Is AI allowed in GMP?

Yes, within limits. Draft EU GMP Annex 22 allows only static, deterministic models to support critical decisions, with stated confidence, and keeps generative AI to non-critical uses with a human in the loop. We design every assist to that line.

Does it compete with SAP Joule?

No. Joule assists inside SAP. These assists work on data held outside SAP and on the sponsor view; both can run side by side.

Where does our data go?

It stays in your cloud tenancy. Models run in your environment or through your approved provider, and no sponsor data is used to train shared models.

Are you an SAP partner?

No. We are an independent studio. We build beside SAP using its released APIs and work with your SAP partner, who stays responsible for SAP itself. If you need the integration certified by SAP, we scope that into the project.

A studio service

AI assist layer for SAP supply chains in GMP plants useful AI that stays inside Annex 22.

AI belongs in a GMP supply chain where it saves people time and never where it takes a decision away from them. We build assists that read documents, predict shortages and excursions, and answer questions in plain language, designed to the draft EU GMP Annex 22: predictive models are static and versioned and show their confidence, generative tools only draft for a person to confirm, and nothing releases, rejects or ships a lot.

See the packages →All supply chain modules

AI assist layer for SAP supply chains in GMP plants

The facts behind this module

Draft EU GMP Annex 22

The draft limits AI in critical GMP applications to static models that “do not continue to learn or adapt after deployment” and give “deterministic outputs.”

Source: European Commission, draft Annex 22 to EU GMP Volume 4, as reported by GMP Insiders, August 2025.

Draft EU GMP Annex 22

“Generative AI and Large Language Models (LLMs) are not acceptable in critical applications.”

Source: European Commission, draft Annex 22 to EU GMP Volume 4, as reported by GMP Insiders, August 2025.

Draft EU GMP Annex 22

For non-critical applications, “there must be a qualified human responsible for evaluating the output, following the ‘human-in-the-loop’ principle.”

Source: European Commission, draft Annex 22 to EU GMP Volume 4, as reported by GMP Insiders, August 2025.

Quoted from the sources linked. We add no estimates of our own.

What it looks like

AI assists — model register

Assists live

Static models

Generative, human-confirmed

Reviews due

Assist	Type	Version	Use	Decision made by	Status
Shortage prediction	Static model	v1.3	Critical support	Planner	Approved
Lane risk score	Static model	v2.0	Critical support	QA	Approved
Certificate reading	Generative	v0.9	Non-critical	QC reviewer	Review

Interface concept with invented sample data, shown to illustrate the design. It is not a screenshot of a live system.

How it connects

SAP S/4HANA

Batches, QM, EWM, planning
Stays the system of record
Core left standard, nothing changed inside

⇄released APIs only

The app, in your tenancy

SAP BTP, Azure or AWS, your region
Audit trail, roles, single sign-on
Read-only in a Pilot; write-back by agreement

⇄scoped access

People and plant systems

Your teams, by role
Sponsors, each to their own data
MES, LIMS, QMS and logger platforms

SAP's clean-core rule, quoted: “Extensions must be developed using only released local or remote public SAP APIs, BAdIs, or ABAP RESTful application programming model business object (BO) extension points.” SAP PRESS, SAP S/4HANA Clean Core: Principles, Benefits, and Best Practices.

N° 01What SAP already covers

We start from SAP, not against it.

SAP Business AI and Joule add assistance and agents inside SAP workflows, with more arriving through 2026. These assists work on data outside SAP: supplier PDFs, MES, LIMS and logger data, and the sponsor and campaign view. This module is a side-by-side extension: it reads and writes SAP only through released APIs and leaves the core standard.

N° 02Where it helps

Four situations this module is built for.

AI pilots that can't go live

An AI assist cannot be used in GMP work until QA can show how the model is controlled. We design to the draft Annex 22 from the start.

Generative AI near GMP decisions

We keep generative tools in non-critical, human-confirmed work and say so in writing.

Models that change silently

Predictive models are versioned and frozen in use; retraining is a change-controlled release.

Answers without sources

Every AI output links to the records behind it.

N° 03What the module covers

Scope, in plain terms.

Document reading

CoAs, delivery notes and labels read and pre-filled for confirmation.

Prediction

Shortage, lead-time and excursion-risk models with confidence and drivers.

Plain-language Q&A

Questions answered from your SAP and plant data, scoped by role and sponsor, every answer sourced.

Model governance

Intended-use statements, test data, version control, confidence thresholds and performance monitoring.

N° 04For each person in the decision

One page, six readers.

Supply chain lead

The scope above, three fixed-scope packages with prices, and a read-only Pilot you can run on one site before committing.

QA and validation

GAMP 5 Category 5 development documents, executed developer IQ/OQ evidence, Annex 11 and Part 11 controls in the GxP-ready package. You run and sign validation.

IT and your SAP team

Released SAP APIs only, no changes inside SAP, one least-privilege technical user, deployed in your own tenancy and region.

Sponsor-facing teams

Each sponsor sees only their own data, enforced in the data layer, with every access logged.

Procurement

Fixed scope and price per package, a written not-included list, and our vendor questionnaire answered in full on request.

Finance

The price shown is ours. SAP licences, including any SAP Digital Access for documents created in SAP, hosting and validation effort are yours and are listed as not included.

N° 05The Pilot, week by week

Eight to ten weeks, read-only from SAP.

Weeks 1–2: scope and access

Requirements workshop, the list of SAP APIs to be read, a technical user from your SAP team, and sample data agreed.

Weeks 3–5: build

The module built in your tenancy against your SAP test system, with a working demo every week.

Weeks 6–7: your data

Connected to the agreed SAP data, tested by your users, findings fixed.

Weeks 8–10: live on one site

Go-live for up to 25 users, a handover session, and a written decision paper for the Production package.

N° 06Your three options

Inside SAP, packaged, or beside SAP.

Customise inside SAP

Done by your SAP partner. SAP's clean-core rule limits extensions to released APIs and extension points, and each change follows your SAP change and validation process.

Buy a packaged product

A vendor's standard product and roadmap, with its own licence, validation package and integration to your SAP.

Build beside SAP (this page)

An app scoped to your process, in your tenancy, connected through released SAP APIs, with the source code and documentation handed to you.

N° 07Investment

Three ways to add AI.

One assist

From $24,000

One assist on one module, proven on your data.

One assist, e.g. certificate reading or shortage prediction
Built on your SAP and plant data, in your cloud tenancy
Accuracy measured on your historical records before go-live
Human confirmation built into the workflow
Not included: GMP-critical use, model validation documentation, SAP AI licences

Start the project

Recommended

Assist suite

From $58,000

Three assists across the modules you run.

Three assists across your modules
Shared retrieval over SAP, MES, LIMS and logger data, scoped by role and sponsor
Usage and accuracy dashboard
Quarterly model review for the first year
Not included: GMP-critical use, validation documentation, SAP AI licences

Start the project

GMP-scoped models

From $85,000

Predictive assists documented for use in critical GMP support.

Static, versioned models with intended-use statements
Independent test data sets and acceptance criteria
Confidence thresholds and explainability on every output
Performance monitoring and change-controlled retraining
Documentation aligned to draft EU GMP Annex 22 and Annex 11
Not included: validation execution and QA sign-off, generative AI in critical use (not permitted)

Start the project

Before you buy

What you get

An independent build that connects to SAP only through SAP's released APIs
A read-only Pilot that creates nothing in SAP, so it carries no SAP licence exposure
Delivery alongside your SAP partner and IT, who keep control of SAP itself

What we don’t do

SAP licences, or SAP Digital Access for documents a two-way build creates in SAP
Changes to your SAP configuration or custom code inside SAP
SAP certification of the integration, unless you ask us to scope it in

DesignKompanie is an independent studio and is not affiliated with, endorsed by or a partner of SAP SE. SAP and SAP S/4HANA are trademarks of SAP SE.

N° 08Questions

The answers we give most often.

Is AI allowed in GMP?: Yes, within limits. Draft EU GMP Annex 22 allows only static, deterministic models to support critical decisions, with stated confidence, and keeps generative AI to non-critical uses with a human in the loop. We design every assist to that line.
Does it compete with SAP Joule?: No. Joule assists inside SAP. These assists work on data held outside SAP and on the sponsor view; both can run side by side.
Where does our data go?: It stays in your cloud tenancy. Models run in your environment or through your approved provider, and no sponsor data is used to train shared models.
Are you an SAP partner?: No. We are an independent studio. We build beside SAP using its released APIs and work with your SAP partner, who stays responsible for SAP itself. If you need the integration certified by SAP, we scope that into the project.

N° 09Related

Start your book

Planning the AI assist layer build?

Tell us your sites, the systems you run and what has to change. You'll get a scoped plan and a fixed price within two business days.

Start the project See the process