<img height="1" width="1" style="display:none" src="https://www.facebook.com/tr?id=323483448267600&amp;ev=PageView&amp;noscript=1">

Solutions by industry / Chemicals & pharma

AI in GxP production, released like any validated change

Models are built in the cloud, approved under your change control and verified where they run. Every decision traceable, every record retrievable, for as long as you must keep it.

21 CFR Part 11and EU GMP Annex 11: the controls our GxP deployments are designed around
One signed releasefrom the model to the unit that runs it
Your systemsidentity, e-signature and QMS, where you already have them
Industry
PharmaceuticalsBiotechFine chemicals
Focus
Data integrityChange controlAudit trailElectronic signatures
With Wizata
Edge deploymentMLOpsVersioned pipelinesClosed loop

The challenge

Why AI stalls at the validation gate

In GxP production, the hard part of AI is rarely the model. It is what quality must be able to answer, sometimes years later: which model ran, on which data, approved by whom, and why the process did what it did.

Models are retrained, cloud platforms change every week and data comes from many systems. Treated as one large system to validate, AI becomes a project nobody wants to own.

Models that keep changingCloud platforms that keep changingRecords kept for decadesWho approved what, and why

What your QA needs to see

  1. A clear boundary between what is validated and what is not.
  2. One release that states exactly what will run, approved under change control.
  3. Proof, at the point of use, that what runs is what was approved.
  4. A complete audit trail, reviewable without the vendor in the room.

The principle

Models are made in the cloud. They are verified at the plant.

Well-trodden GMP ground, applied to AI: a tool does not carry the qualification burden of the system that uses its output, as long as that output is verified at the point of use. The boundary is agreed with your quality function and written into the validation plan.

AuthoringWizata cloud

Data science, training, pipelines and their full version history. Where models are built and tested.

Tool, outside GxP scope
The controlled artifactRelease package

A manifest that states exactly what is inside (models, configuration, source version) and a digest that proves it. Reviewed and signed under your change control.

Approval gate
Point of useWizata edge, on site

Installed unit by unit. Checks the digest before anything runs, and records every recommendation and setpoint in a tamper-evident audit trail.

GxP, qualified on site
Control layerYour validated DCS

Keeps its interlocks and limits. AI setpoints arrive as inputs inside them, never around them.

Unchanged

The release flow

From a trained model to a running unit, one signed release

No model reaches production any other way. A retrained model is a new version, in a new release, approved like any other change.

1Author

Versioned

Pipelines and models versioned in Git, in your own repository if you prefer.

2Package

Described

A manifest lists models, configuration and source version, sealed by a digest.

3Review

Readable

QA and process owners review the manifest in plain language before anything moves.

4Sign

Approved

Electronic signature with re-authentication and its meaning, in your system or ours.

5Install

Verified

Per unit, DEV then QUAL then PROD. The edge refuses anything that does not match.

6Run

Recorded

Each decision logged with its data, model version, approver and time.

21 CFR Part 11 and EU GMP Annex 11

The controls, mapped to the design

How our GxP deployments are designed to meet each control. Qualification itself is done on your site, with your QA and CSV team, against your own test plan.

ControlDesign response
Audit trail§11.10(e) · Annex 11 §9Secure, computer-generated, time-stamped record of who did what and when, in a tamper-evident store. Time taken from your approved time source.
Electronic signatures§11.50 · §11.70 · §11.200The signature is bound to the release record: name, date and meaning, with re-authentication at signing. Where you run a validated e-signature system, the approval happens there.
Access control§11.10(d), (g) · Annex 11 §12Named users through your identity provider, with single sign-on and multi-factor authentication. Roles separate who builds, who reviews and who releases.
Accurate and complete copies§11.10(b)Audit trail and records exported in human-readable form, so QA can run its periodic review independently of Wizata.
Data integrityALCOA+Source timestamps and quality codes kept with every value from the control system; batch or lot identifiers carried through where the plant provides them.
RetentionAnnex 11 §17Records archived into your systems for the retention period you define, not held only on an edge device.
Change control and environmentsAnnex 11 §10 · GAMP 5Separate DEV, QUAL and PROD. Releases scoped per unit, so one unit is qualified and proven before the next.

Questions your QA will ask

Straight answers

Is Wizata 21 CFR Part 11 compliant?

What Wizata brings is a deployment designed around the 21 CFR Part 11 and Annex 11 controls, supplier documentation for your qualification, and engineers who work through it with your QA and CSV team.

Does the cloud platform have to be validated?

With a risk-based approach (GAMP 5), usually not: the cloud is the authoring tool and its output is verified at the point of use. Where the boundary sits is agreed with your quality function and written down before the project starts.

Can we use our own e-signature system and QMS?

Yes. Where you already have a validated system, it is often the best place for the approval: AI releases are then approved exactly where your other changes are.

What happens when a model is retrained?

It becomes a new version in a new release, reviewed and signed like any other change. Nothing updates silently in production.

Does the edge need a connection to the cloud?

Not necessarily. Releases can be installed under your own change-control procedure, and the audit trail exported to your archive. The network design is agreed with your IT and OT teams.

Where should we start?

Where the GMP impact is indirect and the data already exists: clean utilities, energy, cooling. Then a first process unit, qualified on its own, before rolling out to the next.

Next step

Bring AI into GxP production, one validated step at a time

Talk to one of our engineers about your site, your quality system and where AI would pay back first.

Your contact Jérémy Lambert Pharma lead, Wizata jla@wizata.com