed at the wrong decade.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uz9E-S-q3JWDYkl5UUTPrg.png" /></figure><h3>What the HTTP QUERY method actually is</h3><p>Start with the two verbs everyone already knows. GET means “give me this thing.” It is read-only, you can repeat it safely, and its responses can be cached, but everything you want to ask has to fit in the URL. POST means “here, process this.” It can carry a large body of structured data, but HTTP assumes it might change something on the server, so it is treated as neither safe nor cacheable.</p><p>The HTTP QUERY method is the missing third option: a read like GET, but with a request body like POST. You send your query in the body as JSON, SQL, or any other format, and the protocol still guarantees the operation is <a href="https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Methods">safe, idempotent, and cacheable</a>. Think of it this way. GET is shouting a catalog number across a library. POST is handing over a form that might be a request or might be a change, and the librarian cannot tell which. QUERY is handing over a written lookup slip that everyone understands is only a lookup, so the answer can be photocopied for the next person who asks the same thing, and safely redone if the slip gets lost.</p><p>In practice a request looks like this: a QUERY /contacts line, a Content-Type header describing the body, and the actual filter in the body instead of jammed into the query string. A success comes back as a plain 200 OK with results. That is the whole idea.</p><h3>The problem the HTTP QUERY method solves</h3><p>For years, developers had two bad options for search endpoints. The first was sending a body with a GET request, which HTTP leaves undefined and which proxies and load balancers will happily strip in transit. The second, and the one nearly every REST API eventually reaches for, was POST /search. But POST-for-reads is a lie told to the infrastructure. Because POST signals a possible state change, nothing in the request path can automatically retry it after a dropped connection, and no cache will serve it. You end up hand-building idempotency keys and retry logic for what is fundamentally a read. QUERY exists because complex filters, GraphQL queries, and long lists of parameters routinely blow past URL length limits, and the honest fix is a method that says &quot;this is a safe read&quot; while still carrying a body.</p><p>None of this is new as an idea. WebDAV shipped a SEARCH method back in <a href="https://www.rfc-editor.org/rfc/rfc5323.html">RFC 5323 in 2008</a>. It never caught on, because intermediaries, the proxies and gateways and firewalls between client and server, tended to choke on methods they did not recognize. That history matters for what comes next.</p><h3>Why the consensus reading of the HTTP QUERY method is too small</h3><p>The standard take is “finally, REST search is fixed.” True, and underwhelming. If the only beneficiaries were REST APIs that have limped along on POST /search for a decade, this would be a nice-to-have that most teams never bother adopting, exactly like SEARCH before it.</p><p>Here is the part almost no one covering RFC 10008 is connecting. The single largest new source of read-heavy, body-carrying HTTP traffic on the internet right now is not human-facing REST APIs. It is AI agents. Every retrieval call an agent makes, every RAG filter, every structured query against a tool or a data source, is a large, nested, read-only payload that overflows a URL and today rides on POST. Which means every one of those calls loses caching and loses safe automatic retry, at exactly the volume where both matter most.</p><p>Now look at where the agent ecosystem is heading. The <a href="https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/">Model Context Protocol</a>, the dominant standard for how agents call tools and data over HTTP, currently sends effectively everything over POST. And its newer drafts are adding cache-control-style fields and routing headers so that gateways can tell a read from a write without cracking open the body. Read that again. MCP is independently reinventing safe, cacheable, routable read semantics on top of POST, using custom headers, at the same moment the IETF shipped a method that provides exactly those semantics natively. Two ecosystems are converging on the identical need from opposite directions, and as of today, neither has formally acknowledged the other. That is the story. Not the plumbing. The timing.</p><h3>Why the obvious win will not arrive on schedule</h3><p>I want to be honest about the friction, because the same trap that killed SEARCH is still set. Browser fetch() cannot send QUERY yet. HTML forms cannot use it. Most CDNs and proxies do not cache it, and plenty of legacy middleboxes will reject any method they do not recognize outright. The tooling that does support it is real but early: <a href="https://vensas.de/en/blog/http-query-method-dotnet-10">.NET 10 has first-class support</a>, Node.js can send it, and a few API clients have added it. That is a server-to-server story, not a browser story.</p><p>And the caching that makes QUERY attractive is genuinely harder than GET caching, because the cache key has to incorporate the request body, which means a cache has to read the entire body before it can even check for a hit. The <a href="https://news.ycombinator.com/item?id=48640974">developer community has been vocal</a> about this, and they are right to be. Standards adoption and standards publication run on completely different clocks. PATCH took years to become routine. QUERY will not be different just because we want it to be.</p><h3>What actually decides whether the HTTP QUERY method wins</h3><p>Adoption will not be driven by REST nostalgia. It will be driven by traffic economics. The moment a major CDN ships QUERY-aware caching, and watch the two that co-authored the RFC first, agent operators paying for millions of near-identical retrieval calls will have a direct cost-and-latency reason to move off POST. That incentive did not exist for the human web. It exists now.</p><p>The practical move for anyone building data or agent infrastructure is not to rip anything out today. It is to design your read paths to be semantically safe, idempotent, and cacheable now, using explicit headers and content-based cache keys, so that adopting QUERY later is a drop-in rather than a redesign. In our own work running an <a href="https://peliqan.io/">MCP endpoint over enterprise data</a>, the calls that hurt are always the same: high-volume, read-only, body-heavy queries that the transport treats as risky writes. QUERY is the primitive that fixes that category. The teams that quietly prepare for it will look prescient in two years.</p><p>QUERY did not arrive 16 years late. It arrived right on time, just for a different web than the one that first asked for it.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da4fd6b0f6f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MCP won the standards war. MCP governance is the war you are about to lose.]]></title>
            <link>https://medium.com/@jprevanth/mcp-won-the-standards-war-mcp-governance-is-the-war-you-are-about-to-lose-b6026c0e1f17?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6026c0e1f17</guid>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[mcps]]></category>
            <category><![CDATA[model-context-protocol]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Mon, 29 Jun 2026 05:13:32 GMT</pubDate>
            <atom:updated>2026-06-29T05:13:32.555Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The protocol question is settled. The expensive question — who controls the credentials, the audit log, and the API bill — is the one almost nobody is answering.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VMNcVCI8ahhowtPmcDyd1w.png" /></figure><p>In about eighteen months, the Model Context Protocol went from an <a href="https://www.google.com/url?q=https://www.anthropic.com/news/model-context-protocol&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">Anthropic side project</a> to the closest thing AI has to a universal port. OpenAI adopted it in March 2025, Google in April, Microsoft baked it into Azure and Windows, and in December 2025 Anthropic <a href="https://www.google.com/url?q=https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">donated MCP to the Linux Foundation</a>, putting it beyond any single vendor’s control. By its first birthday it was running on more than 10,000 active public servers. As standards wars go, this one is over.</p><p>So the industry has moved on to celebrating adoption. The number of MCP servers, the SDK download counts, the list of vendors who “support” it — this is the scoreboard everyone is watching.</p><p>It is the wrong scoreboard. MCP winning the standards war was never the hard part. The hard part is what happens after you plug it in, and on that question the same body that just blessed the protocol is flashing red. Gartner expects <a href="https://www.google.com/url?q=https://www.gartner.com/en/newsroom&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">over 40% of agentic AI projects to be cancelled by the end of 2027</a> — on escalating cost, unclear value, or inadequate risk controls. None of those three failure modes is a protocol problem. All three are governance problems.</p><h3>Why “should we adopt MCP” is the wrong question</h3><p>The reason MCP spread so fast is also the reason it is dangerous in production: it was built for interoperability first, not security. Security researcher Simon Willison’s <a href="https://www.google.com/url?q=https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">“lethal trifecta”</a> names the trap precisely — an agent with access to private data, exposure to untrusted content, and the ability to communicate externally can be turned into an exfiltration tool by a single poisoned input. His observation is that MCP agents almost always have all three at once.</p><p>This is not a whiteboard risk. The MCPTox benchmark, built on 45 live MCP servers and 353 real tools, <a href="https://www.google.com/url?q=https://arxiv.org/abs/2508.14925&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">successfully poisoned OpenAI’s o1-mini 72.8% of the time</a>, with several leading models above 60%. The most resistant model refused fewer than 3% of attacks. A month earlier, Invariant Labs had shown a <a href="https://www.google.com/url?q=https://invariantlabs.ai/blog/mcp-github-vulnerability&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">working exploit against the GitHub MCP server</a> where a malicious public issue coerced an agent into leaking private repository data — and were explicit that it was an architectural issue, not a bug to be patched. Picking a smarter model does not save you. Governing the assembly does.</p><h3>The cost trap a pilot will never show you</h3><p>The cancellations Gartner predicts will not all be dramatic security incidents. Most will be quiet budget deaths, and they start with a pricing unit nobody modeled.</p><p>An AI agent does not query like a person. It chains dozens of tool calls to finish one task and can fire thousands in a day. That collides badly with consumption pricing: <a href="https://www.google.com/url?q=https://zapier.com/pricing&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">Zapier counts a single MCP tool call as two tasks</a> and <a href="https://www.google.com/url?q=https://www.make.com/en/pricing&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">Make bills per operation</a>. At pilot volume the meter looks cheap. At production volume it does not — roughly 100,000 tasks a month already runs near $500, and a million climbs well into the thousands.</p><p>Underneath that sits a second meter most teams forget. When an agent reads live source APIs directly, it burns the source’s quota too — Intuit now meters QuickBooks Online reads, and Salesforce caps Enterprise orgs near 100,000 API calls a day before it starts returning errors. “Who absorbs the API cost?” is a procurement question, not a footnote.</p><p>And then there is the failure mode with the biggest number attached. IBM’s 2025 report puts the average breach at <a href="https://www.google.com/url?q=https://www.ibm.com/reports/data-breach&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">$4.44M, with ungoverned “shadow AI” adding about $670,000</a> — and unmanaged MCP is a shadow-AI factory. A team running ten agents with five connections each is already juggling more than fifty credentials with no central inventory. Governed access is a cost control, not just a security one.</p><h3>Why more servers is the opposite of the fix</h3><p>The instinct, when an agent lacks context, is to stand up another MCP server. Need accounting data? Add the accounting server. Need the CRM? Add another. Each one arrives with its own authentication, its own governance model, its own audit log — or no audit log at all. This is how you get credential sprawl and “shadow AI”: dozens of ungoverned agent-to-tool connections that no one can see across.</p><p>It is also why Gartner now ties a rising GenAI incident curve directly to MCP adoption, forecasting that <a href="https://www.google.com/url?q=https://www.gartner.com/en/newsroom&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">25% of enterprise GenAI apps will hit five or more minor security incidents a year by 2028</a>. Every server you add to “solve” the context problem widens the surface you have to govern. More plumbing is not more control. It is more attack surface wearing the costume of progress.</p><p>For European teams the math is heavier still. The EU AI Act allows fines up to EUR 35M or 7% of global turnover, GDPR up to EUR 20M or 4%, and DORA — in force since January 2025 — now names its first <a href="https://www.google.com/url?q=https://www.eba.europa.eu/&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">19 Critical ICT third-party providers</a>, a list dominated by US hyperscalers. Where your data is processed, and whether each connection is governed, is a financial question with a regulator attached.</p><h3>What actually closes the MCP governance gap</h3><p>If the protocol is settled and the danger is in the assembly, then the work is architectural, not aspirational. The pattern that survives scale has four properties: a single governed endpoint instead of a dozen ad-hoc servers, central credential custody so the AI client only ever holds a scoped key it can revoke in one click, a cached layer the agent queries instead of hammering live production APIs, and an immutable audit trail that records every read and write.</p><p>This is the part I see up close. At <a href="https://www.google.com/url?q=https://peliqan.io/&amp;source=gmail&amp;ust=1782796263577000&amp;sa=E">Peliqan</a> we built our MCP layer around exactly this gap — one EU-hosted endpoint over a warehouse the connectors sync into, so agents query a cached copy rather than your Salesforce or QuickBooks quota, writes are logged with the originating prompt and authorizing user, and the underlying credentials never leave the platform. I am not claiming that is the only valid answer. A Microsoft Fabric or Snowflake estate may be better served by its native server, and a pure long-tail automation need may fit a workflow tool. The point is the question. You are no longer choosing whether to adopt MCP. You are choosing how to govern it — and that choice is the entire bet.</p><p>MCP being a settled standard does not make your agentic project safe, cheap, or compliant. It just means the easy decision is behind you, and the one that actually decides whether you join the 40% that get cancelled is still sitting on the table.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6026c0e1f17" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI observability gap won’t be closed by more observability]]></title>
            <link>https://medium.com/@jprevanth/the-ai-observability-gap-wont-be-closed-by-more-observability-ca6d72aa6875?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca6d72aa6875</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-integration]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 13:28:06 GMT</pubDate>
            <atom:updated>2026-06-09T13:28:06.827Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Three in four enterprises now report double-digit AI failure rates. The industry’s answer is to buy more dashboards. That’s like installing a better speedometer in a car with no engine.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SYLYA3KXSZINtz6CqCcGfg.png" /></figure><p>The numbers landed within a few weeks of each other, and they all say the same thing. A <a href="https://www.google.com/url?q=https://www.businesswire.com/news/home/20260309160253/en/New-Study-Reveals-75-of-Enterprises-Report-Double-Digit-AI-Failure-Rates-as-Fragmented-Observability-Hits-Its-Breaking-Point&amp;source=gmail&amp;ust=1781097811022000&amp;sa=E">Virtana study of 351 senior IT leaders</a> found that 75% of enterprises now report AI job failure rates above 10%, and one in three report failure rates above 25%. The <a href="https://www.google.com/url?q=https://hai.stanford.edu/ai-index/2026-ai-index-report&amp;source=gmail&amp;ust=1781097811022000&amp;sa=E">Stanford AI Index 2026</a> reported that agents now clear 66% of real enterprise tasks on benchmarks, up from 12% a year earlier, yet 89% of enterprise agent projects never reach production. <a href="https://www.google.com/url?q=https://www.gartner.com/en/newsroom/press-releases/2026-05-12-gartner-predicts-40-percent-of-organizations-deploying-ai-will-use-ai-observability-to-monitor-model-performance-by-2028&amp;source=gmail&amp;ust=1781097811022000&amp;sa=E">Gartner</a> says only 4% of organizations piloting AI have reached full production maturity, while 62% are still stuck in pilots.</p><p>The consensus reading of all this is now firmly set. The models work. The deployments don’t. And the thing standing between the two is visibility. We can’t see what our agents are doing, so we can’t trust them, so we can’t ship them. The fix, therefore, is observability. Gartner projects that 40% of organizations deploying AI will adopt dedicated AI observability tools by 2028, and that large language model observability will jump from 15% of GenAI deployments to 50% in the same window.</p><p>This reading is comfortable, well-funded, and wrong.</p><p>Observability does not close the AI observability gap. Observability is a mirror, not a fix. It shows you the failure in higher resolution. It does not touch the thing producing the failure. And the thing producing the failure sits one layer below where every observability tool is pointed.</p><h3>Why the observability framing fails</h3><p>Start with what the same studies say when you stop reading the headline. The Virtana report did not actually find that executives lacked dashboards. It found a disconnect: 59% of executives believed their organizations were ready for AI-scale operations, while 62% of practitioners reported fragmented systems and persistent visibility gaps. That is not a tooling deficit. That is two groups looking at the same estate and seeing different things, because the underlying data is scattered across systems that were never designed to agree with each other.</p><p>Forrester’s root-cause work on failed agent deployments is even more direct. It attributes 33% of failures to insufficient tool or data access and another large share to unclear success criteria. <a href="https://www.google.com/url?q=https://hbr.org/sponsored/2026/04/bridging-the-readiness-gap-to-the-agentic-enterprise&amp;source=gmail&amp;ust=1781097811022000&amp;sa=E">HBR Analytic Services</a> found that 94% of leaders say well-connected data is highly important to AI success, while only 27% say their data, processes, and applications are actually well-connected today.</p><p>Read those three findings together. The failures are not happening because nobody is watching. They are happening because the data feeding the agent is fragmented, stale, or contradictory before the agent ever runs. An observability tool sits downstream of that. It can tell you, beautifully, in real time, that your agent just made a confident decision on bad inputs. It cannot make the inputs good.</p><h3>What is actually breaking</h3><p>Picture the failure in motion. An agent is asked to flag at-risk customer accounts. It pulls subscription status from the billing system, support history from the ticketing tool, and usage from the product database. The billing record says active. The product database, synced on a different schedule, says the account churned out eleven days ago. The agent has no way to know which source is right, so it picks one, produces a confident answer, and is wrong about a quarter of the accounts.</p><p>Your observability stack will capture all of this perfectly. The traces, the latency, the token spend, the decision path. You will have a flawless recording of a failure you could not have prevented by recording it. The agent did not hallucinate. It read exactly what you gave it. You gave it two versions of the truth.</p><p>This is why the failure rate is stubborn even as model quality climbs. The Stanford jump from 12% to 66% task success is real. It means the model is no longer the weak link. The weak link is the seam between the eight systems the agent has to reconcile, and no model improvement touches that seam.</p><h3>Why buying more observability won’t help</h3><p>The obvious objection: surely better observability at least helps you find and fix the data problems faster. Sometimes. But there is a trap in the spending pattern itself.</p><p>Most enterprises already run fragmented observability. Virtana found 62% of practitioners citing exactly that. So the realistic 2026 purchase is not the first observability tool. It is the fourth, bolted onto an estate that is already a patchwork. You end up with more dashboards watching more fragments, and a higher monthly bill for the privilege of watching your data contradict itself in greater detail. Visibility into chaos is not order. It is just chaos you can no longer claim you didn’t see.</p><p>There is a quieter cost too. Every dashboard you buy reinforces the belief that the problem is a monitoring problem, which delays the work that actually moves the failure rate: consolidating the data the agents depend on.</p><h3>What actually closes the gap</h3><p>The unglamorous answer is that reliability is built upstream, in the data layer, before observability ever gets a turn. An agent is only as trustworthy as the consistency of the data it reads. That means one place where billing, product, support, and CRM data are reconciled on the same schedule, against the same definitions, so the agent is never handed two versions of a customer.</p><p>This is the part I see up close. At <a href="https://www.google.com/url?q=https://peliqan.io&amp;source=gmail&amp;ust=1781097811022000&amp;sa=E">Peliqan</a> we work with teams trying to make agents reliable, and the pattern is consistent: the teams that get agents into production are not the ones with the best monitoring. They are the ones who consolidated their data into a single governed layer first, so the agent reads one coherent picture instead of stitching together six conflicting ones. Observability comes after that, and at that point it finally earns its keep, because it is watching a system that can actually be trusted.</p><p>Order matters. Foundation, then visibility. Do it the other way around and you have spent 2026 buying very expensive proof that your data was broken all along.</p><p>The AI observability gap is real. It is just misnamed. It is not a gap in what you can see. It is a gap in what you have connected, and no dashboard has ever connected anything.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca6d72aa6875" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The agent system of record is a roster. It is not a data strategy.]]></title>
            <link>https://medium.com/@jprevanth/the-agent-system-of-record-is-a-roster-it-is-not-a-data-strategy-6c749f78ebe9?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c749f78ebe9</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-integration]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 06:02:01 GMT</pubDate>
            <atom:updated>2026-06-02T06:02:01.559Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Every enterprise vendor now wants to be the place your AI agents live. Almost none of them are fixing the reason those agents keep failing.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aPH5K356WHTOA8hvCJse0A.png" /></figure><p>In the space of a few months, the same idea has been launched four different ways. Workday made its <a href="https://www.google.com/url?q=https://blog.workday.com/en-us/managing-ai-powered-future-of-work.html&amp;source=gmail&amp;ust=1780466334239000&amp;sa=E">Agent System of Record</a> generally available, a hub to register, govern, and measure every agent touching Workday data. Airtable published a guide arguing that an <a href="https://www.google.com/url?q=https://www.airtable.com/articles/agent-system-of-record&amp;source=gmail&amp;ust=1780466334239000&amp;sa=E">agent system of record</a> is “the key ingredient to AI agent success in 2026.” No Jitter declared that the <a href="https://www.google.com/url?q=https://www.nojitter.com/data-management/ai-the-system-of-record-wars-have-begun&amp;source=gmail&amp;ust=1780466334239000&amp;sa=E">system of record wars have begun</a>, with ServiceNow and Boomi partnering on data activation. Josh Bersin watched Workday move <a href="https://www.google.com/url?q=https://joshbersin.com/2026/04/the-reinvention-of-workday-from-system-of-record-to-platform-of-agents/&amp;source=gmail&amp;ust=1780466334239000&amp;sa=E">from a system of record to a platform of agents</a> and called it the smartest repositioning the company has made in a decade.</p><p>The consensus reading is straightforward. Agents are proliferating, nobody can see them, and so the winner of the AI era will be whoever owns the registry where all those agents are governed. Control the agent system of record, the thinking goes, and you control the blended workforce.</p><p>I think that reading is exactly backwards, and the confidence behind it is the most expensive mistake in enterprise AI right now.</p><h3>An agent system of record answers the wrong question</h3><p>Look closely at what these products actually do. An agent system of record tells you which agents exist, who owns them, what they cost, and whether they are compliant. That is genuinely useful. Agent sprawl is real, and a registry is a reasonable response to it.</p><p>But notice the question it does not answer: can your agents actually do the work? A registry that says “you have 240 agents, here is the org chart” tells you nothing about whether any single one of them can answer a question that spans your CRM, your ERP, and your support desk. Governing the agents is not the same as governing the data the agents have to read. The industry is conflating the two, and the conflation is the product.</p><p>This matters because the failures are not happening at the registry layer. When <a href="https://www.google.com/url?q=https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027&amp;source=gmail&amp;ust=1780466334239000&amp;sa=E">Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027</a>, it did not blame a missing roster of robots. It blamed escalating cost, unclear value, and the technical complexity of integrating agents into systems that were never designed to be read by a machine. The agents fail because they cannot get to the data. No system of record for the agents themselves changes that.</p><h3>What an agent actually hits when it tries to work</h3><p>Picture the simplest valuable question an enterprise agent gets asked: which customers with active deals and rising support load are at renewal risk this quarter? The deals live in the CRM. The support load lives in a ticketing system. The contract terms live in finance. The product usage lives in a database nobody on the revenue team has credentials for.</p><p>To answer, the agent has to authenticate against four systems, understand four different schemas, respect four sets of rate limits, and reconcile four definitions of “customer.” Most agents simply cannot, so they answer the half of the question that fits inside one application and quietly drop the rest. The demo looked brilliant. Production is a shrug.</p><p>An agent system of record sits politely above all of this and records that the agent ran. It does not make the four systems queryable. It cannot, because it was built to manage agents, not to unify data.</p><h3>Why owning a system of record is the incumbents’ last move, not their best one</h3><p>It is worth being honest about why this category appeared so fast. The companies racing to be your agent’s system of record are the companies most exposed by the agent era. For twenty years their moat was being the place the data lived. An agent that can read across every system, on demand, turns each of those applications into something closer to a commodity source. Reuters reported that the same vendors pouring billions into agents, names like Salesforce and Oracle, are the ones Gartner accuses of “agent washing,” with only around 130 of the thousands of self-described agentic vendors judged to be real.</p><p>Declaring yourself the agent system of record is how an incumbent stays central when the data stops being a moat. It is a defensive land grab dressed as a platform strategy. And it works on buyers because the framing flatters the application they already own: of course the agents should live where the people live, where the work already happens. The trouble is that agents do not work like employees. They do not have a home department. Their whole value is reading sideways across the org, and a system of record built around one application pulls them back into the silo they were supposed to escape.</p><h3>What actually closes the gap</h3><p>The scarce layer in the agent era is not a registry of agents. It is a governed, queryable layer underneath all of them, one place where the CRM, the ERP, the support desk, and the product database have already been synced, cleaned, joined, and made safe to read. Get that right and the agent system of record becomes a thin bookkeeping concern, because every agent is reading from the same trusted surface instead of scrambling across thirty live APIs.</p><p>This is the part I have watched up close. The teams that get real value out of agents are the ones who put a warehouse and a governance layer between the agent and the source systems first. At Peliqan we saw one operator, CIC Hospitality, consolidate more than 50 sources into a single governed warehouse and cut over 40 hours a month of manual board reporting, with agents, dashboards, and reverse ETL all reading from the same layer rather than each re-solving integration on its own. The unglamorous pairing that makes this work is something like <a href="https://www.google.com/url?q=https://peliqan.io/connectors/Pipedrive/Exact%2520Online/&amp;source=gmail&amp;ust=1780466334239000&amp;sa=E">a CRM and an accounting system in one queryable place</a>, so “at-risk revenue” finally means the same thing to the agent that it means to the CFO.</p><p>None of that requires anyone to own a new system of record. It requires the data to stop being thirty islands.</p><p>So when the next vendor offers to be the system of record for your agents, ask the only question that predicts whether the project survives to 2027: not “where will my agents be governed,” but “where will my agents read.” If the answer is still “from thirty live systems, one at a time,” the roster is not going to save you.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c749f78ebe9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The MCP security crisis isn’t a vulnerability. It’s a design choice.]]></title>
            <link>https://systemweakness.com/the-mcp-security-crisis-isnt-a-vulnerability-it-s-a-design-choice-3021c0e18b62?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/3021c0e18b62</guid>
            <category><![CDATA[security]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[mcps]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Tue, 26 May 2026 11:04:24 GMT</pubDate>
            <atom:updated>2026-05-27T08:37:23.072Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Anthropic’s protocol crossed 97 million downloads and joined the Linux Foundation. Then a researcher showed 200,000 servers can be hijacked with a single line of code. Anthropic’s response: this is how it’s supposed to work.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4f_VV6LmAY0PNf6Ck4-LUA.png" /></figure><p>In December 2025, Anthropic <a href="https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation">donated the Model Context Protocol</a> to a new Linux Foundation effort called the Agentic AI Foundation. The press releases were the usual coronation language: vendor-neutral governance, open standards, the future of AI agents. OpenAI, Google, Microsoft, AWS, Cloudflare, and Bloomberg signed on.</p><p>Four months later, <a href="https://www.ox.security/blog/the-mother-of-all-ai-supply-chains-critical-systemic-vulnerability-at-the-core-of-the-mcp/">a research team at OX Security</a> published a paper they called “The Mother of All AI Supply Chains.” They had found a design flaw at the core of MCP that lets an attacker run arbitrary OS commands on any of the roughly 200,000 servers running the protocol. The researchers demonstrated four working attack chains, including one that successfully poisoned nine of the eleven public MCP marketplaces they tested.</p><p>OX repeatedly asked Anthropic to patch the underlying design. <a href="https://www.theregister.com/2026/04/16/anthropic_mcp_design_flaw/">Anthropic’s response, on the record</a>: the protocol works just fine. The behavior is “expected.”</p><p>That sentence is the entire story.</p><h3>Why everyone shipped MCP in the first place</h3><p>For the executives who have not had to care about this until now: MCP is the protocol that lets AI agents call tools and access data. Connect Claude to your Salesforce. Wire ChatGPT into your warehouse. Give your coding agent access to your repo. Every one of those connections is an MCP call. It is the USB-C of agent infrastructure, which is exactly what Anthropic set out to build.</p><p>The numbers explain why everyone shipped it. <a href="https://www.digitalapplied.com/blog/mcp-97-million-downloads-model-context-protocol-mainstream">97 million monthly SDK downloads in March 2026, a 970x increase in eighteen months</a>. More than 12,000 public MCP servers. Every major model vendor ships MCP-compatible tooling by default. It is, by any honest measure, the most successful new protocol launched in enterprise software since GraphQL.</p><p>It is also, by the same measure, the most under-secured.</p><h3>The “vulnerability” is the protocol working as designed</h3><p>Here is the part executives keep missing. The MCP “vulnerability” is not a bug in a particular implementation. It is a property of the spec.</p><p>MCP servers communicate over standard input/output using a launch model where the client process spawns the server as a subprocess. The launch string is passed in by configuration. The launch string can be any shell command. The protocol does not constrain it.</p><p>To run a legitimate MCP server, you execute the launch string. To run an attacker’s payload, you execute the launch string. The protocol does not distinguish.</p><p>That is what OX Security found, and what Anthropic was unwilling to change. To “fix” the flaw would be to redesign the launch model. Which would break every MCP server published in the last eighteen months. Which is precisely why Anthropic told the researchers the protocol works fine.</p><h3>Four attack vectors. The last one should keep CISOs up.</h3><p>OX demonstrated four working exploits.</p><p><strong>One.</strong> Unauthenticated command injection against MCP-integrated frameworks including LangFlow and GPT Researcher. CVE-2025–65720 was issued.</p><p><strong>Two.</strong> Hardening bypass. Even where developers had added input sanitization, the underlying protocol still permitted command execution.</p><p><strong>Three.</strong> <a href="https://thehackernews.com/2026/04/anthropic-mcp-design-vulnerability.html">Zero-click prompt injection across the major AI coding agents</a> — Claude Code, Cursor, Windsurf, Gemini-CLI, GitHub Copilot. CVE-2026–30615 was issued for Windsurf. The “zero-click” part is important. The developer does not have to do anything wrong. Opening a file is enough.</p><p><strong>Four.</strong> Marketplace poisoning. The researchers successfully published malicious MCP servers to nine of the eleven public MCP marketplaces they tested, including registries enterprises are using to discover and install tools.</p><p>That last one is the one that turns a security incident into a supply chain crisis. The moment an enterprise’s AI platform team installs an MCP server from a registry — which is the default behavior MCP was designed to encourage — they are running unknown code with access to the agent’s full toolset. That toolset usually includes the warehouse, the CRM, internal APIs, and developer credentials.</p><h3>We have seen this movie. It was called npm.</h3><p>The OX paper’s title is not just dramatic flourish. The architectural pattern MCP just standardized — frictionless publishing to public registries, deep system access by default, opt-in security after the fact — is the same pattern that produced the npm catastrophe of the last several years.</p><p>In 2024 and 2025, npm absorbed worm attacks like Shai-Hulud and the axios compromise. <a href="https://unit42.paloaltonetworks.com/monitoring-npm-supply-chain-attacks/">Tens of thousands of packages were backdoored. Credentials were stolen across the open-source ecosystem.</a> Mitigation work is still ongoing.</p><p>The difference with MCP is that npm packages mostly run inside developer build pipelines. MCP servers run with the explicit purpose of giving an autonomous agent permission to act on real systems in production. An npm exploit gets you developer credentials. An MCP exploit gets you the agent’s authority to write to the database, send the email, and place the order.</p><p>The blast radius is not the same.</p><h3>The Linux Foundation handoff looks different now</h3><p>Run the timeline again. December 2025: Anthropic donates MCP. April 2026: OX Security publishes the disclosure. The handoff happened four months before the world found out the protocol it was now governing had a design flaw the original maintainers had refused to fix.</p><p>This is not an accusation that Anthropic knew specifically about the OX work. It is an observation that the optics of donating a flawed protocol to an industry foundation, weeks before its biggest design problem becomes public, are not great. The Agentic AI Foundation is now the institution responsible for fixing this. Anthropic is not.</p><p>That is a successful platform handoff. It is also a successful hot potato.</p><h3>What to do this quarter</h3><p>If you are running an agent program in 2026, three things matter.</p><p>One. Assume every MCP server you have not personally code-reviewed is hostile. Treat the marketplaces the way you would treat npm — a publishing platform, not a vetted distribution channel.</p><p>Two. Agents that call MCP servers must run in sandboxed execution environments with explicit allowlists. The MCP tunnels feature Anthropic shipped in research preview on May 19 moves in this direction. The default configuration does not.</p><p>Three. Stop letting “MCP-native” be a procurement checkbox. If a vendor’s entire pitch is that they expose their product as an MCP server, that is a description of their attack surface, not their security posture.</p><h3>The protocol that won is going to lose something</h3><p>MCP won. There is no realistic scenario where a different protocol displaces it now. Linux Foundation governance, model-vendor consensus, 97 million downloads — that race is over.</p><p>What has not been settled is whether MCP wins on the original terms, or on terms set by the first major MCP-driven enterprise breach. Right now, MCP’s security story is the story Anthropic chose — the protocol works fine, the behavior is expected, the rest is your problem.</p><p>The first time a Fortune 500 finds out their AI agent’s MCP server was the entry point for the breach, that story will not hold.</p><p>The protocol will not be redesigned. The defaults will.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3021c0e18b62" width="1" height="1" alt=""><hr><p><a href="https://systemweakness.com/the-mcp-security-crisis-isnt-a-vulnerability-it-s-a-design-choice-3021c0e18b62">The MCP security crisis isn’t a vulnerability. It’s a design choice.</a> was originally published in <a href="https://systemweakness.com/">System Weakness</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Enterprise AI inference cost is the only line item the CFO can no longer ignore]]></title>
            <link>https://medium.com/@jprevanth/enterprise-ai-inference-cost-is-the-only-line-item-the-cfo-can-no-longer-ignore-d22e3296b9d4?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/d22e3296b9d4</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-cost]]></category>
            <category><![CDATA[ai-bill]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Mon, 18 May 2026 12:18:13 GMT</pubDate>
            <atom:updated>2026-05-18T12:18:13.667Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Token prices collapsed 280x in two years. Enterprise AI bills tripled. The consensus says efficiency is winning. The data says agentic workflows just rewrote the math, and the CFO is about to start canceling projects the CIO can’t justify.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hanVETQtZKHJaa-K11uU-Q.png" /></figure><p>Two numbers landed in front of CFOs this week, and they do not fit together. Token prices have fallen 280x in two years, according to <a href="https://analyticsweek.com/inference-economics-finops-ai-roi-2026/">AnalyticsWeek’s 2026 Inference Economics report</a>. Over the same period, average enterprise AI budgets climbed from $1.2 million to $7 million per year. Some Fortune 500 names are now reporting <a href="https://www.cloudzero.com/blog/nvidia-ceo-keynote-ai-inference-cost/">monthly inference bills in the tens of millions</a>. Microsoft’s research drop, covered by Fortune <a href="https://fortune.com/2026/05/11/what-microsoft-research-tells-cfo-roi-ai/">this Monday</a>, told CFOs the active-agent count in Microsoft 365 grew 15x year over year, 18x for large enterprises. Anthropic and OpenAI are both <a href="https://aiautomationglobal.com/blog/ai-inference-cost-crisis-openai-economics-2026">losing money on inference</a> at current pricing — OpenAI was estimated to be spending $1.35 for every $1 it earned in 2025. Bain’s CFO note from this month was blunt: <a href="https://www.bain.com/insights/cfos-funded-ai-revolution-now-they-are-joining-it/">CFOs funded the AI revolution. Now they are joining it.</a></p><p>The consensus reading is that AI got cheaper. Efficiency won. Token deflation is the proof, and 83 percent of CFOs plan to increase AI budgets above 15 percent over the next two years.</p><p>That reading dies on contact with the actual budget. Token prices collapsed and bills tripled in the same window. Anyone celebrating efficiency is reading the wrong line item. The real story is this: agentic workflows consume 5 to 30 times more tokens per task than chatbots, vendor pricing is propped up by VC cross-subsidy that is now visibly breaking, and the people in the room who can do this math are not the CIO or the Chief Data Officer. They are the CFO. And the CFO is about to become the most important AI buyer in your company.</p><h3>Why the “token deflation” narrative misleads on enterprise AI inference cost</h3><p>The cheap-token narrative is correct on its face. Inference cost per token has dropped 280-fold. The story does not survive the second derivative. An agentic workflow does not make one model call. It makes 10 to 20, according to <a href="https://oplexa.com/ai-inference-cost-crisis-2026/">multiple inference economics breakdowns</a>. A simple chatbot query costs roughly $0.001. A multi-step agent that plans, retrieves context, invokes tools, reflects, and self-corrects can cost $0.10 to $1.00. That is a 100x to 1,000x multiplier sitting on top of a 280x deflation in unit price.</p><p>Two-hundred-and-eighty divided by a thousand is less than one.</p><p>Combine the multiplier with always-on patterns — agents that monitor email queues, log streams, market data, and operational systems continuously instead of on demand — and the unit-economic story flips. CIO reported this quarter that CFO cloud bills are <a href="https://www.cio.com/article/4163877/the-inference-bill-nobody-budgeted-for.html">running 40 percent above forecast</a>. That gap is the gap between the press release and the invoice.</p><h3>The vendor side of the math is breaking too</h3><p>OpenAI’s 2025 numbers, surfaced by independent analysis, show the company spent $1.35 for every dollar earned. That is not a sustainable race-to-zero pricing model. It is venture capital and hyperscaler subsidies financing your inference contract. The Bain note this month framed it carefully: token economics today are not the token economics enterprises should plan against in 2027.</p><p>Translation: the API price your CIO budgeted in March is not the API price your CFO will see in November.</p><p>Andreas Timm <a href="https://andreastimm.com/inference-cost-collapsed-enterprise-business-cases/">wrote it directly</a> two weeks ago. The collapse in inference cost happened. Enterprise business cases did not catch up. They are still written against 2024 token prices and chatbot-shaped usage patterns. Replace either assumption and the ROI flips negative.</p><h3>Why “wait for it to get cheaper” will not work</h3><p>The instinct in most boardrooms is to wait. Inference is getting cheaper every quarter. The right move is to deploy now and let the cost curve catch up.</p><p>It will not. Two things are working against the wait.</p><p>First, agent token consumption is growing faster than per-token price is dropping. NVIDIA’s own keynote framing — which CloudZero translated into the actual budget impact — was that <a href="https://www.cloudzero.com/blog/nvidia-ceo-keynote-ai-inference-cost/">inference is the next big cost problem</a>, not the next big efficiency win. Second, the cross-subsidy that is keeping API prices low has a clock on it. When OpenAI and Anthropic raise prices to reach unit economics, the enterprise inference line jumps overnight — and the contracts that were sized against 2025 pricing become the budget incident that gets your AI program killed in committee.</p><p>The CFO is the only person in the room who tracks both lines. The CIO sees agent count growing. The CDO sees data infrastructure investments paying back. The CFO sees a cloud bill that broke the forecast by 40 percent and an OpenAI invoice that grew while the unit cost technically fell.</p><h3>What actually works: instrument enterprise AI inference cost like you instrument the data layer</h3><p>Three moves matter. First, treat per-agent cost as an observability metric, not a finance report. Most enterprises can tell you how many agents they have deployed. Very few can tell you what each agent cost to run last month. That asymmetry is what kills projects in 2027 budget season. Second, separate steady-state inference from experimental inference in the budget. The first is operational and predictable. The second is R and D and should be capped. Third, model the next contract against vendor unit economics, not current promotional pricing. If your vendor is losing money on inference, plan against a 2x-3x repricing within 24 months.</p><p>This is one of the conversations we keep having at Peliqan with mid-market customers who run their data integration, warehouse, and agent activation in one place. The reason it lands is not feature parity. It is the FinOps story. When the warehouse, the pipelines, and the agent runtime share a single bill, the CFO can actually do the math. When they live in four vendors with four pricing models, the only thing the CFO can do is freeze the budget.</p><p>The CFO did not used to be in this conversation. They are now. The companies that will still be running agents in 2028 are the ones whose CFOs were doing the unit economics in 2026.</p><p>The token price dropped. The bill went up. Both are true. Only one of them is on the agenda.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d22e3296b9d4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI agent context layer is the real reason agents aren’t employees]]></title>
            <link>https://medium.com/@jprevanth/the-ai-agent-context-layer-is-the-real-reason-agents-arent-employees-69e79bcc2e03?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/69e79bcc2e03</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-context-layer]]></category>
            <category><![CDATA[data-integration]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Thu, 14 May 2026 06:14:56 GMT</pubDate>
            <atom:updated>2026-05-14T06:14:56.163Z</atom:updated>
            <content:encoded><![CDATA[<p><em>HBR and BCG just published research saying companies shouldn’t treat AI agents like coworkers. Their evidence is right. Their explanation isn’t. The reason isn’t accountability or role design. It’s that enterprise data pipelines break 4.7 times a month, and every break is an agent that just lost its memory.</em></p><p>Three things happened in enterprise data infrastructure this week. Harvard Business Review and BCG <a href="https://hbr.org/2026/05/research-why-you-shouldnt-treat-ai-agents-like-employees">published research</a> drawn from a study of more than 1,200 managers, finding that framing an AI agent as an “employee” cut error detection by 18 percent and pushed individual accountability down by 9 percentage points. Boomi rebranded itself from an integration company to <a href="https://siliconangle.com/2026/05/13/data-activation-age-agentic-ai-boomiworld/">“the data activation company for AI”</a> at BoomiWorld 2026, explicitly positioning to fight Snowflake and Databricks for the layer that feeds enterprise agents. And Fivetran’s <a href="https://www.fivetran.com/press/data-pipeline-failures-cost-enterprises-3-million-per-month-fivetran-benchmark-finds">2026 enterprise benchmark</a>, drawn from 500 senior data leaders at companies above 5,000 employees, found that 97 percent of them said pipeline failures have slowed their analytics or AI programs, with an average of 4.7 failures per month and 13 hours to resolve each one.</p><p>The consensus reading is comfortable. HBR’s prescription is governance. Publish what each agent does. Name a human owner. Document the override path. Run quarterly reviews. Treat the agent as a system, not a coworker, and accountability returns.</p><p>That fix does not survive contact with the actual failure mode. The reason agents should not be treated like employees is not accountability or role design. It is context loss. Human employees carry institutional memory, tribal knowledge, the conversation from three weeks ago, the customer preference from last quarter. They walk into a problem with continuity. Agents do not. Agents retrieve their context from data, and that data layer breaks — according to Fivetran’s number — 4.7 times a month, every month. Every break is an agent that just lost its memory.</p><h3>The HR framing makes the symptom look like the cause</h3><p>Karim Lakhani at Harvard has spent two years arguing that AI adoption is fundamentally an organizational problem. He is not wrong, and the BCG study from this week corroborates him. 18 percent fewer errors caught when an agent is anthropomorphized is a real effect. Treating a probabilistic system as if it had a coworker’s judgment is a measurable management failure.</p><p>But the governance fix — documentation, ownership, quarterly review — presumes the agent has a stable context to operate inside. It does not. Fortune <a href="https://fortune.com/2026/04/30/agentic-ai-data-infrastructure-readiness-scale/">put this plainly</a> two weeks ago: your data infrastructure, not your model, determines whether agentic AI scales. The Fivetran benchmark put a number on that gap. 4.7 pipeline failures per month per enterprise. 13 hours mean time to resolve. $3 million in average monthly business exposure for large organizations.</p><p>You can publish a perfect job description for an agent and still have that agent answering a customer with stale inventory data because a CDC stream broke at 2 AM and will not be fixed until 3 PM.</p><h3>The real failure mode is silent context drift</h3><p>Picture a procurement agent inside a mid-market manufacturer. It is supposed to flag suppliers whose lead times are slipping. It pulls context from three pipelines — the ERP, a SaaS supplier scorecard tool, and a freight feed. The freight pipeline silently fails on Tuesday morning. No alert fires because the schema did not change. It just stopped backfilling new rows. By Thursday, the agent has approved six purchase orders against stale freight assumptions. Nobody notices until accounts payable reconciles two weeks later.</p><p>That is not a role design failure. The agent did exactly what its documented owner asked it to. It is a context layer failure. The data underneath the agent lost continuity, and the agent had no way to know.</p><p>OpenAI’s engineering team published <a href="https://atlan.com/know/context-layer-for-ai-agents/">the design behind their internal data agent</a> in January. Six explicit context layers, version-controlled and observable. Time to answer a complex business question dropped from 22 minutes to 1 minute 22 seconds. The interesting number is not the speedup. It is what it took to get there. Not better governance. Six layers of explicit, queryable context underneath the model. That is the AI agent context layer most enterprises are now realizing they need — and most do not have.</p><h3>Why “buy a data activation platform” will not fix it</h3><p>Every major data platform vendor has pivoted to “data activation for AI” in the past 90 days. Boomi rebranded. Snowflake launched <a href="https://www.snowflake.com/en/product/features/openflow/">Openflow</a>. Databricks expanded Lakeflow. Each is racing to own the semantic layer agents read from.</p><p>This is a vendor consolidation play dressed up as a category. Whoever wins captures the next decade of enterprise lock-in. The CIO who buys the warehouse, the activation layer, and the agent runtime from the same vendor has just bought a closed loop that will be expensive to leave in 2029.</p><p>The same Fivetran benchmark found that 70 percent of enterprises discovered their data infrastructure was inadequate only after launching ambitious AI initiatives. Buying a tighter loop from one of those vendors solves nothing if the underlying integration architecture is still the bottleneck. You are activating broken data faster. The activation layer is downstream of the failure. The failure is in pipeline reliability, lineage, and silent-failure detection — and none of that is on the launch deck at BoomiWorld.</p><h3>What actually fixes the AI agent context layer</h3><p>Three moves matter. First, treat the context layer as portable infrastructure, not a vendor feature. If your agent’s context is locked inside the same platform that holds your warehouse, you have built a single point of vendor failure. Second, instrument every pipeline for freshness, lineage, and silent-failure detection — not just schema change alerts, which are what most teams stop at today. Third, give agents observable access to their own context state, so they can refuse to act on stale data instead of confidently activating it.</p><p>This is the thesis we keep coming back to at Peliqan, and it is why mid-market enterprises — who cannot absorb a Boomi-plus-Snowflake-plus-orchestration stack — keep telling us they want the integration layer, the warehouse, and the activation layer in one place but exportable. Not because any single capability is better in isolation, but because the moment your context layer is locked inside one vendor, your agents are too.</p><p>HBR and BCG are right that you should not treat your AI agents like employees. They are wrong about why. Employees do not need 4.7 freshness checks a month to remember what business they work in. Agents do. Until the context layer is portable, observable, and reliable, every governance memo in the building is a Post-it note stuck to a server room that is on fire.</p><p><em>I write about data integration, AI infrastructure, and the gap between what the industry promises and what enterprises actually experience. If this resonated, follow along — I publish a few times a week.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69e79bcc2e03" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Belgium turned on mandatory e-invoicing in January.]]></title>
            <link>https://medium.com/@jprevanth/belgium-turned-on-mandatory-e-invoicing-in-january-1e00ac58a046?source=rss-149486c4464------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e00ac58a046</guid>
            <category><![CDATA[mcps]]></category>
            <category><![CDATA[billit]]></category>
            <category><![CDATA[peppol]]></category>
            <category><![CDATA[peppol-e-invoicing]]></category>
            <category><![CDATA[belgium]]></category>
            <dc:creator><![CDATA[Revanth]]></dc:creator>
            <pubDate>Mon, 11 May 2026 05:44:49 GMT</pubDate>
            <atom:updated>2026-05-11T05:49:14.539Z</atom:updated>
            <content:encoded><![CDATA[<h3>Belgium turned on mandatory e-invoicing in January. Most CFOs still can’t tell which invoices cleared Peppol.</h3><p><em>Belgium became the first major EU economy to flip the switch on mandatory B2B Peppol invoicing. The grace period ended April 1. Fines are active. Here’s what the gap looks like on the ground — and why France, Germany, and Spain should care.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_9-_lCDBbexmM2lhP1eIwg.png" /></figure><p>On January 1, 2026, every VAT-registered business in Belgium became legally required to send and receive its B2B invoices as structured UBL documents over the Peppol network. The change is generational. Paper invoices and PDF attachments — the format Belgian SMEs have used for forty years — are no longer compliant by default.</p><p>Three months later, on April 1, the FOD Financiën quietly let the tolerance period expire. Fines moved from “theoretical” to “being levied.” The fine schedule, published by <a href="https://www.loyensloeff.com/insights/news--events/news/belgium--mandatory-b2b-e-invoicing-as-of-1-january-2026/">Loyens &amp; Loeff</a> and confirmed by <a href="https://www2.deloitte.com/be/en/pages/tax/articles/mandatory-b2b-e-invoicing-belgium.html">Deloitte Belgium</a>, runs €1,500 to €5,000 per offence — plus proportional VAT penalties of 60% to 100% on the underlying transaction. The first FOD Financiën audits referencing Peppol message logs have started circulating among Brussels tax advisors.</p><p>And yet, walk into a 50-person Belgian SME this week and ask the controller a simple question: <em>“Of the invoices we sent in April, which ones haven’t been acknowledged by the buyer’s Peppol endpoint yet?”</em> — and you will get a shrug, a half-hour of Excel work, or, more often, no answer at all.</p><p>The mandate is live. The visibility into the mandate is not.</p><h3>What “mandatory Peppol” actually means</h3><p>To recap for non-Belgian readers: Peppol is an EU-backed network for exchanging procurement and invoicing documents in structured XML format. Sender → Access Point → SMP lookup → Access Point → Receiver. It’s been used voluntarily across the EU for a decade, mostly for public procurement.</p><p>Belgium’s mandate, codified in the law of 6 February 2024 and effective from 1 January 2026, makes this the <em>only</em> legal way for two Belgian VAT-registered businesses to exchange invoices. There are narrow exemptions (very small businesses, certain B2C flows, foreign suppliers without a Belgian VAT number), but for the bulk of the Belgian economy, Peppol or non-compliant are now the only two options.</p><p>The piece that catches Belgian CFOs off guard is the acknowledgement model. Sending an invoice through Peppol is not the same as the buyer having received and validated it. The buyer’s Access Point fires a Message Level Response (MLR) back through the network. That MLR can take minutes — or, when buyer-side validation queues are saturated, hours or even days. According to <a href="https://www.vatcalc.com/belgium/belgium-e-invoicing-update/">vatcalc’s tracker</a>, roughly 12% of Belgian Peppol traffic still shows ack delays exceeding four hours in the post-April 1 period.</p><p>This is the asymmetry the mandate created. “Did I file my invoice?” is no longer a yes/no question. It’s a stateful, time-dependent reconciliation problem, and most finance teams have no tooling for it.</p><h3>Why Billit sits at the centre of this story</h3><p><a href="https://www.billit.eu/">Billit</a> is a Ghent-headquartered invoicing platform and Peppol Access Point. By their own published figures, they serve roughly 120,000 active users, process more than 2 million invoices per month, and handle a majority share of all Belgian Peppol traffic — the company <a href="https://www.billit.eu/blog/">estimates</a> over 50%, with corroborating figures from the <a href="https://efactuur.belgium.be/">Belgian e-invoicing government portal</a>.</p><p>For most Belgian SMEs, Billit <em>is</em> the Peppol layer. Their accountant uses Billit. Their ERP integrates with Billit. Their first encounter with the mandate, in practice, is a Billit dashboard.</p><p>Billit’s own UI gives you the standard things: send invoice, see status, download UBL. What it does not give you — because no transactional UI ever does — is the analytical layer. The CFO question is never “what does this one invoice look like?” The CFO question is always “which of these 3,200 invoices are non-compliant, slow-acking, or stuck — and what is my exposure?”</p><p>That’s the question this post is about.</p><h3>The four questions every Belgian CFO is now asking</h3><p>Across the Belgian finance leaders we’ve spoken to over the last six weeks, the same four questions surface in every conversation:</p><p><strong>1. Compliance attestation.</strong> <em>Which of our invoices in the last 90 days didn’t make it through Peppol at all?</em> This is the question your tax advisor will ask in the next audit. There’s no Billit report that answers it cleanly because the events live in three places: the invoice record, the PeppolEvent stream, and the buyer’s MLR response.</p><p><strong>2. Working capital exposure.</strong> <em>Which buyers haven’t acknowledged invoices that are now overdue?</em> Peppol non-acknowledgement is a red flag for working capital — it usually means the invoice never landed in the buyer’s AP system. The collection conversation needs to start two weeks earlier than it currently does.</p><p><strong>3. Counterparty readiness.</strong> <em>Which of our suppliers still aren’t Peppol-enabled?</em> You can’t legally pay an invoice that arrived outside Peppol from a Belgian counterparty. Procurement needs a list. Nobody has one.</p><p><strong>4. Multi-client reporting (for accountancy firms).</strong> <em>Across our 200 clients, which clients have the highest non-compliance rate this month?</em> This is the question every Belgian fiduciary is asking about its own book of business right now. The answer requires fan-out across 200 separate Billit organisations and a unified rollup.</p><p>All four of these are SQL questions. None of them are UI questions. And none of them are answered today in any tool any Belgian SME or accountancy firm has installed.</p><h3>What we built</h3><p><a href="https://peliqan.io/">Peliqan</a> is a managed data platform — Postgres + Trino warehouse, governed connector layer, MCP server on top. We’ve been EU-hosted (Frankfurt + Brussels), SOC 2 Type II, and GDPR-native from day one.</p><p>On the back of the mandate, we shipped what we believe is the first MCP server tied specifically to Belgian Peppol compliance: the <a href="https://peliqan.io/model-context-protocol/Billit-mcp/">Billit MCP</a>. Two-line elevator pitch:</p><ol><li>We sync every Billit endpoint — invoices, customers, products, Peppol participants, PeppolEvents, acknowledgements — into a managed warehouse on a 15-minute cadence.</li><li>We expose that warehouse to Claude, ChatGPT, Cursor, and n8n via the Model Context Protocol.</li></ol><p>The architectural piece that matters for the Peppol use case specifically is what we call <em>stateful materialisation of asynchronous events</em>. Peppol acks arrive async via webhook. To answer “which invoices we sent last week haven’t been acknowledged,” you need a stateful reconciliation table that LEFT JOINs sent invoices to the latest matching MLR. We materialise this as a peppol_delivery_state view, refreshed on its own schedule, and expose it to the MCP layer as a single queryable table.</p><p>The full technical write-up — including the Trino federation layer, the writeback safety model, and how we handle multi-administration fan-out for accountancy firms — is in our companion piece, <a href="https://peliqan.io/blog/billit-claude-peppol/">Connecting Billit and Claude for Peppol compliance</a>.</p><h3>What this looks like from a Claude prompt</h3><p>A Belgian controller, sitting in Claude Desktop with the Peliqan MCP connected, can now type:</p><blockquote>“Show me all April invoices we sent through Billit where the buyer hasn’t acknowledged within 48 hours. Group by customer, sort by total exposure.”</blockquote><p>Behind that prompt, Claude composes a SQL query against the Trino warehouse, joins three Billit tables, applies the 48-hour threshold, and returns a ranked list — typically in under two seconds.</p><p>The follow-up is the part that matters. The controller can then ask:</p><blockquote>“For each of those customers, what’s the open opportunity value in HubSpot, and the upcoming payment date in our Mollie reconciliation?”</blockquote><p>That second query crosses Billit, HubSpot, and Mollie in a single Trino plan. Three minutes of work where there used to be three days.</p><p>The writeback closes the loop. The controller can ask Claude to draft a corrected UBL invoice, validate it, and resend it through Peppol. Every mutation routes through our governed connector layer — the AI never hits Billit’s API directly, which means we can apply rate limits, audit logs, and rollback semantics that the live API doesn’t provide on its own.</p><h3>Why this matters beyond Belgium</h3><p>Belgium is the first big EU economy to ship a mandatory Peppol regime. It will not be the last.</p><p>France’s <a href="https://www.economie.gouv.fr/cedef/facturation-electronique">domestic e-invoicing reform</a> is phasing in across 2026 — phase one for large enterprises in September, broader rollout into 2027. Germany’s <a href="https://www.bundesfinanzministerium.de/">e-invoicing mandate</a> is staged through 2026–2028, with mandatory B2B receipt already live since January 2025 and mandatory issuance phasing in. Spain’s <a href="https://www.boe.es/eli/es/l/2022/09/29/18">Crea y Crece law</a> has a B2B e-invoicing component pending regulatory specification, expected to land late 2026 or 2027. The EU’s <a href="https://taxation-customs.ec.europa.eu/vat-digital-age_en">VAT in the Digital Age (ViDA) package</a> sets 2030 as the bloc-wide harmonisation target for cross-border B2B Peppol.</p><p>The Belgian mandate is, in other words, the dress rehearsal. Whatever tooling and workflows shake out for Belgian businesses over the next twelve months will be the template that French, German, and Spanish CFOs adopt eighteen months later. The Belgian experiment is the European experiment, on a fast timeline.</p><p>The lesson, if there is one: regulatory mandates create three things at once — a compliance problem, an analytical problem, and a workflow problem. The compliance problem is well-served by incumbent invoicing platforms like Billit. The analytical and workflow problems are the gap. AI-grade tooling tied to mandate-native data is the bet we’re making, and the bet we think every EU SaaS company building near a regulated workflow should be making.</p><h3>Try it</h3><p>The Billit MCP is live. EU-hosted, SOC 2 Type II, GDPR-native, free tier available.</p><ul><li>Product page → <a href="https://peliqan.io/model-context-protocol/Billit-mcp/">peliqan.io/model-context-protocol/Billit-mcp/</a></li><li>Technical write-up → <a href="https://peliqan.io/blog/billit-claude-peppol/">peliqan.io/blog/billit-claude-peppol/</a></li><li>Umbrella server (Billit + 249 more EU apps) → <a href="https://peliqan.io/model-context-protocol/">peliqan.io/model-context-protocol/</a></li></ul><p>If you’re a Belgian CFO, fiduciary, or systems integrator trying to answer the four questions above and not finding an answer, drop me a line on LinkedIn or at rev@peliqan.io. We’ll find an hour this week.</p><p>The mandate isn’t going to un-mandate itself. But the visibility into it is now solvable.</p><p><em>Rev Periyasamy is Senior Marketing Manager at </em><a href="https://peliqan.io/"><em>Peliqan</em></a><em>, the data platform and MCP server for European business applications. Peliqan is based in Gent, Belgium, EU-hosted, and SOC 2 Type II certified.</em></p><p><em>Sources cited in this article: </em><a href="https://www.loyensloeff.com/"><em>Loyens &amp; Loeff</em></a><em>, </em><a href="https://www2.deloitte.com/be/"><em>Deloitte Belgium</em></a><em>, </em><a href="https://www.vatcalc.com/"><em>vatcalc</em></a><em>, </em><a href="https://efactuur.belgium.be/"><em>Belgian government e-invoicing portal</em></a><em>, </em><a href="https://www.billit.eu/"><em>Billit</em></a><em>, </em><a href="https://taxation-customs.ec.europa.eu/vat-digital-age_en"><em>European Commission ViDA package</em></a><em>. All compliance figures and fine schedules as published at time of writing (11 May 2026); readers should consult a Belgian tax advisor before making compliance decisions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e00ac58a046" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>