Using AI as a team: which working agreements do you need?

Start with one task and an owner

Choose a task for which the team can explain the input and intended output. For example, draft answers to product questions, not a broad instruction to handle customer contact with AI. Define what remains outside the agreement, such as committing to prices or handling exceptions.

Assign someone responsibility for the working practice. They do not need to perform every task themselves, but should be able to determine who changes instructions and where questions go. Without an owner, a shared template can quickly become a collection of different versions.

Make seven practical agreements

Record the agreements briefly somewhere the team can find them. A useful agreement describes not only what is allowed, but also when a colleague should stop and ask for help. Avoid a long list of rules that nobody can apply during everyday work.

The seven areas below are a starting point. They are not a complete organisational policy or legal approval. Align them with existing rules for information, access and responsibility. Microsoft, for example, provides resources for preparation and employee adoption [M3]; these do not replace agreements about your own task.

• The permitted tool and account.
• The information that may be used for the task.
• The shared instruction or template.
• Checks a person must perform before using the output.
• Exceptions and reasons to stop.
• The owner, substitute and contact for questions.
• When to evaluate outcomes and update the agreement.

Fictional example: a service team

A service team uses an approved AI environment to prepare draft answers from public manuals. The task is limited to product questions with an available current source. The team does not submit customer files, and AI does not send anything to the customer automatically.

The colleague first records the product and version, selects the corresponding manual and uses a shared instruction. They then check whether the source supports the explanation, whether unauthorised commitments were added and whether the question was actually answered. Missing sources or conflicting versions are referred to the product owner.

The owner of the working agreement maintains the instruction and has changes checked against representative examples. This is a fictional setup, not a client case. The amount of work and checking required must be investigated in the actual situation.

On a narrow screen, scroll the table sideways. With a keyboard, focus the table and use the arrow keys.

Make human review practical

“A colleague checks it” is not specific enough. State what they check: the correct source, current product version, amounts or commitments and missing information. The reviewer needs access to the source and time to investigate discrepancies.

If nobody can assess the output, the working practice is not ready for that task. Persuasive wording is not evidence. Agree when a colleague completes the task manually or hands it over. Keep information needed to investigate errors within the existing agreements on data use.

Manage access and changes

Use accounts and access rules permitted within your organisation. Do not casually share a personal account to save a licence. Assess which information colleagues may access and whether connected sources unintentionally broaden that access. Microsoft’s documentation, for example, illustrates why existing permissions deserve attention with business Copilot [M2].

Also agree how to handle a new colleague, a departing colleague, a changed role and a changed tool. Who updates permissions or instructions, and who checks the consequences? Use existing management processes where they are sufficient; this article introduces no new management platform.

Assess outcomes, not just usage

Frequent AI use does not automatically mean a better working practice. Look at total handling time, corrections required and what happens with exceptions. Include reading source material and repairing incorrect drafts.

Discuss with colleagues whether the output is useful to the next person in the process. A faster draft may create extra checking elsewhere. Agree in advance which improvement you want to investigate and when to reassess the approach. Released capacity is not the same as financial profit.

Assess capacity and costs in a business case

When is an integration relevant?

An agreed working practice can be the lasting solution. Investigate an integration when recurring handovers between systems or copying information remain a substantial part of the work. An integration also requires attention to authority, errors, recovery and ownership.

An unclear instruction does not automatically improve when a system executes it. First consider whether the task and its exceptions are described well enough. The existing article on a standalone AI tool or workflow helps compare the approaches; it is not a mandatory maturity ladder.

Try the agreement on a limited task

Choose representative, fictional where necessary, examples before expanding use. Include an ordinary question, missing information, a contradiction and a question outside the task. Discuss results with the people performing and checking the work.

Record what needs changing, who will do it and when you will review it again. If further investigation, technical setup or guidance is needed, first agree scope, responsibilities and costs. Ongoing management or training is not automatically included in an exploratory conversation.

Frequently asked questions

Who keeps the instructions current?

Assign an owner and a substitute. Agree who checks changes before the team uses an updated instruction.

What if colleagues receive different answers?

Check both against the source and agreed task criteria. The most persuasive answer is not automatically the correct one.

Do we need to connect systems now?

No. An agreed practice with human review may remain sufficient. Investigate an integration only when the remaining work warrants it.

Sources and conditions

Official documentation checked on 12 October 2026. Capabilities and terms depend on the product, account, subscription and setup. No customer account or supplier agreement was tested.

• [M2] — Microsoft Copilot — Data, Privacy, and Security
• [M3] — Microsoft Copilot — adoption and onboarding resources