Human in the Loop in Pharma: Why AI Ends With a Signature

Human in the loop is the most used and least explained phrase in AI for pharma. This post explains it in plain words, for QA Heads first and for everyone else who works under good manufacturing practice (GMP) rules. The short version is in the box; the long one runs through 21 CFR 211.22(c), the FDA’s first AI warning letter, and draft EU GMP Annex 22, and it ends, every time, at a signature.

On 2 April 2026, the U.S. Food and Drug Administration published Warning Letter 320-26-58 to Purolea Cosmetics Lab, a small Michigan manufacturer of over-the-counter drug products. Investigators had found the firm using AI agents to generate its specifications, standard operating procedures (SOPs), and master production and control records, all put into use without quality unit review. Asked why process validation had never been performed, staff said the AI agent never told them it was required. FDA rejected that answer, and the firm has since ceased drug production. No new AI rule was needed. A quality unit clause from 1978 was enough.

Here is the argument this post defends: in a GMP plant, AI output is a draft until a qualified person has reviewed it and signed. The machine drafts, suggests, and flags. A person reviews, decides, and signs. Regulators on both sides of the Atlantic are converging on that line, and the signature is where they draw it.

What does "human in the loop" actually mean in GMP manufacturing?

Human in the loop means a qualified person sits between the machine’s output and anything the plant acts on. The AI proposes; the person disposes. In GMP terms: no AI-generated specification, procedure, or record has any standing until someone with the training and the authority to reject it has reviewed it and signed for it.

In vendor decks the phrase has picked up a loose life, sometimes meaning little more than a person being somewhere in the building. The working definition is stricter, with three tests. The person must be able to see what the machine did, including where its inputs came from. The person must be able to stop it: change or reject the output, on a schedule that assumes rejections happen. And the person must sign for it, taking named accountability. See it, stop it, sign it. Fail one, and you have a human nearby, not a human in the loop.

Takeaway: if the person can’t see it, stop it, and sign for it, the loop is marketing.

What does the machine do and what does only a person do?

The cleanest way I know to hold the whole concept is six verbs. Call it the Six-Verb Boundary. The machine gets three: draft, suggest, flag. The person keeps three: review, decide, sign. Any AI feature you’re shown, in any system, sorts onto one side of that boundary in about ten seconds, and everything regulators have written so far respects it.

On the machine’s side: drafting a batch manufacturing record (BMR) section from the master formula record; suggesting an SOP cross-reference or a deviation category; flagging every field where a draft differs from its master. On the person’s side: reviewing against the effective master, deciding what a discrepancy means, and signing — the verb the other five exist to serve.

The boundary, applied across the batch documentation lifecycle:

Takeaway: a demo where the machine performs a verb from the third column isn’t innovation. It’s the Purolea condition.

New to this series? Start at the pillar:
AI in pharma documentation for small and mid-size QA teams.

Has a regulator actually enforced this?

Yes. Warning Letter 320-26-58, issued 2 April 2026, is the first FDA warning letter to carry a dedicated section on inappropriate use of artificial intelligence in pharmaceutical manufacturing. The citation underneath it wasn’t new. It was 21 CFR 211.22: the quality unit’s responsibility over specifications, procedures, and records.

Purolea had used AI agents to help it comply with FDA regulations; the agents’ output (specifications, procedures, master production and control records) went into use without further review. FDA’s position was direct: “you must review the AI generated documents to ensure they were accurate,” and skipping that review violates 211.22(c). The second finding is the one I’d put on a slide. The firm hadn’t performed process validation, required under 21 CFR 211.100, and explained that the AI agent had never identified the requirement. FDA rejected the explanation. Reliance on a tool does not relieve a firm of knowing the rules.

“A warning letter never names the model. It names the firm that let a draft become a record without review.”

Takeaway: the precedent isn’t that AI is dangerous. It’s that accountability never moved.

What do the new AI rules expect from human oversight?

Both major regulators are writing the same expectation from different directions. In the EU, draft Annex 22 to EudraLex Volume 4 (published for consultation on 7 July 2025 alongside a revised Annex 11 and Chapter 4; consultation closed 7 October 2025) is the first GMP text written specifically for AI. As of August 2026 it remains a draft; the European Medicines Agency (EMA) ran a workshop on 30 June and 1 July 2026 to shape the next version. Its architecture is already clear. Critical GMP applications are confined to static, deterministic models. Generative AI and large language models (LLMs) should not be used in critical applications at all. And in non-critical applications, qualified personnel are expected to confirm outputs are suitable for intended use.

The U.S. side rhymes. FDA’s January 2025 draft guidance on AI in regulatory decision-making, still draft after comments closed in April 2025, builds a seven-step, risk-based credibility framework around each model’s context of use, informed by 500-plus AI-containing submissions to CDER since 2016. In January 2026, FDA and EMA jointly published ten guiding principles of good AI practice, human-centered design among them. And EFPIA’s 2024 position paper says the quiet part plainly: the extent of the human in the loop on critical steps is a risk decision, and it sets the validation load.

Takeaway: the direction of travel is one-way. The more critical the application, the more human the loop.

Where does a draft legally stop being a draft?
The deep dive:
When does an AI-generated draft become a GMP record?

Why does it always end with a signature?

Because the signature is where accountability attaches: legally, procedurally, in front of an inspector. Under 21 CFR 11.50(a), a signed electronic record must display three things: the signer’s printed name, the date and time, and the meaning of the signature, such as review, approval, responsibility, or authorship. Meaning is the load-bearing word. A model can produce text. It cannot mean anything by it.

Part 11 goes further. 11.100(a) requires each electronic signature to be unique to one individual, never reused, never reassigned. 11.70 binds the signature to its record so it cannot be lifted and moved. Every link in that chain assumes a human being with a name, a role, and something to lose. That is why every AI workflow in a GMP plant terminates at the same event: a person, reading a draft, deciding it is right, and signing it into existence as a record. Everything before that moment was a draft. Everything after it is the plant’s word.

“In ten years of audits and vendor qualifications, I’ve never heard an inspector ask which software drafted a document. They ask who reviewed it and who signed it. The tools have changed three times in my career. The questions haven’t changed once.”

— Kinjal Patel, Senior Consultant — QMS, Validation & GMP Compliance

“Human in the loop isn’t a checkbox in a vendor deck. It’s the difference between a draft and a record.”

How a draft crosses that line mechanically  page states, watermarks, what QA actually signs  is its own territory, and this post stays out of it. Start with when an AI-generated draft becomes a GMP record and how QA release review turns an AI draft into a signed record.

Takeaway: the signature isn’t the formality at the end of the workflow. It’s the destination the workflow exists to reach.

When does "human in the loop" fail?

Three ways, and none of them is a model failure.

1. The rubber stamp.
   Automation bias, the habit of over-trusting automated output, is one of the best-documented effects in human-factors research; Goddard, Roudsari and Wyatt’s 2012 systematic review in the Journal of the American Medical Informatics Association tied it to reduced vigilance exactly where review matters most. Industry discussion of AI in GMP has named the same drift: when outputs are consistently good, verification quietly erodes. The defense is structural: design review as interrogation, with flags to clear and differences to explain, not a re-read of clean-looking text.
2. The loop with no authority.
   A reviewer who can’t reject the output, or gets no time to, is decoration. If the release calendar assumes zero rejections, the loop exists on paper. Real loops budget for the word “no.”
3. The loop nobody defined.
   No SOP names who reviews AI output. No training covers it. No signature meaning maps to it. This was Purolea’s condition: the failure wasn’t review done badly. It’s that no quality process existed for the AI’s output to enter.

Takeaway: a loop you can’t audit is a loop you don’t have.

I’ve reviewed batch records for a decade. When something goes wrong (a deviation, a complaint, an audit finding) the first question in the room is always the same: who signed this? Not which template, which system, or, lately, which model. The pattern is always identical: the plants that struggle aren’t the ones with the oldest tools; they’re the ones where that question takes a meeting to answer. The fix is rarely a smarter tool. It’s a loop where the answer takes one second.