The Question Every Inspector Will Eventually Ask

AI features are arriving in regulated software faster than governance frameworks can classify them. Sooner or later, every life sciences company deploying them will face a version of the same question from an auditor or health authority: ‘This content was generated by AI — who is accountable for it, and how do you know it is correct?’

Most vendors answer with adjectives: responsible, trustworthy, human-centric. We think the defensible answer is architectural, and it fits in one sentence: AI output is always a draft; only a human can promote content into the record. Everything else in our AI design — provenance, guardrails, signatures, triage queues — is machinery for enforcing that single boundary. This post lays out how it works, because we think the pattern matters more than the product.

Why the Draft/Record Line Is the Right Line

GxP regulation is, at its core, about records: who created them, when, why, and whether they can be trusted. 21 CFR Part 11 and ALCOA+ do not prohibit assistance in creating records — regulatory professionals have always used templates, precedents, and medical writers. What the framework requires is that accountability for the record is attributable to a person and that the record’s history is transparent.

Drawn this way, the boundary resolves the AI governance problem cleanly:

  • AI can read, retrieve, summarize, suggest, and draft — activities where an error is caught before it matters.
  • Only an identified, authenticated human can approve, sign, release, or discharge — state changes where an error becomes a regulatory fact.

The corollary is that the interesting engineering is not in the AI at all. It is in making the promotion gate impossible to bypass.

The Machinery: Five Enforcement Mechanisms

1. Drafts are structurally draft

In the DnXT platform, AI-generated content — a cover letter draft, a suggested response to a health authority question, a proposed document classification — is created in a state the workflow engine treats as non-authoritative. It cannot be referenced by a submission, cited as an approved response, or trigger downstream obligations. This is not a UI label; it is a state machine constraint. There is no code path from ‘generated’ to ‘effective’ that does not pass through a human approval step.

2. Provenance travels with the content

Every AI-drafted document carries its origin with it: that it was AI-generated, from what inputs, and under what approval it entered the record. Exported drafts carry a visible provenance footer. The audit trail records the generation event, the retrieval sources that grounded it, the human who reviewed it, and the edits made between draft and approval. When the inspector asks ‘how was this created?’, the answer is a record, not a reconstruction.

3. Anti-fabrication guardrails on the generation side

The draft boundary does not excuse sloppy generation — a reviewer drowning in plausible fabrications will eventually miss one. Our document intelligence features generate from retrieved, cited sources rather than open-ended invention; drafts that require facts the system does not have get explicit placeholders demanding human input, rather than invented values. An automated quality pass checks drafts for unresolved placeholders and unsupported claims before a human ever sees them — AI checking AI, with the human still owning the verdict.

4. Signature-grade gates for consequential state changes

Where promotion into the record carries regulatory weight — approving a controlled document, closing a quality issue, signing off a review — the human gate is a Part 11 two-component electronic signature: re-authentication, recorded meaning, cryptographic binding to the exact content signed. The signature seals the specific version the human saw. If the AI draft changes afterward, the signature visibly no longer covers it.

5. Automated ingestion lands in quarantine

The boundary also governs data entering the system without a human author. Our correspondence tracking can ingest inbound agency email and auto-attribute it — which application, which agency, what interaction type. Attractive automation, and a perfect vector for silent errors. So nothing auto-ingested becomes authoritative on arrival: items land in a proposed state in a triage queue, and a human confirms attribution before the record exists as a record. The pattern generalizes: automation proposes; people promote.

The Objection: Doesn’t This Throttle the Value?

The pushback writes itself: if a human must review everything, have you just moved the bottleneck? Two answers from operating experience.

First, reviewing a well-grounded draft is dramatically faster than authoring from blank. The human hour shifts from assembling boilerplate to exercising judgment — which is both faster and a better use of scarce regulatory expertise.

Second, the boundary is precisely what lets you deploy AI broadly. Because no AI output can become a record unnoticed, the blast radius of a model error is bounded by review. Organizations that skip the boundary end up in one of two bad places: quietly trusting unreviewed output (until an inspector finds the fabricated citation), or banning AI outright after the first scare. The draft/record line is what makes the sustainable middle path available — consistent with the platform’s broader security and compliance posture.

A Test You Can Run on Any Vendor

Evaluating AI features in regulated software? Ask four questions:

  • Can any AI output reach an effective/approved/released state without an identified human action? (The only acceptable answer is no — enforced by the system, not by policy.)
  • Where do I see, per document, that it originated from AI and who approved it?
  • What happens when the AI does not know a fact — placeholder or plausible invention?
  • What state does automatically ingested data land in?

Vendors with a real boundary answer in mechanisms. Vendors without one answer in principles.

The Point

AI in GxP does not need less ambition; it needs a hard line in the right place. Draft versus record is that line: machines accelerate everything up to it, humans own everything past it, and the audit trail proves which side every action happened on. That is an answer an inspector can accept — and an architecture a regulatory team can actually build on.

To see the boundary working end to end — generation, provenance, approval, signature — talk to a regulatory expert.