The six-layer AI architecture checklist

Two per layer, in the order a request passes through them. A yes counts only when the control runs in code today.

Each yes scores one point.

1. 1.The AI feature works inside the signed-in user's permissions. Layer 1, product interface. It calls the same APIs as the rest of the product, so it can't read or change more than the person using it.
2. 2.The interface says it's AI and marks what it generated. Layer 1. A notice where people talk to the system, and a machine-readable mark on generated content (art. 50(1) and 50(2) Regulation (EU) 2024/1689).
3. 3.Every request passes authentication and validation before a prompt exists. Layer 2, access and policy. Size limits, schema checks, and rate limits run in ordinary code, ahead of the model.
4. 4.A policy step removes or pseudonymises the personal data a task doesn't need. Layer 2. Data minimisation in practice (art. 5(1)(c) Regulation (EU) 2016/679). The model sees a token, and our own system swaps the real value back in.
5. 5.Retrieval reads only the sources we approved. Layer 3, context retrieval. A named list of data sources, each with an owner. A new source goes through review like a code change.
6. 6.The retrieval service applies the tenant filter in code. Layer 3. The tenant ID comes from the authenticated session. A prompt that says "only this customer's data" isn't a control.
7. 7.Every model call goes through one gateway. Layer 4, model gateway. Our code talks to one interface. The gateway holds the EU-region endpoints, the quotas, the retries, and the model routing.
8. 8.The feature still works when the model is slow or down. Layer 4. A queue, a timeout, and a rules-based fallback. A failed model call delays a suggestion and never blocks the core product.
9. 9.An evaluation set from real cases scores every change. Layer 5, quality checks. It runs in CI on each prompt or model change, and a lower score blocks the merge.
10. 10.A named person approves the outputs that matter. Layer 5. Anything that moves money, changes a record, or reaches a customer waits for a human or a deterministic rule.
11. 11.Every request leaves a trace we can query. Layer 6, operate safely. The request ID, the model and prompt version, the retrieved sources, and the approver, in a log the feature can't edit.
12. 12.We know what one AI action costs, per customer. Layer 6. Tokens, model, and cache hits per request, joined to the tenant. Pricing and plan limits rest on that number.

The three patterns from the talk are the same six layers, with a different pair under the most pressure.

I highly recommend bringing your score and the one AI feature on your roadmap to a thirty-minute call. I'll show which layer I'd expect to break first, and whether the problem is the one it looks like. Which layer scored zero for your team?

These takeaways are an engineer's reading of the rules and the tools. I wrote them to help teams decide what to build. They aren't legal advice. I recommend checking the current text with your own counsel before planning around a date or a threshold.