Hello,

Eight field notes went up on Navata Insights in August, across AI evidence architecture, a three-part series on AI reliability, Veeva delivery authority, and the opening piece of a new series. All eight were circling the same question:

When the evidence stays green, the audit trail reads clean and the test still passes, is the decision underneath any of that still true?

The month opened with two pieces on where AI-generated content actually becomes a GMP record and what has to survive afterwards. Conventional audit trail review catches a visible change to a record. It was never built to catch a person adopting AI-generated content unchanged, because that moment produces no change for the review process to see. And even where that adoption is captured properly, as a distinct event with its own attributable actor, the evidence explaining it doesn't necessarily live as long as the GMP record itself, because the model provider, the QMS and the integration layer are rarely governed by the same retention schedule. I ended up naming that gap the evidence-retention envelope: the minimum set of records, metadata and relationships that has to stay intelligible for as long as the record it explains needs defending.

From there, a three-part series asked what it actually takes for Quality to trust an AI agent, rather than just watch it work. Capability is evidence the system can perform a bounded task. Confidence is evidence that justifies Quality actually relying on it. Maturity is the organisation's capacity to correct the system later without losing the evidence trail behind the reliance it's already been granted. Those are three separate claims, and confusing them in either direction is expensive. The clearest recurring failure across all three parts is what I kept calling the Legacy Precedent Trap: an agent applies a historical decision correctly, exactly by the rules it was given, and still gets it wrong, because what counts as acceptable precedent has moved on since that decision was made. One of the six categories in the diagnostic framework behind this, the Navata Quality AI Failure Map, sits entirely outside the AI system: Quality Intent, the organisation's own decision logic that experienced people apply but nobody ever wrote down. The series closes with the Correction Path, four steps for the moment a correction actually needs to happen: classify whether it's material, attribute it to the right layer, propagate the decision everywhere it needs to exist, and re-anchor the evidence that justified reliance in the first place.

The same ownership gap reappeared in delivery. A recognised implementation partner and a tightly negotiated SOW can still leave nobody holding independent authority to reject a design that technically satisfies the contract while weakening the architecture underneath it. A good partner can propose that authority. Structurally, they can't hold it on the client's own behalf, because it means being able to reach a conclusion that works against their own delivery plan.

The month closed on the piece that I think sets up where this goes next, and the opening entry in a new series. A design can pass every test and keep a fully green traceability matrix, and still stop being a decision the organisation can defend, because something it depended on, entirely outside the configured system, quietly changed. I'm calling that a decision validity condition.

A passing test and a still-true decision are not the same claim.

August's field notes:

• The AI Audit Trail Illusion: The Exception Nobody Configured

• The Complete AI Record Does Not Exist in One System

• Your AI Agent Went Live in Four Weeks. What Exactly Went Live? (Time to Capability, Part 1 of 3)

• The Agent Works. Can Quality Trust It Yet? (Time to Confidence, Part 2 of 3)

• AI Doesn't Mature With Time. It Matures Through Quality Intent. (Time to Maturity, Part 3 of 3)

• Buying Veeva Is Easy. Buying the Wrong Implementation Is Expensive.

• The Configuration Still Passes. Does the Rationale?

• The Architecture of a Workaround

  
  

If AI evidence architecture is your immediate concern, begin with "The AI Audit Trail Illusion." If you're working through Veeva delivery authority, begin with "Buying Veeva Is Easy. Buying the Wrong Implementation Is Expensive."

Part 2 of "Where the Evidence Breaks" is next.

If this was useful, forward it to a colleague who'd want it too.

Rohith

Founder, Navata Ltd

https://navata.ai