Translating ALCOA+ for AI: data integrity when a model writes the record

When a model writes part of a GxP record, ALCOA+ still holds, and the evidence you need to satisfy it gets much bigger. The record now holds the output plus the prompt, the model version, the timestamp, the source data and the named person who accepted it.

I've spent more than 25 years in pharma quality and IT, and I've watched data integrity come through paper, spreadsheets, validated systems, the cloud and the move from CSV to CSA. ALCOA+ came through all of it largely unchanged. AI is the first thing I've seen that tests the assumptions underneath it.

This piece grew out of a long conversation with the regulatory and validation team at BGO Software. They asked the questions well, so I've kept their shape and answered in my own words. Part 2 covers model drift, shadow AI and decision integrity, which is where the record ends and the decision begins.

The stakes are real. Data integrity is one of the most cited categories in FDA drug GMP warning letters, named in roughly 61% of the letters issued in 2021 by one European Pharmaceutical Review analysis. And AI is now drafting deviation narratives, summarising batch records and shaping quality decisions on that same ground.

Key takeaways

• The principles hold and the evidence gets heavier. AI asks you to show a great deal more to prove the same thing.
• Attribution becomes a chain: a person, a model version, a prompt and a dataset. The accountable human is the person who accepted the output.
• The original is a bundle of the output, the prompt, the model version, the timestamp and the source data.
• Accuracy is judged against a tolerance you set in advance for the intended use.
• Human review counts as a control only when the reviewer has the time, the information, the competence and the authority to say no.

Why does AI test assumptions that held for decades?

ALCOA+ lasted because it asks a human question: can I trust this record, and can I rebuild what happened? That question doesn't much care what technology sits underneath, which is why it carried us through paper, spreadsheets and validated systems without needing to change.

Underneath it sat a few assumptions we never said out loud. A record gets created once, by a person you can name, at a moment you can point to, and then it sits there unchanged. AI unsettles each of them.

• The author might be a model.
• The original is a bundle of the prompt, the model version and the data behind it, so there's no single thing you can hold.
• The output is probabilistic, so the same question asked tomorrow can come back a little differently.

The principles are as sound as they ever were. What has changed is how much evidence you have to put on the table to satisfy them.

How does the AI shift compare with paperless, cloud and CSA?

Every big shift in my career moved trust to a new place. Going paperless moved it from the wet signature to the audit trail behind it. The cloud moved it again, onto a line we now shared with a vendor. The move from CSV to CSA took it away from the sheer weight of documentation and towards honest thinking about risk.

AI moves trust into the reasoning itself. With everything before, the system did what we told it to, and our job was to prove that it had. With AI, the system reaches a conclusion of its own, and we have to show that the conclusion holds up.

The good news is that the discipline we built in the CSA years is what this moment needs: intended use, risk-based thinking, and testing the things that matter. A team that took the move from CSV to CSA seriously is a long way from starting cold.

I'll be honest about one part. This is the first shift I've lived through where the system can change its own behaviour without anybody raising a change request. That's new ground for all of us.

Which ALCOA+ principle is hardest to defend with AI?

Attributable is the hardest, with Original close behind. Accurate, the one everyone expects to struggle, turns out to be the most manageable.

Attributable gets hard because the honest answer to “who made this record?” is no longer one name. It's a person, a particular model version, the prompt that was used and the data the model learned from. Attribution becomes a chain.

Original is hard for a related reason. The source isn't one clean artefact any more, so you have to decide, deliberately and in advance, what your original will be.

If you made me pick one, I'd say Attributable. Accountability is what regulators care about most, and AI is the technology that smudges it.

Does ALCOA+ need a rethink, or a better interpretation?

A better interpretation, as long as it's real interpretation. The principles are about trust and reconstruction, and neither of those changes because the author happens to be a model.

A full rethink would throw away decades of hard-won understanding between industry and regulators, and it would signal that the old discipline no longer applies. It still applies. What we need is new evidence under the same principles, plus one honest addition on top: decision integrity, layered above data integrity. That adds a floor to the house and leaves the rest of it standing.

We've done this kind of translation before. When we went paperless, nobody rewrote Contemporaneous. We agreed that the audit-trail timestamp was the contemporaneous record and moved on. AI asks for the same kind of move.

What does translating ALCOA+ for AI look like in practice?

Ask one question of every principle: what would I show an inspector now?

None of this is exotic. It lives where good practice always has: the URS, the intended-use statement, the audit-trail design and the monitoring plan. CSA already gave us the skeleton, and we're hanging model-aware evidence on it.

To make it concrete: alongside the line “root cause: equipment fault”, the record now holds the prompt, the model version, the timestamp and the name of the analyst who accepted it. The conclusion on the page is the same, and you can stand behind it 2 years later when somebody asks.

What does attribution look like when AI contributes to a decision?

It becomes a chain of accountability. The word I'd underline is “contributes”, because the AI never owns the decision. A person always does.

At minimum, capture these 4 things.

1. Who ran it.
2. The input or prompt.
3. Which model, and which version, produced the output.
4. Who reviewed it and accepted it into the GxP process.

The accountable human is the person who accepted the output, who is often a different person from whoever built the model. Make that explicit, because that's the name that belongs on the record.

It goes wrong when an AI output drifts into a batch record with no thread back to the model, the version or the prompt. A record like that can't be attributed to anyone, and it won't hold up when an inspector pulls on it.

FDA has already put this in writing. In April 2026 it issued warning letter 320-26-58, the first citing AI use in drug manufacturing, to a firm that let an AI agent write specifications and procedures and took the agent's word for what the regulations required. The letter says that “any output or recommendations from an AI agent must be reviewed and cleared by an authorized human representative of your firm's QU.”

The simplest way to think about it is to treat the AI like a junior analyst's draft. The junior can write it, and a named, qualified person reads it, signs it and owns it.

When does human review actually reduce risk?

When the reviewer can catch the error, has what they need to catch it, and has a real route to say no. Take away any one of those and the review turns into a stamp.

It becomes a stamp the moment someone is signing off more than they could possibly read, or can't see the reasoning behind the output, or has no real authority to overrule it. A name in the loop on an organisation chart doesn't make a control on its own.

To make review mean something:

• Give the reviewer the inputs and a sense of how confident the model was.
• Tell them clearly what they're looking for.
• Make it easy to record a disagreement.

A review process where nobody ever overrules anything worries me, and I read it as a warning sign. A good deal of what passes for human oversight today sits closer to the stamp, and inspectors are getting better at telling the two apart.

Picture a model screening 400 records a day and one reviewer approving all 400 inside an hour. That person didn't read them. Real oversight looks like the model passing the dozen uncertain cases to a person who has the time to look at them properly.

How should accuracy be defined for probabilistic AI outputs?

As fitness for the intended use, within a tolerance you set before deployment. We already work this way in other corners of quality.

Set acceptance criteria before you go live and tie them to the risk of what the model does. A model drafting a summary for a person to review can live with far more uncertainty than one feeding an automated release decision, and the criteria should reflect that.

Then test against a known, representative dataset, state plainly the performance you'll accept, and keep watching that it holds. The useful questions are whether it's right often enough, in the ways that matter, and what happens on the occasions when it's wrong.

We never demanded a perfect analytical method. We set limits for accuracy and precision and validated against them. A probabilistic model is the same idea applied to a new kind of system.

What do Original and Contemporaneous mean when AI writes the record?

Both the prompt and the model version are part of the record. The record is the output plus enough context to rebuild how it came to exist.

• Contemporaneous means recorded at the time of the event. For AI the event is the generation, so the timestamp, the prompt and the model state all belong to that moment.
• Original means the first capture of the data. Keep only the tidy final output and you've kept the conclusion and thrown away the evidence behind it.

The test I give teams is easy to remember. If you couldn't regenerate the output, or at least explain exactly how it came about, 6 months later in front of an inspector, you haven't captured the original.

Take an AI that writes a deviation narrative. Keep only the paragraph, and a year from now you can't say what facts it worked from or which model wrote it. Keep the prompt, the version and the timestamp with it, and you can reproduce it and stand behind it.

Where this leaves ALCOA+

AI doesn't ask us to give up the principles that have governed pharmaceutical data integrity for decades. It asks us to read them against a longer chain of authorship, evidence and accountability. The record may now include the prompt, the model version, the source data, the timestamp, the acceptance criteria and the qualified person who accepted the result.

The next problem sits beyond the record. Once an AI output starts to influence a GxP decision, you also have to show that the model stayed in its validated state, the training data was governed, vendor changes were controlled and the decision itself was sound. The draft EU GMP Annex 22 on artificial intelligence has started to map that ground, alongside the draft revisions of Annex 11 and Chapter 4, and my reading of it is in what the Annex 22 draft asks of you. If you want the inspector's side, I've written up what an inspector asks about your AI.

Part 2 picks it up from there: decision integrity, model drift and shadow AI.

Sachin Bhandari runs TrustBridge Compliance, a senior-led advisory for AI in GxP governance, CSV to CSA and digital validation in pharma, biotech and CDMOs. He sits on the Global GAMP Steering Committee, contributed to the EU GMP Annex 22 (AI) industry review, and wrote Validating AI in GxP Environments: A Practitioner's Handbook, which works through the evidence for each risk tier. The AI Governance Stack shows where the decision record sits.

A version of this article appears on BGO Software's blog as What ALCOA+ Means When AI Writes the GxP Record. My thanks to BGO Software's regulatory and validation team for the conversation behind it.

Sources: European Pharmaceutical Review on FDA warning letters · FDA data integrity and compliance with drug CGMP · MHRA GxP data integrity guidance · FDA warning letter 320-26-58 · European Commission consultation on the Annex 22, Annex 11 and Chapter 4 drafts.

Questions leaders ask

Does AI make ALCOA+ obsolete?

No. The principles still hold. What changes is the amount and type of evidence needed to show them: the prompt, the model version, the timestamp and the named person who accepted the output all become part of the record.

Who is accountable for an AI-generated GxP record?

The named, qualified person who reviewed the output and accepted it into the GxP process. The model can't be accountable, and neither can the team that built it.

Is the prompt part of the GxP record?

Yes, together with the model version. Without them you keep the conclusion and lose the evidence behind it, and you can't reconstruct the record for an inspector.

How do you validate a probabilistic model in a GxP environment?

Set acceptance criteria tied to the intended use and the risk before deployment, test against a known, representative dataset, state the performance you'll accept, and monitor that it keeps holding.

Is a human in the loop enough to control AI risk?

Only when that person has the time, the information, the competence and the authority to overrule the output. One reviewer approving 400 records in an hour is a stamp.

Which regulations cover AI in GMP?

The draft EU GMP Annex 22 on artificial intelligence, published for consultation alongside draft revisions of Annex 11 and Chapter 4, together with the existing MHRA and FDA data integrity guidance and 21 CFR Part 11. The Annex 22 text is still a draft.