GMP validation with continuous AI evidence — IQ/OQ/PQ alive, with no sprints before every audit.

The problem

Validation as an event: sprints, paper, and the uncertainty of whether the validated state is still there.

The QA manager of a pharmaceutical plant knows the cycle: when a line is qualified or a change is introduced, a validation project is set up that runs for months. IQ, OQ, PQ, signed protocols, PDF reports, file into the cabinet. And between one validation and the next, the "validated state" rests on a declaration from the quality system — not on continuous evidence that the process is still inside the validated ranges.

1. The initial IQ/OQ/PQ is done properly — a dedicated validation team, protocols in a specialist tool, RP signatures. Then it is frozen in the cabinet.
2. The quarterly change control — a PLC parameter changes, an equipment component is replaced, a new SKU is introduced — triggers a mini-project of partial revalidation. Each one gets its own sprint.
3. The FDA/EMA audit arrives — the QA team spends three weeks gathering evidence that the process has stayed in range between the PQ and today. The evidence exists (SCADA, BMS, LIMS), but it is scattered.

The result in a GMP plant: the real cost of validation is not the first one (that one gets budgeted) — it is the continuous work of keeping the validated state alive. And the QA manager's recurring feeling is: “if the FDA turns up tomorrow, can I prove in a week that the process has been compliant on every one of the last 365 days?”. The second edition of GAMP 5 asks for exactly that: continuous validation.

How it fits the IRIS system

iLEAN does not replace your quality system — it turns the validated state into an always-on layer.

The problem is not a lack of protocols or validation tools: there are plenty, and good ones. The problem is that the real process evidence that supports the validated state lives in separate systems (SCADA, BMS, LIMS, MES) and is only cross-referenced during the audit sprint. iLEAN acts as the putty that fills that crack, without replacing your eQMS, your DMS, your LIMS or your validation system.

Tracer keeps IQ/OQ/PQ linked to the live process. The Agent sees every deviation against the validated range before it affects the batch. Writer prepares the revalidation or audit dossier. The QA person signs — always, under 21 CFR Part 11.

The three iLEAN pieces applied to GMP validation:

• iLEAN Tracer — keeps the traceability URS → design → IQ/OQ/PQ protocol → real process evidence → batch produced alive. When something changes (a component, PLC software, a new SKU), Tracer identifies which protocol is affected and prepares the revalidation scope.
• iLEAN Agent — sees the critical process parameters (temperature, pressure, mixing, cycle) continuously, compares them with the ranges validated in the PQ, detects drift trends before a batch is at risk and alerts the QA manager through the right channel. It proposes; the person signs. The Agent does not release batches or authorize restarts on its own.
• iLEAN Writer — prepares the revalidation dossier when there is a change control, and the continuous evidence package when the FDA/EMA inspection arrives. Formatted against GAMP 5, EU GMP Annex 11 and 21 CFR Part 11. The QA team reviews and signs; the reconstruction sprint disappears.

See the full IRIS architecture →

Before and after

Classic sprint-based validation vs. continuous validation with iLEAN

Aspect	Classic validation (event + sprints)	With iLEAN Tracer + Writer + Agent
IQ/OQ/PQ	Frozen in a PDF at the end	Linked to the live process
Validated state between events	A declaration from the quality system	Continuous evidence under Part 11
Change control	A revalidation mini-sprint	Scope identified, dossier prepared
Detection of process drift	In the OOS after the batch	As a trend, before it affects the batch
FDA/EMA audit	3 weeks of QA team time	Package prepared, human review
CSV of new systems	A heavy project per system	CSA approach, risk-proportional

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• GMP plant (oral solids, sterile products, biotech) with 2-5 lines, active quarterly change control, periodic FDA/EMA audits, a QA team partly dedicated to gathering evidence by hand.
• Tracer + Writer + Agent pilot on one critical process (a granulation/compression line, a lyophilization cycle, an aseptic filler). First value expected within a few weeks.
• Reduction in manual QA effort spent gathering evidence: ≥30% in the first 6 months.
• Indicative payback between 4 and 9 months, driven by the cost of the QA team in validation sprints and, above all, by the avoided risk of a launch delay when a revalidation blocks a new SKU.

And the QA manager's reasonable doubt

“What if the AI invents a conclusion about a process parameter, or misses a real deviation?” — hallucination is a problem of free generation, not of anchored tasks. When the AI merely reads the real data from SCADA/BMS/LIMS and compares it with the range validated in the PQ (pure recontextualization, one of the tasks AI performs best at), the best models brought error below 1.5% ^[1]. And even so, the Agent proposes; the QA person signs, under 21 CFR Part 11. The three safety rings protect critical OT — batch release is not done by an AI.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.