The 2026 rules for AI in pharma quality are here

When I joined Veriqa as one of the co-founders, I’d just spent a few years shipping AI products at AWS. I thought I knew what “production-grade AI” meant. Pharma quality taught me I didn’t.

AWS taught me to ship fast and monitor in production. If something went sideways, you’d see it in your dashboards within minutes and you’d patch it. The whole engineering culture was built around the assumption that you could iterate your way to correctness.

Pharma quality manufacturing is the inverse of that culture. You don’t get to “iterate to correctness” when a regulator is in the room. You don’t get to “ship now, fix later” when the cost of a mistake is a recall, a 483 letter, or a lost manufacturing license. The discipline is older, slower, and — once you understand it — far more rigorous than what most software engineers ever encounter.

So when we started building Veriqa, the question wasn’t “can we make an LLM draft an APQR?” That part is comparatively easy. The question was: how do you build an AI-assisted quality system that an FDA or EMA inspector will accept — not just today, but five years from now, when the regulators have caught up to the technology?

In the last twelve months, the regulators have caught up. Almost nobody is talking about what that actually means.

The line that changes everything

In July 2025, the European Commission published a draft of GMP Annex 22 — the first dedicated regulatory guidance on AI in pharmaceutical manufacturing. The final version lands in 2026. Enforcement begins the year after. There is one line in it that should be on every pharma quality leader’s screen.

Generative AI, large language models, and continuously-learning models are not permitted for GMP-critical decisions.

Read it again. The regulator hasn’t said “AI is bad.” It has said something more precise: the kinds of models that have flooded the pharma quality market for the last twenty-four months — the “ChatGPT-for-APQRs,” the “GPT-powered deviation triage,” the “AI-assisted CAPA” pitches — are now structurally on the wrong side of the line for any decision that matters under GMP.

If you’ve sat through a dozen vendor demos in the last year, you’ve probably watched a lot of them just point an LLM at your data and call it “AI for quality.” Most of those vendors built their entire product around a single model API. When Annex 22 finalizes, those products won’t be quietly upgraded — they’ll need to be re-architected.

And this isn’t a hypothetical concern. The new EU GMP Annex 11 revision — also finalizing in 2026 — quadruples in length and explicitly extends to AI/ML and cloud systems. The FDA and EMA jointly released ten Guiding Principles of Good AI Practice in January 2026, with human oversight, transparency, and risk management as the spine. The ISPE GAMP AI Guide — the closest thing the industry has to a validation playbook — runs 290 pages and treats AI as a lifecycle problem, not a feature.

The direction of travel is unambiguous. AI is welcome in pharma quality. Generative-AI-as-the-decision-maker is not.

What a system built for this future looks like

When we were designing Veriqa, we kept coming back to a principle our design partner reinforced for us: a regulator doesn’t care that you used AI. They care whether your process stayed in control and whether you can prove it. Once we internalized that, every architectural decision got easier. Five of those decisions turned out to matter most.

1. The model is a swappable component, not the system of record

The most counterintuitive lesson from the last year: a validated workflow can’t be tied to whichever LLM is winning this quarter. Model providers raise prices. Models get deprecated. Regulators reclassify what’s permitted. Any system whose compliance posture depends on a single model is structurally fragile.

We architected Veriqa from day one around the assumption that the underlying model layer is interchangeable. The validated workflow lives one level above it. When the model behind the system changes, the process doesn’t. When a new regulation lands — Annex 22 included — we don’t have to re-validate end-to-end. Model risk is insulated from process risk. That’s not a marketing claim; it’s an architectural choice we made because we thought the regulators would eventually require it. They have.

2. Each decision boundary is deterministic

Annex 22 draws a sharp line between AI assists and AI decides. A generative component can help a human reach a decision faster — that’s permitted, with the right architecture. But the output that crosses a regulatory decision boundary has to be reproducible, validated, and traceable. That’s the part most vendors are about to discover they don’t have.

Our pipeline treats every transition between stages as a validated gate. The work that happens inside a stage may involve generative components; the output that leaves the stage is constrained, schema-locked, and produces consistent downstream behavior given the same inputs. The model assists the human inside the gate. The human signs at the gate. The next stage gets a deterministic input.

This is what regulators actually mean when they say “AI as a tool, not a decision-maker.” It’s not a posture. It’s a property of the system.

3. Every claim traces to a source. No exceptions.

21 CFR Part 11 requires complete and tamper-evident records. EU GMP Annex 11 requires audit trails for every GMP-relevant change. The new ISPE GAMP AI Guide extends both expectations to AI-enabled systems by requiring traceable model lifecycle documentation. The shared spirit across all three is the same: a regulator should be able to follow any claim back to the data it came from, without you in the room to explain it.

We treat citations as load-bearing, not decorative. Every line of every report Veriqa produces carries a traceable link back to the batch record, deviation log, or stability study it came from. There is no path through our system that can produce an unsourced claim. If our agent gets something wrong, a reviewer sees exactly where the wrongness came from. If it gets it right, the audit trail is already in place. We didn’t build this for elegance. We built it because we couldn’t see how the system would otherwise survive an inspection.

4. Human-in-the-loop is enforced by the architecture, not by policy

The FDA/EMA Joint Guiding Principles released in January 2026 require human oversight. Annex 22 requires human-in-the-loop. The PDA Letter on AI and audit readiness, published in February, frames it more sharply: every AI suggestion should be advisory until approved by a qualified person.

None of that is controversial. The interesting question is what HITL actually means in practice. “HITL” can mean almost anything — from “a human clicks approve at the end” to “a human reads every claim and verifies the citation.” The first is approval theatre. The second is real review. The difference is whether the system surfaces what the reviewer needs to see, or just asks for a signature.

Our reviewer interface is designed to surface disagreement. Where the model’s confidence is low, where its draft contradicts a prior decision, where the source data is ambiguous — those are the moments we draw the reviewer’s eye. Approval is the easy case; disagreement is what the architecture is built around. A reviewer who only sees what’s likely correct is a reviewer who’ll eventually rubber-stamp something that isn’t.

5. Validation is continuous, not a one-time event

Both the new Annex 11 and the ISPE GAMP AI Guide reflect a shift away from validate at qualification toward validate across the lifecycle. That makes sense for AI-enabled systems in a way it doesn’t for traditional software: the inputs evolve, reviewer expectations sharpen, regulatory interpretations shift. A snapshot of compliance from twelve months ago doesn’t tell you anything reliable about today.

We designed Veriqa around the assumption that every reviewer interaction is data — a labeled signal we can use to monitor whether the system’s outputs are still aligned with what reviewers actually accept. When the rules update, we propagate the change through the validated workflow without re-touching the underlying model. The validation lives in the workflow. The model just does its job inside it.

A regulator doesn’t care that you used AI. They care whether your process stayed in control and whether you can prove it.

Four questions to ask any AI quality vendor — including us

If you’re a pharma quality leader evaluating AI vendors right now — whether you’re layering AI onto an existing QMS/eQMS or replacing the point tools around it — there are four questions that will tell you very quickly whether their system is built for where the regulators are going.

1. What happens to your validated system if you have to swap the underlying model? If the answer involves re-validating end-to-end, the model is the system of record. That’s structurally fragile, and Annex 22 is going to make it worse.
2. What’s the deterministic output of each stage in your pipeline? If they can’t answer this — if the answer is some version of “the LLM produces it” — the system isn’t deterministic at the gate. That’s an Annex 22 problem.
3. Show me how a single claim in a generated report traces back to its source. If they can’t trace it, it’s not auditable. If they can, watch how many clicks it takes. The faster the trace, the more inspection-ready the system.
4. What does your reviewer interface surface — what’s likely correct, or where the model is least confident? If it’s the former, you’re buying approval theatre. If it’s the latter, you’re buying a real second pair of eyes.

If you ask us those four questions in a working session, we’ll show you the answers in our system. We’ll also tell you the cases where our answer isn’t yet what we want it to be — because there’s no AI quality vendor in 2026 who has finished this problem, and any vendor who tells you otherwise should be your first signal to keep looking.

What’s coming, and what to do about it

Annex 22 is going to land. The Annex 11 revision is going to land. The FDA and EMA principles are already aligned. The vendor pitches that survive the next two years are going to be the ones built for what’s coming — and they will look meaningfully different from the AI tools that have been pitched to your team in 2024 and 2025.

The first post on this blog argued that the APQR is a symptom — that the pharmaceutical industry has a quality data problem its existing tools were never designed to solve. This is the regulatory companion to that argument. The systems that solve it have to be built for the rules that are coming, not the rules that were.

If your team is working through this — evaluating vendors, validating internal projects, or just trying to figure out where the line is — we’d like to hear what you’re stuck on. Send us your validation questions. We’ll send back what we’d ask if we were on your side of the table.

We promised an engineering deep-dive in our first post — that’s still in the works. We wanted to start with the question quality teams asked us first: what changes when the regulator catches up?