destill.ai · All sectors · Deutsch · Trust Center
For Medical Affairs, medical information and pharmacovigilance at large pharmaceutical companies
Hundreds of enquiries by morning, sorted by evening. Who shows how each answer came about?
In the film’s fictional day, an AI agent reads each enquiry, checks it against approved sources, drafts in the right tone and passes only the uncertain cases to a person. Proof of Agent Work gives every one of those steps a signed receipt that your QA and auditors can check without us. Large pharma teams add questions about GxP, audit trails, fabricated images and leaks; the section below takes each one and says what we cover and what we do not.
Try it yourselfRequest a pilotBook a call
See also: Healthcare and pharma: QC steps, audit trails, sealed images
Medical information
An answer to a healthcare professional must rest on approved sources. Afterwards you need to show which ones, in which version.
Pharmacovigilance
If an enquiry contains a suspected adverse reaction, it must not get lost. You want to show when it was flagged and to whom it went.
Compliance
Medical answers must not turn into promotion. A trail of what the agent drew on helps you show where the line was.
Where large pharma teams are stuck with AI today
Public regulatory drafts, industry roadmaps and research-integrity studies, with the source on each card. For every problem we say what Proof of Agent Work and Fortress cover today and what they do not.
AI answers need an audit trail
Medical Information teams are asked whether they can show a complete audit trail for AI-assisted answers. Hallucinated content and the spread of off-label information are named as the main risks, and accountability stays with the human team.
What we cover today: A receipt per step names the approved source document and version the agent drew on, the model and its version, and the person who approved the answer.
What it does not cover: Whether the answer is medically right, and whether it is information or promotion. That stays with your reviewers.
Sources: Medical Affairs roadmap (PDF) · IntuitionLabs: GenAI in Medical Affairs
EU GMP Annex 11 and Annex 22 (drafts)
The draft revised Annex 11 and the new Annex 22 on AI went to consultation until 7 October 2025 and are not in force. They point to stricter audit-trail expectations for GMP computerised systems and, for critical GMP uses, to static, deterministic models with a human in the loop.
What we cover today: A signed, independently time-stamped record of what each agent step did, which your QA team can check offline.
What it does not cover: It is not a validation and does not make a model static. Annex 22 is about GMP manufacturing, not every Medical Affairs workflow; whether it applies is your assessment.
Sources: gmp-compliance.org: Annex 22 draft · European Commission: consultation page
21 CFR 11.10(e): audit trails
Part 11 expects secure, computer-generated, time-stamped audit trails for operator actions that create, modify or delete electronic records.
What we cover today: A hash-chained, time-stamped record of agent steps in which any later change breaks the check. The pilot adds a mapping document as a working basis for QA and RA.
What it does not cover: It does not replace the audit trail of your validated system (for example the eQMS) or the Part 11 rules on electronic signatures, and it is not a validation.
Sources: eCFR: 21 CFR 11.10
Model changes and context of use
The FDA draft guidance of January 2025 on AI in regulatory decision-making asks for a credibility assessment per context of use and for documenting changes that could affect model output. Its scope is AI that generates data for decisions on safety, effectiveness or quality.
What we cover today: Every step records model and version, so a model change shows up in the trail.
What it does not cover: The credibility assessment itself, and anything outside that draft’s scope.
Sources: gmp-compliance.org: FDA draft guidance on AI · HHS: FDA draft guidance
Pharmacovigilance and human oversight
The CIOMS XIV work on AI in pharmacovigilance scales human oversight to risk; regulators expect a human in the loop for case processing. A suspected adverse reaction must not get lost.
What we cover today: A receipt shows when the agent flagged a case, to whom it went, and the signed approval by the person who decided.
What it does not cover: Case coding, reporting deadlines and submissions. We do not file or track reports.
Sources: Qualio: CIOMS XIV summary · ProPharma: AI in PV
Fabricated study and microscopy images
Generative AI can produce realistic western blots, and in a 2025 study free AI detectors performed poorly on them. Paper mills are named as the main concern. It matters for your own labs, for due diligence and for in-licensing.
What we cover today: Fortress SEAL: a photo taken live in our camera app is sealed at capture; after any change the seal breaks and the check shows where. Useful for images your own labs capture from now on.
What it does not cover: Archives, uploaded and third-party images cannot be sealed, there is no retrospective detection, and the SEAL never says a picture is “not AI-generated”.
Sources: PeerJ: AI detectors and western blots · arXiv: explainable artifacts for synthetic western blots
Leaks and IP tracing
Pre-launch documents, dossiers and training material go to many recipients. If a copy turns up, you want to know which copy it was.
What we cover today: Fortress TRACE: one invisibly marked copy per recipient; a found copy points to the recipient it was sent to.
What it does not cover: A mark can be removed with effort, and no hit is not proof of no leak. It is evidence for a conversation, not a verdict.
Model and data provenance
Auditors ask which model produced which answer, in which version, from which sources.
What we cover today: The receipt records model, version and the source documents named in the step.
What it does not cover: Where a model’s training data came from, and whether its output was right. A receipt covers what was logged.
Estimate or check: probabilistic versus deterministic
Deepfake detectors estimate
A detector looks at a finished picture and returns a probability. It can be wrong in both directions, and every new generator model makes its job harder.
Fortress SEAL checks
A photo taken live in our camera app is sealed at the moment of capture. The check compares the photo with that seal and answers yes or no: changed since capture, and if so, where. Existing or uploaded photos cannot be sealed, and the badge says "Captured live & unaltered", nothing more.
POAW: a receipt per agent step
Every step of an AI agent gets a receipt: signed with ML-DSA-65, time-stamped per RFC 3161, checkable offline with OpenSSL. It supports the evidence for the AI Act logging duties (Art. 12, 19 and 26(6)); it does not make anyone compliant.
Planned, not built: one evidence chain from the sealed photo to the decision. Pilot partners wanted.
The film
The opening scene, as a short clip.
The full explainer
Chapters
Try it yourself: a signed receipt you can check
This is the kind of receipt you would get from every run of your agent after the pilot. Synthetic data, real cryptography: an ML-DSA-65 signature and an RFC 3161 time stamp from an independent authority. Check it in your browser, then check a copy where one value in step 3 was changed.
Visual quality-control step (fictional batch)
| Step | What the agent did | Time (UTC) |
|---|---|---|
| 1 | Batch receivedbatchId: B-26-0417 · product: tablet 50 mg (fictional) · line: L3 | 06:15:00 |
| 2 | Images capturedstation: inspection camera 2 · images: 3 × hash | 06:15:42 |
| 3 | Defects classifiedmodel: vision-qc 1.7 · defectsFound: 2 · classes: chipped edge, spot · confidence: 0.94 | 06:15:44 |
| 4 | SOP limit checkedsop: SOP-QC-112 v5 · acceptLimit: 3 · result: within limit | 06:15:44 |
| 5 | Disposition proposeddisposition: forward to QA review | 06:15:45 |
| 6 | Human approvalrole: QA reviewer · approver: hash | 07:02:19 |
| 7 | Record archivedsystem: eQMS · record: hash | 07:02:21 |
Check this receiptCheck a tampered copy
Download the receipt (JSON) · Or drag the downloaded file into the verifier yourself, edit one number in a text editor, and drag it in again.
Your engineers can repeat the check offline with OpenSSL 3.5 or later and the published script: check pack.
What you get
Each item carries its status from our feature register: Built, Pilot or Roadmap. Hover a label for the evidence behind it.
One signed receipt per agent run: every step, in order, with model and version. Post-quantum signature ML-DSA-65 (FIPS 204). Time stamp from an independent RFC 3161 authority; nobody can backdate the receipt, us included. The human approval signed with the reviewer’s own key. Fortress TRACE: one invisibly marked copy per recipient; if a copy turns up, you see whom it was sent to. AEGIS: a prompt-injection screen on CPU, before the model. Self-reported figures only; no public reproducible re-run yet, so this page prints none. Fortress SEAL (images only): a photo sealed at live capture; after any change the seal breaks and the check shows the area.
The legal frame
Where a receipt helps you evidence an obligation. It does not fulfil the obligation for you and does not make anyone conform with the law by itself. Not legal advice.
AI Act Art. 12, 14 and 26
Logging and human oversight for high-risk AI. Whether your use case is high-risk is an assessment for your legal team. Where it is, a receipt helps you evidence that a log existed unchanged at a fixed time and that a person decided.
Source: eur-lex.europa.eu
AI Act Art. 50
Marking of AI-generated or manipulated content. Fortress marks can help you evidence that content was marked. Whether and how Art. 50 applies to you is a legal assessment.
Source: eur-lex.europa.eu
Directive 2001/83/EC, Art. 86 to 88
Rules on advertising of medicinal products. The receipt shows which approved source an answer drew on; the line between information and promotion stays your call.
Source: eur-lex.europa.eu
EMA GVP Module VI
Collection and reporting of suspected adverse reactions. The receipt helps you evidence when the agent flagged a case and to whom it was escalated.
Source: www.ema.europa.eu
GDPR Art. 9
Health data are special-category data. The receipt stores hashes, not the data themselves.
Source: eur-lex.europa.eu
The honest limit
A receipt proves who signed which steps, and when. It does not prove that the agent was right, and it does not prove that nothing was left out: it covers what was logged. The time stamp proves the receipt existed before the stamped time, not when the actions happened. The SEAL shows a change made after capture; it cannot show that a picture was already wrong before it was sealed, and it never says a picture is "not AI-generated". Not legal advice.
Planned, not built yet
What comes next, by demand and investment. Nothing here is part of the pilot.
One evidence chain across systems, e.g. photo → agent → decision. Keys in hardware (HSM). - Time stamp from an authority in your own house, so that signing needs no external connection.
- A hosted interface (API) and a validation package for regulated environments.
- Fortress SEAL in a dedicated smartphone app (a ProofCam demo exists). A stronger proof of origin would come from camera makers embedding the resilient and the fragile mark in the hardware itself. We want to build this with them; it depends on investment and demand.
The pilot: scope and duration set together, price after scoping
Scope
One enquiry workflow, one team, one product area. Runs in your network or cloud tenant. POAW receipts for the agent steps and the human approval; optionally TRACE for documents sent out. Final report with go/no-go. Your own implementation partner can connect it: we hand over the SDK and the sandbox set-up (one machine, Docker). The time frame is fixed in scoping.
Price frame: Price after scoping, based on your process and environment.
Success criteria, agreed in writing at kick-off
- At least 99 % of AI steps in the pilot workflow produce a verifiable receipt.
- Your team verifies a receipt on its own, without our help (OpenSSL).
- Extra effort per work step is measured; we set the target together at kick-off.
Source: our written pilot offer; the targets are what we commit to measure, not results from an earlier pilot.
What you provide
- One concrete workflow and an owner
- Test data or test images and access to a test environment
- Contacts from IT and compliance or audit
- A connection to a time-stamp authority, at signing or later (signing itself works offline)
- Your own legal assessment
- Approved source documents and test enquiries without real patient data
Want the written pilot offer? Choose "Talk about a pilot" below and we send it with our reply.