IT systems · Practical guide

Assessing Microsoft 365 Copilot for a GxP workflow

Copilot can find information and draft a convincing answer in seconds. Before using that answer for a quality task, check whether it draws on the correct effective document and whether the task permits this use. Start the assessment with the intended use.

Table of contents

At a glance

Can Microsoft 365 Copilot support a GxP activity?

Assess the specific task, authorised sources, access permissions and how a qualified person will check the output. Approval of a bounded use does not approve the product for every GMP activity. The 2025 Annex 22 consultation draft excludes generative AI and large language models from its proposed scope for critical applications. Evaluate the actual use under the applicable requirements, including the consequence of a reviewer accepting an incorrect answer.

Identify the product and the specific task

Microsoft documentation now also uses the name Microsoft Copilot for Microsoft 365 Copilot. Earlier names remain in use. Identify the enterprise service, licence, enabled features and data access that your assessment covers. Assess consumer products, extensions and custom agents separately where their scope or data flows differ.

Describe the task more precisely than “assistance”. Structuring meeting notes, assessing a deviation and preparing a batch decision have different implications. Define the inputs, output, users and permitted follow-up action. Effects on quality and data integrity may be indirect. If staff read only a generated summary, the assistant may become the effective basis for a decision.

Sources: [1]

Further reading:

Assess the regulatory position for the actual use

As of 14 October 2026, Annex 22 was a 2025 consultation draft. It excluded generative AI and LLMs from its proposed model scope and advised against their use in critical GMP applications. The draft conferred no general authorisation and did not itself create a binding prohibition. Assess the actual process against the applicable regulatory framework.

A statement that a human reviews the output does not amount to approval for general use. Actual impact determines criticality. Describe which content the model can influence and how errors would be detected. The responsible process owners decide the allowed use and its limits. Resolve open questions before relying on an output. This guide does not recommend assigning autonomous quality decisions to a general-purpose language assistant.

Sources: [2] [3] [4]

Review access permissions before introducing Copilot

Microsoft describes an enterprise service that makes organisational content available according to existing user permissions. If those permissions are already too broad, users may discover information they should not have. Review SharePoint libraries, groups, guests and other connected sources. A sharing link that drew little attention for years may become consequential once the content is searchable in a new way.

Use representative roles. A person authorised for site A might have no right to see confidential matters at site B. Test direct questions and summaries. Seeing a citation does not establish that permissions were enforced. Assess extra data paths created by extensions and agents separately. Check that the actual enterprise configuration supports the vendor's description of access controls.

Sources: [1] [2]

Check the source, effective revision and answer

An answer can accurately summarise the wrong source. Decide which documents govern the task. Keep drafts, historical revisions and current approved instructions distinguishable. Check whether users can tell which source the answer relies on and where it applies. Convincing wording alone does not establish that the right document was used.

Compare answers with independently prepared reference information. Include omissions, exceptions, tables and conflicting sources. Open the cited passage and check whether it supports the claim. Define when the assistant must state that it cannot answer reliably and direct the user to the original document or responsible contact. Unsupported additions drawn from general model knowledge can mislead users in a regulated process.

Sources: [2] [5]

Worked example of a question about an internal SOP

Suppose an employee asks how to handle an illegible label. The collection contains an effective SOP, an earlier revision and a draft for Site B. The permitted use is to find the correct source for Site A and direct the employee to the relevant passage for review. The tool must not authorise the process or generate an alternative action. Assess this use case with the responsible process experts before the pilot.

Include an unambiguous question, a question that omits the site, conflicting source documents and a user without access. Expect an appropriate answer or a clear statement of the limitation. Evaluate it against prepared sources and criteria. The matrix is our test aid. It gives no assurance about universal Copilot functionality or regulatory approval of this use case.

Test cases for the SOP question
Test caseExpected controlObservable result
Correct SOP Applicable source and revision Relevant passage in the original
Site unspecified Establish the missing context No unsupported assignment to a site
Earlier revision Distinguish which revision is valid The historical version is not presented as current
Conflicting information Make the uncertainty clear No invented resolution
Missing permission Restrict source access No confidential content is disclosed
Cannot be answered Acknowledge the limitation Identify who can resolve the question

Sources: [2]

Make human review workable

Define what the reviewer checks, including the source, completeness, figures, exceptions, scope and permitted use. A general instruction to check every AI output is inadequate if the reviewer lacks time or suitable references. Reviewers need process knowledge and a clear route for handling deviations. Match the review to the consequences of using the answer.

Test the control with deliberately incorrect outputs. Does the reviewer detect an omitted exception or a wrong number? Do they know when the answer must not be used? A signature without effective review does not adequately control the risk. Record permitted tasks and any prohibited or unapproved uses in terms staff can follow. Informal demonstrations do not establish that routine use is safe.

Sources: [2] [5]

Assess data processing and changes to the service

For the specified enterprise service, Microsoft says it excludes prompts, responses and data accessed through Microsoft Graph from foundation model training. Check the product, contractual terms and features in use. Also assess storage, retention, web search, extensions and actual data flows. This enterprise commitment does not extend automatically to consumer products.

A cloud assistant can change when its service, model, features or source connections change. Identify which changes matter and how you will learn about them. Rerun representative questions after suitable triggers. A test report covers the assessed scope and date. Assign responsibility for monitoring the service and investigating new failure patterns.

Sources: [1] [2]

Document the pilot and the decision on use

Record the task, criticality, inputs, sources, roles, output, review and permitted follow-up. Identify open points and actions. Define participants, data and boundaries for the pilot. Evaluate results against criteria set beforehand. Time saved can measure benefit, alongside measures of substantive quality.

For routine use, assign ownership, training, error handling and change assessment. Fit Copilot into existing operational processes so staff can act on the controls. The Annex 22 article explains the draft status in detail. Here the decision concerns the actual Microsoft service, the uses your controls can support and the limits that remain.

Sources: [2]

Project experience and collaboration

Assess the limits of Copilot through one defined task

Start with a bounded task, such as finding an internal SOP, and consider the consequences of a wrong answer. With the process owner and IT, we check source versions, actual access permissions and how a knowledgeable person can verify the result. A suitable, limited use may be tested in a pilot. If critical consequences or weak source controls cannot be managed adequately, the task needs a different solution. We record the basis for the decision and the limits of use. Human confirmation alone does not establish suitability.

FAQ on Copilot in GxP processes

Is Microsoft 365 Copilot already validated for GxP use?

Validation evidence relates to a specific task, its data, sources, users and controls. A general product statement cannot establish suitability for your process. Assess relevant vendor evidence and document what your own assessment covers, including its limits.

Does human review make every Copilot use non-critical?

No. The actual effect on decisions and records remains decisive. The reviewer must be able to detect errors and perform the review in practice. A confirmation without suitable references, competence and criteria cannot establish control. Assess critical uses separately.

Can Copilot surface information that was shared too broadly?

The enterprise service described here relies on existing permissions. If they are too broad, Copilot can make the affected information easier to find. Review groups, libraries and guest access before use. Your organisation remains responsible for its access model.

Sources: [1]

Are enterprise data excluded from model training?

Microsoft makes a specific commitment for the enterprise service described here. Verify the product, contract and enabled features. The same commitment does not automatically apply to consumer products, third-party agents or every extension. Also assess storage and other processing.

Sources: [1]

How does draft Annex 22 relate to Copilot?

The 2025 consultation draft is not in force. It excludes generative AI and LLMs from its proposed model scope and does not envisage their use in critical GMP applications. For non-critical use, it proposes that suitably qualified and trained people ensure outputs are suitable for their intended purpose. This grants no approval for Copilot. Assess applicable requirements against the actual process independently of the draft.

Sources: [3]

How do we evaluate a Copilot pilot?

Set criteria for source accuracy, completeness, error detection, access permissions and permitted use. Measure the actual effort, including review. Fast responses and high satisfaction alone cannot establish suitability. Use the findings and observed limits to decide whether to proceed.

Sources and review date

Sources were checked on 14 October 2026. The regulatory interpretation and product capabilities are supported separately. Adapt the examples, matrices and templates to the intended use. On 18 October 2026, the scope of the 2025 Annex 22 consultation draft concerning generative AI and LLMs was checked against its EudraLex status. The draft was not in force.

  1. Microsoft, Data, Privacy, and Security for Microsoft CopilotMicrosoft documentation also uses the name Microsoft Copilot for Microsoft 365 Copilot. Distinguish the enterprise product from consumer products and extensions. The access model and vendor privacy commitments were checked on 14 October 2026.
  2. European Commission, EU GMP Annex 11, Computerised Systems (2011)EudraLex listed this version as current on 14 October 2026. It covers application, risk, operation and the data lifecycle. The proposed tests are our own.
  3. European Commission, draft Annex 22 consultation text (2025), section 1, ScopeDraft checked on 14 October 2026. Generative AI and LLMs fall outside its proposed model scope. Its position on critical GMP use does not establish a newly binding prohibition.
  4. European Commission, EudraLex Volume 4 official document indexOn 14 October 2026, EudraLex listed the January 2011 Annex 11 and did not list Annex 22 as an annex in force. Separate EU implementing regulations for veterinary medicinal products have applied since 16 July 2026.
  5. PIC/S, PI 041-1, Good Practices for Data Management and Integrity (2021)Inspectorate guidance on data integrity, particularly electronic data, transfer, access and review.

Sources checked: .

Daniel Herrmann, Daniel Herrmann Consulting

Daniel Herrmann Consulting

How we can help

Daniel Herrmann Consulting combines assessment of the process with work on the actual system. Tell us the intended use, the evidence you have and the decision you need to make.

  • AI system validation →

    Bring your AI system into GxP use with traceable evidence and keep it under control in operation. Once the system and intended GxP use are defined, we validate according to risk, addressing data, supplier, tests, release and change control.

Send a project enquiry Customer testimonials →