How to Build a Team AI Assistant in Copilot, ChatGPT, Claude or Gemini

In our programmes, every working group builds at least one AI assistant for its own work. Some of them are still in daily use months later. Many assistants we see elsewhere quietly die within weeks.

The ones that die share three habits: they try to do five jobs, nobody tested them, and nobody owns them. The ones that survive are boring in the right way. This article shows how to build the second kind, in whichever tool your company already licenses.

What is a team AI assistant?

A team AI assistant is a saved set of instructions, reference files and a few starter prompts that the whole team uses for one recurring task. Everyone gets the same behaviour without writing the same long prompt again.

Each vendor has its own name for it:

• Microsoft Copilot: an agent, built in Agent Builder.
• ChatGPT: a project, and for repeatable workflows in business plans, skills and plugins. Custom GPTs are on their way out (more below).
• Claude: a project.
• Gemini: a Gem.

The mechanics differ, the method does not.

Which tool should you build your assistant in?

Build where your licence and your data already are. A great assistant in a tool half the team cannot open is not a team assistant.

Tool	What you build	Building and sharing	Watch out for
Microsoft Copilot	An agent in Agent Builder	Copilot Chat users can create agents if your admin allows it. Without a Microsoft 365 Copilot licence they use web knowledge; grounding in SharePoint or OneDrive needs the paid licence or pay-as-you-go billing set up by your admin.	Agents can have several owners. Add a second owner so the agent does not depend on one person.
ChatGPT	A project with instructions and files; skills and plugins in business plans	Projects and project sharing exist on every plan; Business, Enterprise and Edu can also share with workspace groups.	Custom GPTs retire on 11.12.2026. Use a business workspace for company data.
Claude	A project with instructions and knowledge files	Sharing projects with colleagues needs a Team or Enterprise plan.	On personal plans each person builds their own copy, which drifts.
Gemini	A Gem with instructions and files	Gems are shared like Drive files in Google Workspace, within your admin's settings.	Shared files follow Drive permissions, so check who can see them.

If you are still choosing the tool itself, our comparison of ChatGPT, Copilot, Claude and Gemini covers the decision. In Microsoft 365 companies, the difference between Copilot Chat and the paid licence decides what your agents can see.

Custom GPT retirement: what happens on 11.12.2026?

OpenAI announced on 11.09.2026 that it is retiring custom GPTs. The standard retirement date is 11.12.2026; some Enterprise workspaces with an approved deferral have until 11.02.2027. After that date, GPTs and their pages are no longer accessible.

The replacement is plugins. A GPT can be migrated: its instructions become a skill inside a plugin. Several things do not carry over, according to OpenAI and early migration reports: the selected model, the sharing settings (the plugin starts private) and custom actions. Past conversations stay in your chat history but do not move into the plugin. Personal Free, Go, Plus and Pro accounts can already no longer create new GPTs.

If your team relies on GPTs, list the ones used in recurring work, migrate them, then run each one through the test kit below before you switch over. Do not assume a migrated assistant behaves the same.

When is a task worth an assistant?

Build one only when all four are true:

1. The task recurs at least weekly. Monthly tasks are better served by a saved prompt.
2. The input looks similar every time. An enquiry email, a supplier invoice, meeting notes.
3. You can describe a good result. If two colleagues disagree on what good looks like, settle that first.
4. Someone in the team cares. That person becomes the owner.

Typical first assistants we see built:

Function	Assistant	Input	Output
Sales	Offer drafter	A client enquiry	A draft offer with [CHECK PRICE] placeholders
Customer service	Complaint reply	A complaint thread, names removed	A reply in house tone for a person to check
Finance	Variance commentary	Budget versus actual table	Five bullets for the management report
HR	Job ad writer	A role brief	A job ad in house style, with questions for missing details
Operations	Supplier letter	A delivery deviation	A clear letter with the next steps
Public sector	Citizen reply	An enquiry, name and address removed	A plain-language reply with the legal reference to check

The 12 rules we use when building assistants

These come from the assistants groups build in our programmes, in all four tools.

1. One assistant, one job. An assistant that writes offers and answers complaints does both badly.
2. Weekly task or no assistant. Otherwise nobody remembers it exists.
3. Keep instructions tight. Copilot agents accept up to 8,000 characters; 3,000-4,000 are followed more precisely in our experience. The same logic holds in every tool.
4. Background material goes in files. Price lists, tone guides and templates are knowledge, not instructions.
5. Placeholders, not guesses. Tell it to write [CHECK PRICE] or [CHECK DATE] when it does not know. An invented figure that looks right is the most expensive failure.
6. Set length limits. "At most 150 words" beats "be concise".
7. Add three or four starter prompts. They show colleagues what the assistant is for.
8. New task, new chat. Long chats drift. A new chat starts again from the saved instructions.
9. Let the AI improve its own instructions. After a bad answer, ask it which instruction was unclear and to suggest a fix.
10. Build v2 next to v1. Never overwrite a working version until the new one passes the tests.
11. Test before release. A small test kit with pass criteria. One bad answer means one new rule in the instructions, then rerun the whole kit.
12. One owner edits the instructions. Everyone else reports what broke.

Copy-paste template: assistant instructions

We build instructions on the same six parts we teach for prompts (ROCSRS).

ROLE. You are [role] at [company], helping the [team] team with [task].

OBJECTIVE. Produce [output] that [reader] can [use for what] without rewriting it.

CONTEXT. Our clients are [who]. We write in [language] in a [tone] tone. Use the attached files: [price list], [tone guide], [three good examples]. [Any rule that is specific to this task.]

STRUCTURE. [Format, sections, order.] At most [number] words.

REASONING. Do not explain your reasoning unless asked. If the input is unclear or contradictory, first list what is missing, then stop.

STOP. Never invent prices, dates, quantities, names or legal references: write [CHECK PRICE], [CHECK DATE] and so on instead. Do not promise anything to the client that is not in the input. Do not answer questions outside [task]; say what this assistant is for.

How do you test an assistant before the team uses it?

Write five to eight test cases from real work, with fictional names and numbers. Each case gets a pass criterion that anyone can check.

#	Test case	Pass criterion
1	A typical, complete input	Usable with small edits only
2	Input with the price missing	Writes [CHECK PRICE], invents nothing
3	A contradictory input	Lists the contradiction first
4	A long, messy email thread	Picks out the actual request
5	A request outside the assistant's job	Declines and says what it is for
6	Input in the other working language	Answers in the language set in the instructions

Run the whole kit after every change to the instructions. It takes ten minutes and saves the week in which nobody trusts the assistant any more.

How do you keep it alive?

• The owner checks it monthly. Five minutes: is it used, what broke, what should change.
• Keep a short change log. Date, what changed, why.
• Review assistants quarterly. Retire what nobody uses. Fewer, better assistants beat a long list.
• Mind licence reviews. In Copilot, give every agent a second owner with edit rights. Before you remove a licence or someone leaves, check that someone else can still edit it.

An assistant that is documented, owned and in daily use is what we call a shipped AI solution. Anything less is a demo.

What to do this week

Next 15 minutes. Pick one task your team does every week. Collect five real examples of the input and one example of a good result.

Next 7 days. Fill in the instruction template, build v1 in the tool you license, write the test kit and run it. Ask two colleagues to try it on real work.

Next 30 days. Name the owner, add the starter prompts and release it to the team. If you depend on custom GPTs, finish the migration and rerun every test kit before 11.12.2026.

In our AI Onboarding Partner programme each group builds and tests at least one assistant like this with us. Start with a discovery call.

Sources

OpenAI Help Center: Custom GPT retirement and migration FAQ, Projects in ChatGPT, Skills in ChatGPT. Microsoft Learn: Agents in Microsoft 365 Copilot Chat, Share and manage agents built in Agent Builder. Anthropic: Creating and managing projects. Google Workspace Updates: Gem sharing. Product facts checked on 30.10.2026.