Skip to main content

​
The problem, as your Quality team sees it

AI agents are now being introduced into regulated biotech workflows — clinical data entry, batch documentation, trial monitoring, regulatory submission drafting. The vendors pitching them point at model accuracy and SOC 2. Your Quality team asks a simpler question:
“Can I prove, to an FDA inspector, who authorized this specific action, against what policy, and that the policy was active at that moment?”
Today, the honest answer from most AI vendors is no. They give you:
  • Training-data provenance (orthogonal — it’s about the model, not the action).
  • Prompt / output logs (evidence of what was asked, not what was authorized).
  • Access control on the UI (keeps people out — doesn’t apply to the agent itself).
None of these answer the auditor’s question. They all collapse the moment the agent is wrong — and agents are wrong constantly.

​
The missing layer

21 CFR Part 11, EU Annex 11, and ICH E6 GCP were written before AI agents existed, but their requirements translate cleanly:
  • §11.10(e) — secure, computer-generated audit trails on all record changes.
  • §11.10(g) — authorized signatories only for electronic signatures.
  • §11.50 — signature manifestation linked to the record inseparably.
The regulation calls for a layer between the agent and the record that:
  1. Decides whether the agent is authorized to take each action.
  2. Produces a signed, attributable authorization record.
  3. Preserves the record tamper-evidently.
  4. Makes the record retrievable on demand, verifiable offline.
That is what AtlaSent is.

​
The shape of the fix

1

Policy as code

Your Quality team authors a RACI policy in JSON. Actions an agent might take (healthcare.record.amend, data.export, trial.unblinding.execute) map to allow, deny, hold, or escalate with a reason citing the governing regulation. Policies are versioned; an active version is not edited in place, and a change publishes a new version.
2

Authorization at execution time

The agent calls AtlaSent before it takes the action. The decision is returned in milliseconds — one of allow (permit issued), deny (action cannot proceed), hold (waits for an external signal — typically an approval webhook), or escalate (routed to a human reviewer queue). On hold / escalate, an approver resolves it; the original evaluation then becomes either allow or deny.
3

Evidence by construction

Every decision writes a tamper-evident, hash-linked, Ed25519-signed audit entry. The chain exports as a self-contained bundle. FDA inspectors can verify it on an air-gapped machine with our source-open CLI — no AtlaSent staff involved, no dependency on our uptime.

​
Why “execution time” matters

You can have a policy document. You can have a design review. You can have a model eval. None of those answer the auditor’s question about a specific action taken on a specific Tuesday. Only a decision at the moment of execution, recorded tamper-evidently, does. This is the distinction AtlaSent is built around. The rest of the product (the Console, the SDKs, the MCP server, the GitHub Action) is just convenience — the core is the pre-execution decision and the signed record.

​
Next