On August 2, 2026, the EU AI Act’s obligations for high-risk AI systems become binding. Most of the attention has gone to model providers — but one requirement lands squarely on teams that use AI with tools: Article 12 requires automatic logging of events over the system’s lifecycle, with records retained so that operation can be traced and audited. Several analyses now argue that when AI agents invoke APIs or MCP servers, that action layer is in scope.
One disclaimer before anything else: this is not legal advice, and whether your specific system counts as high-risk is a question for your counsel. What this post covers is the practical side: if you need to log what your AI did, what does that actually take — and what does it mean for MCP?
Why the action layer is the hard part
Logging a chat transcript is easy. Logging actions is where it gets real: when your team uses Claude or ChatGPT with MCP servers connected, the consequential events are tool calls — a ticket closed, a record updated, a file read, an email drafted. Those calls happen client-side, spread across every employee’s individual configuration. If ten people each wired their own MCP servers into their own AI clients, there is no place where those logs exist at all — let alone one you could hand an auditor.
That’s the structural problem: compliance requires a chokepoint, and per-employee MCP configs have none.
What a compliant action-layer log looks like
- Automatic — recorded by infrastructure, not by asking people to remember.
- Complete — every tool call, including the failed and the blocked ones.
- Attributable — which person’s AI client did it, when, to which system.
- Retained — Article 12 logs must be kept (six months as a baseline in the Act).
- Readable — an auditor (or your own postmortem) can follow it without parsing JSON.
The gateway pattern
This is exactly the problem an MCP gateway exists to solve, and why we think the regulation quietly makes gateways the default architecture for teams in the EU: route every AI client through one endpoint, and the chokepoint you need for compliance is the same chokepoint that gives you access control and policies. In gate, every tool call any team member’s AI makes is logged automatically with a plain-language summary — “Claude closed JIRA-142 at 14:03” — searchable, exportable, and attributable per client and per person.
The same chokepoint handles the other half of responsible AI-tool use: per-tool allow/ask/block rules, security scanning of servers before they go live, and instant revocation when someone leaves. Compliance and control turn out to be the same piece of infrastructure.
What to do this week
- Inventory which AI clients your team runs and which MCP servers are connected — including the ones nobody approved (shadow MCP is real).
- Ask your counsel whether your use cases touch the high-risk categories.
- Regardless of the answer: put the action layer behind one logged endpoint. Even if the Act never applies to you, “what did our AI actually do last Tuesday?” is a question you eventually have to answer for yourself.
gate gives you the logged endpoint in minutes, free to start — here’s the team setup.