Practical guides · Practical guide

How to establish AI governance in a pharmaceutical company

Who may try which AI tool, and with which data? For a first pilot, appoint an owner, define permitted tasks and give staff a contact for questions. When AI produces quality-relevant output in a specific system, assess that system’s intended use and what evidence it needs.

Table of contents

At a glance

Where should a pharmaceutical company start with AI governance?

Create an inventory of uses. For each one, record its purpose, data, technology, effect on the process and accountable owner. Decide whether to restrict the use, investigate it further or approve it within defined limits. Connect that decision to supplier assessment, training, change control and incident handling. Document the GxP rationale alongside other applicable assessments.

  • Record the activity and its owner for each use case.
  • GMP process criticality and categories under the AI Act require separate assessments.
  • Define a pilot's boundaries, data rules and decision owner.
  • After the first decision, monitor changes and actual use.

Record what people do with each AI product

For each application, identify its purpose, users, input data and how people use the output. One product can support several tasks. An inventory entry called “assistant” fails to distinguish drafting, search and decision support.

Ask which tools people already use. An accurate inventory allows the company to make decisions. An empty register is of little use if staff are already using AI. For an initial assessment, describe the data types and their use at an appropriate level. There is no need to transfer confidential information into new tools simply to classify the application.

Sources: [1] [2]

Assess each regulatory and operational question on its own terms

A GxP assessment examines the effect on relevant processes, records and decisions. Classification under the EU AI Act follows its own definitions, roles and scope. Data protection, information security and contractual issues may also apply. A single internal traffic-light rating cannot replace these assessments.

Involve the relevant specialists and record their conclusions separately. Using an application in a pharmaceutical company does not by itself make it high-risk under the AI Act. A lower legal risk category does not by itself establish suitability for a particular GMP process either. Where a product or safety component is involved, assess Article 6(1), the legislation listed in Annex I and the need for third-party conformity assessment. The applicable text sets out further qualifications. An industry classification cannot replace this assessment.

Sources: [3] [1] [2]

Who decides whether a new AI use may proceed?

Name an owner who can explain the purpose and limits of the proposed use. Agree which specialist assessments are needed. The matrix is our suggested organisational model. Adapt the roles to the company.

Define who resolves disagreements. Passing every question between committees makes the procedure hard to use. Preserve the independence and responsibilities required for the assessment.

Responsibilities for a new AI use
DecisionInput needed fromOutcome
Describe the use Process team and application owner Defined purpose and excluded tasks
Assess the process impact Relevant specialists and QA where GxP is involved Rationale for the assessment or testing required
Assess the data and the service IT, information security and relevant specialists Suitable use of data and supplier services
Determine the legal position Legal and data protection specialists A separate rationale for applicability and obligations
Set pilot boundaries Process team, IT and the people assessing the results Scope, data, criteria and the decision at the end
Assign operational responsibility Named operational owner Routes for handling changes, incidents and reviews

Sources: [4] [1]

Give a pilot clear limits and a decision at the end

A pilot should resolve a specific uncertainty. Name the permitted data, the users involved and what may happen with the outputs. Prevent pilot results from entering a critical workflow unintentionally. Identify the evidence needed to take the next decision.

A successful pilot establishes only what was examined under its stated conditions. Moving into operation can require additional work. Record the absence of reference data or assessable acceptance criteria as a finding. The pilot may lead to a narrower task or a decision to stop.

Sources: [2] [1]

Further reading:

Resolve the boundaries of use before approving a project

The 2025 draft Annex 22 consultation text grants no general permission to use generative AI in critical GMP applications. Before investing in detailed testing, assess whether the purpose, model type and controls support the proposed use. A human confirmation click cannot establish that suitability.

Show how an application idea becomes a brief for expert review or testing. Our AI validation guide and Annex 22 articles discuss technology limits and specific checks. At governance level, identify who initiates the assessment, which decision is needed and how an unsuitable use will be limited.

Sources: [5] [6] [7]

Further reading:

The same assistant used for two different tasks

Suppose a company is assessing a tool that revises the wording of general internal texts. Another team wants it to generate content that directly informs quality decisions. The product name and licence are the same. The process impact and the evidence needed can differ substantially.

Record the two uses separately. Approval of the first task must not silently authorise the second. Assess and document whether a narrower purpose, additional controls or another solution would be suitable in each case. This example does not establish that either use is permitted.

Sources: [1] [2]

Track changes, incidents and actual use

A change to the model version, prompts, knowledge sources, roles or interfaces can affect outputs. Agree who will detect and assess such changes. Supplier information is one input. Assess the effect on how your company uses the application.

Define a route for unsuitable or unexpected outputs. Who can limit use, who investigates affected transactions and who decides whether use can resume? Base the response on impact and risk. Those responsibilities continue after the initial inventory has been signed.

Sources: [1] [8]

Give people rules they can use in their daily work

Article 4 of the AI Act, as consolidated on 27 July 2026, requires providers and deployers to take measures supporting the development of AI literacy among staff and others acting on their behalf. Consider their technical knowledge, experience, education and training and the context of use. This version does not require a guaranteed level of individual competence.

Staff should understand the limits relevant to their own tasks. A general webinar may provide a foundation. People in specific roles often need further instruction. Practise when to check an output against the original source and when an application falls outside the task proposed.

Start with a manageable procedure and try it on different applications. Check that decisions are understandable, documented and retrievable. The number of entries in the inventory says little about effectiveness. Check whether each critical open question reaches an owner and receives a traceable assessment.

  • Record each application, its purpose, data and process impact.
  • Document the reasons for the GxP assessment and other legal classifications separately.
  • Define the conditions for the pilot and the decisions to be taken afterwards.
  • Report changes and unsuitable outputs to the people responsible.
  • Prepare employees for their actual tasks and limits of use.

Sources: [4] [1] [3]

Project experience and collaboration

Test your AI governance process on actual uses

With the departments involved, we identify where AI is already being used and where its use is planned. For selected applications, we clarify the purpose, the data flows and who can make each decision. That gives staff a procedure they can use in everyday work. A small number of tightly defined applications may need a simple starting point. Multiple suppliers and different uses require a more detailed inventory, clearer ownership and agreed change procedures.

FAQ on AI governance in pharmaceutical companies

Does every AI tool need the same approval procedure?

Match the assessment to intended use, risks and applicable requirements. Not every application needs the same depth of review. A common set of initial questions and clear ownership make decisions easier to explain and compare.

Sources: [2]

Is every pharmaceutical AI application automatically high-risk under the AI Act?

No. The legal classification depends on the actual scope and on the roles defined in the regulation. The industry alone does not determine it. Record this separately from the GxP risk assessment.

Sources: [3]

Does the AI Act replace a GMP assessment?

No. Both frameworks may be relevant. A legal classification does not establish the suitability of a particular GMP configuration and use. Record the basis for each assessment and who is responsible.

Sources: [3] [1]

Can an employee simply approve an output from generative AI?

A human confirmation step cannot make an unsuitable use acceptable. Assess the purpose, technology, consequences of error and actual effectiveness of the control. Draft Annex 22 is not a general authorisation for critical GMP use of generative AI.

Sources: [5] [2]

Which documents should a small organisation start with?

Start with an understandable use-case inventory, decision rules and records of the assessments actually performed. Intended use and applicable requirements determine the appropriate detail. Assign ownership so people can act on the records.

How can we tell whether AI governance works?

Check whether people identify new uses in time, assess them and operate them within defined limits or restrict them where needed. Changes and incidents should reach the responsible owner. Test this with actual applications. Counting policies cannot establish effectiveness.

Sources and review date

Primary sources were checked on 17 October 2026. The AI Act source is the EUR-Lex consolidated version of 27 July 2026, incorporating Regulation (EU) 2026/1744. Determine obligations, roles and dates of applicability for the specific case. The Annex 22 text is a consultation draft. We developed the decision process, matrix and example as organisational aids. The EU GMP discussion concerns human medicinal products under EudraLex Volume 4.

  1. European Commission, EU GMP Annex 11, Computerised Systems, January 2011 revisionGMP for human medicinal products. Relevant provisions include sections 1–4, 7, 10–13 and 16–17, as identified in the article.
  2. ICH Q9(R1), Quality Risk Management, official text published by the European CommissionRisk-based decisions, appropriate formality and management of subjectivity. The matrices in this article are implementation suggestions.
  3. Regulation (EU) 2024/1689, AI Act, consolidated text of 27 July 2026This framework classifies uses and roles on its own terms. GxP criticality is a separate matter. Assess time-specific obligations for the particular case.
  4. European Commission, EU GMP Chapter 2, Personnel, effective 16 February 2014Personnel, role-specific training and evaluation of its practical effectiveness. The official EudraLex entry links to the original chapter.
  5. European Commission, Annex 22 Artificial Intelligence, 2025 consultation draftConsultation draft. Its proposals are not in force. Do not treat generative AI as an approved central use case.
  6. European Commission, 2025 consultation on Chapter 4, Annex 11 and the proposed Annex 22Identifies the consultation and the associated drafts. Publication for consultation does not establish entry into force.
  7. European Commission, EudraLex Volume 4 official document indexChecked on 17 October 2026. Human medicines and the veterinary medicines rules changed from 16 July 2026 are distinct.
  8. European Commission, EU GMP Chapter 7, Outsourced ActivitiesOutsourced GMP activities, suitability, responsibilities, agreements and oversight.

Sources checked: .

Daniel Herrmann, Daniel Herrmann Consulting

Daniel Herrmann Consulting

Compare three different AI applications

For each application, bring its purpose, data types, users and intended action after an output. Compare the three to see which decisions can be governed together and which require separate expert assessment.

  • AI consulting →

    Use AI for defined purposes with rules and solutions that fit your processes. We prioritise use cases, develop governance and policies, and assess providers and technical options.

Send a project enquiry Customer testimonials →