Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Manufacturing and quality

AI in a GMP plant pays off only when it is validated for a defined intended use and measurably beats the manual process it replaces. Sort uses into critical and non-critical first, because regulators set a much higher bar of evidence for anything that directly affects product quality.

Part of AI for Life Sciences & Healthcare. 8 sources, checked .

of 1,309 FDA drug quality inspections in fiscal year 2025, including 61 by mutual recognition partners, were classified official action indicated
18%
of 1,309 FDA drug quality inspections in fiscal year 2025, including 61 by mutual recognition partners, were classified official action indicated
U.S. Food and Drug Administration, Center for Drug Evaluation and Research, Office of Pharmaceutical Quality, 2026(opens in a new tab)
of nearly 10,000 US drug applications and supplements needing a facility assessment (received fiscal years 2024 to 2025) drew a complete response letter due to a facility withhold, as of May 2026
12%
of nearly 10,000 US drug applications and supplements needing a facility assessment (received fiscal years 2024 to 2025) drew a complete response letter due to a facility withhold, as of May 2026
U.S. Food and Drug Administration, Center for Drug Evaluation and Research, Office of Pharmaceutical Quality, 2026(opens in a new tab)
initial Field Alert Reports on quality problems in marketed drugs received by FDA's Center for Drug Evaluation and Research in fiscal year 2025
4,059
initial Field Alert Reports on quality problems in marketed drugs received by FDA's Center for Drug Evaluation and Research in fiscal year 2025
U.S. Food and Drug Administration, Center for Drug Evaluation and Research, Office of Pharmaceutical Quality, 2026(opens in a new tab)

The problem

FDA's Center for Drug Evaluation and Research mapped the main uses in a 2023 discussion paper: process design and scale-up, advanced process control, process monitoring and fault detection, and trend monitoring.7 Its examples include models that watch equipment and trigger maintenance, vision systems that check packaging, labels or glass vials, and text analysis of complaints and deviation reports to find clusters of problems.7 The same paper noted that there are limited industry standards and FDA guidance for developing and validating models that affect product quality.7 Until that gap closes, each manufacturer has to build its own case that a model is fit for use.

Weak quality systems have a measurable cost. Of the 1,309 drug quality inspections FDA counted in fiscal year 2025, 18% were classified official action indicated.8 Of nearly 10,000 applications and supplements that needed a facility assessment, 28% had received a complete response letter by May 2026, and 43% of those letters (12% of all submissions) were due to facility withholds.8 The top inspection citations behind those withholds included failure to thoroughly review discrepancies or batch failures, and quality control unit procedures not written or followed.8 Investigations and quality unit oversight are the work that deviation triage and batch review tools have to improve, with evidence.

The European Union has drafted AI-specific GMP rules, and they are still a draft. From July 7 to October 7, 2025 the European Commission consulted on a new Annex 22 on Artificial Intelligence, together with revised versions of Annex 11 (Computerised Systems) and Chapter 4 (Documentation), drafted by the EMA GMDP Inspectors Working Group in cooperation with PIC/S.2 As of November 18, 2026 the Commission's EudraLex Volume 4 page still lists Annex 11 in its January 2011 revision and has no Annex 22.3 The draft applies to AI models in critical applications with direct impact on patient safety, product quality or data integrity. It covers only static models with deterministic output, and says that models that keep learning during use, models with probabilistic output, generative AI and large language models should not be used in critical GMP applications.1

In the United States the existing rules already reach AI. FDA states that AI in manufacturing must be implemented in accordance with current good manufacturing practice, and names 21 CFR 211.22 and 211.68 as applicable.6 Section 211.68 requires computers and related systems to be routinely calibrated, inspected or checked under a written program, and their input and output to be checked for accuracy.4 FDA's January 2025 draft guidance adds a seven-step, risk-based credibility assessment. Its manufacturing example, an AI vision system assessing vial fill level, is rated medium risk because release testing on a sample of each batch independently verifies fill volume.6 For device makers, FDA's final Computer Software Assurance guidance of February 3, 2026 says its risk-based approach can be applied to artificial intelligence and machine learning tools used in production or the quality management system.5

What we build

  1. 01

    Process monitoring and predictive maintenance

    Models on historian and sensor data that flag equipment and process behavior moving away from normal, frozen and version-controlled, with alerts routed to a named role and every response logged.

  2. 02

    Visual inspection model qualification

    We measure your current inspection performance first, build a locked test set covering each defect type and severity, and qualify the vision model against acceptance criteria your process experts approve before testing.

  3. 03

    Deviation and CAPA triage

    Classification and clustering of deviation and complaint text that surfaces similar past events and recurring causes. A qualified investigator decides the root cause, and every suggestion and decision is recorded.

  4. 04

    Batch record review by exception

    Checks that flag missing entries, out-of-range parameters and open deviations in executed batch records, so reviewers start with the exceptions. The quality unit still approves or rejects the batch.

  5. 05

    Quality document drafting with a human author

    Drafts of investigation reports, product quality reviews and procedure revisions built from your approved records, showing the source passage for each statement, with a named author who reviews and signs.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • The wrong kind of model in a critical step

    The EU's draft Annex 22 says models that keep learning during use, generative AI and large language models should not be used in critical GMP applications.1 Control: classify every use as critical or non-critical before design, use frozen deterministic models for critical steps, and keep language models in non-critical tasks where a qualified person is responsible for each output.

  • No baseline for the process being replaced

    The draft requires acceptance criteria at least as high as the performance of the process the model replaces, which means that performance has to be known.1 Control: measure the manual process first, for example inspector detection rates on a known defect set, then have the process expert approve metrics and acceptance criteria by subgroup before testing starts.

  • Test data that leaked into training

    A model tested on data its developers have seen will look better than it is. The draft calls for test data that was never used in development, protected by access control and an audit trail, with a record of which data was used for testing, when and how many times.1 Control: lock the test set in a separate repository before training starts and keep developers out of it.

  • Silent change after go-live

    FDA's draft guidance warns that model performance can change when inputs differ from the training data, and says changes to the model, or to manufacturing that may affect it, should be evaluated by the manufacturer's change management system.6 FDA has also noted that quality agreements with third parties, for example for cloud data management, may have gaps in managing AI risks.7 Control: put the model under change and configuration control, monitor performance and input drift against defined retest triggers, and write model updates into supplier agreements.

Rules that apply

  • EU GMP Annex 22: Artificial Intelligence (consultation draft, July 2025; not adopted as of November 2026)

    Primary text on health.ec.europa.eu(opens in a new tab)

    Covers static, deterministic models in critical GMP applications. Requires a documented intended use, acceptance criteria approved before testing, independent test data, a review of the features behind decisions, and change control and performance monitoring in operation.

  • 21 CFR 211.68, Automatic, mechanical, and electronic equipment

    Primary text on ecfr.gov(opens in a new tab)

    Computers and related systems used to manufacture drug products must be routinely calibrated, inspected or checked under a written program. Input and output must be checked for accuracy, with the degree and frequency based on the system's complexity and reliability.

  • FDA Computer Software Assurance guidance for production and quality management system software (February 2026)

    Primary text on fda.gov(opens in a new tab)

    Device makers should scale assurance of production and quality system software to risk. Software that determines product acceptability or adjusts process parameters with limited or no human review is generally high process risk.

  • FDA draft guidance on AI to support regulatory decision-making for drugs (January 2025)

    Primary text on fda.gov(opens in a new tab)

    Recommends a seven-step, risk-based credibility assessment for AI models that produce information supporting decisions on drug safety, effectiveness or quality, with life cycle maintenance plans available for review in the site's pharmaceutical quality system.

How to measure it

  • Model performance against pre-approved acceptance criteria, overall and by subgroup
  • Performance of the AI-assisted step against the measured manual baseline, such as false rejects and missed defects
  • Deviation investigation cycle time and the rate of repeat deviations
  • Batch record review time, and how often reviewers overrule the model's flags
  • Input drift alerts, retests triggered and unauthorized model changes detected

Questions buyers ask

Can generative AI or large language models be used in GMP manufacturing?

Under the EU's draft Annex 22, not in critical applications, meaning those with direct impact on patient safety, product quality or data integrity. The draft limits critical uses to static models with deterministic output. Language models may be used in non-critical GMP tasks if qualified, trained personnel are responsible for checking that each output is suitable. The draft was not yet adopted when we last checked, so confirm its status before relying on it.

How do you validate an AI model as a GMP computerized system?

Start by describing the intended use and the input data in detail. Set test metrics and acceptance criteria before testing, at least as high as the process being replaced. Test on independent, representative data that developers never had access to, and review the features behind the model's decisions. Then place the model under change and configuration control, and monitor its performance and input drift in routine use.

Is EU GMP Annex 22 on artificial intelligence in force?

Not as of November 18, 2026. The European Commission consulted on the draft, together with a revised Annex 11 and Chapter 4, from July 7 to October 7, 2025. The Commission's EudraLex Volume 4 page still lists the January 2011 versions of Annex 11 and Chapter 4 and shows no Annex 22. Plan against the draft, and check the Commission's page for the final text and its application date.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.Annex 22: Artificial Intelligence (draft guideline for stakeholders' consultation, EudraLex Volume 4)(opens in a new tab)European Commission, Directorate-General for Health and Food Safety, July 7, 2025 · health.ec.europa.eu
  2. 2.Stakeholders' Consultation on EudraLex Volume 4 - Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22(opens in a new tab)European Commission, Directorate-General for Health and Food Safety, July 7, 2025 · health.ec.europa.eu
  3. 3.EudraLex - Volume 4 - Good Manufacturing Practice (GMP) guidelines(opens in a new tab)European Commission, Directorate-General for Health and Food Safety, November 2026 · health.ec.europa.eu
  4. 4.21 CFR 211.68: Automatic, mechanical, and electronic equipment(opens in a new tab)Electronic Code of Federal Regulations (FDA), September 8, 2008 · ecfr.gov
  5. 5.Computer Software Assurance for Production and Quality Management System Software: Guidance for Industry and Food and Drug Administration Staff(opens in a new tab)U.S. Food and Drug Administration, February 3, 2026 · fda.gov
  6. 6.Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products: Draft Guidance for Industry and Other Interested Parties(opens in a new tab)U.S. Food and Drug Administration, January 2025 · fda.gov
  7. 7.Artificial Intelligence in Drug Manufacturing (Discussion Paper)(opens in a new tab)U.S. Food and Drug Administration, Center for Drug Evaluation and Research, 2023 · fda.gov
  8. 8.FY2025 Report on the State of Pharmaceutical Quality(opens in a new tab)U.S. Food and Drug Administration, Center for Drug Evaluation and Research, Office of Pharmaceutical Quality, 2026 · fda.gov