The managing director of an engineering firm signs the quote for Microsoft 365 Copilot. IT assigns the licences; the SharePoint permissions have been cleaned up. Three weeks later a project lead asks whether she may summarise a client contract in Copilot, and whether she can still use ChatGPT for her English emails. Nobody knows, because there is a licence, and there is no policy.
That is the difference this article is about. The licence settles what Microsoft does with your data. The AI policy settles what your own people may do with it, and no licence covers that. Six agreements are enough, and they fit on one page.
Why is a licence not yet a policy?
A Copilot licence answers a vendor question. Microsoft 365 Copilot does not use prompts, responses or Graph data to train its underlying models, processing stays within the Microsoft 365 service boundary, and a data processing agreement sits underneath. What that means exactly, and where its edges are, is set out in the privacy risks of Copilot.
The policy answers a different question: what an employee may do with it on a Tuesday afternoon. May a client contract go in, a job application, a medical report? Who switches on a connector to an external source? And what applies to the tools people use alongside it? Microsoft has no answer to those, because they are questions about your organisation.
The wider decisions around Copilot, from permission structure to who assesses a new use, are in Microsoft Copilot governance. This article is about the policy that follows from them: short enough to read at the moment it matters.
Which six agreements do you make?
| Agreement | What you record | Why it is in there |
|---|---|---|
| Who | Who holds a licence for the paid variant, who works with Copilot Chat | A licence is a choice per person, and that choice determines what Copilot can read for that person |
| Which variant | Only signed in with a work account; the consumer version gets its own status | Three services are called Copilot, and one of them has no business data protection |
| Which data | Per category: may go in, may go in under a condition, never goes in | Microsoft's training commitment says nothing about what may be sent |
| Which connectors | Which connectors, agents and plug-ins are on, and who signed off | Every connector widens what Copilot can read, and so what can end up in an answer |
| The other tools | A status per AI tool next to Copilot, with Copilot as the preferred route where that fits | A Copilot policy that is silent on ChatGPT leaves most of the use unaddressed |
| The measurement | Which figures you look at, how often, and that no person is in the data | Without measurement compliance is an assumption; with the wrong measurement it is surveillance |
Each row is one or two sentences in the policy, with a name and a date. Where no name fits, the agreement has not been made yet.
How short can it be?
Shorter than most organisations dare. The engineering firm's policy ended up like this: licence holders are on IT's list; Copilot only with a work account, the consumer version is not allowed; client data may go into the paid variant, personal data of applicants and staff may not; connectors are switched on by one administrator and listed with an owner; ChatGPT and Perplexity have limited use with public information only, Copilot is the preferred route; the share of use through allowed tools is reviewed each quarter, without names.
That is half a page. A lawyer checked it, which is sensible, because the sentence about applicants touches the GDPR and the sentence about measurement touches the works council rules. But the sentences themselves belong to the firm, not to a template. How to build a policy like that from the questions people really ask is in writing an AI policy.
What do you do about the tools next to Copilot?
This is the agreement most often missing, because the policy is called "Copilot policy". When the firm looked, a month after rollout, at which AI tools were being opened in the browser, Copilot was at the top. Below it were ChatGPT, Perplexity, a translation tool and two tools for reading technical drawings. Nothing had ever been decided about the last two, because nobody knew they existed.
Nobody had broken the policy. The policy was about Copilot, and these tools were not Copilot. Why people use other tools alongside Copilot is its own question; for the policy, what matters is that every tool in use gets a status. Allowed, limited use with a condition the organisation writes itself, no decision yet, or not allowed. Copilot is then one row in that register, and the rest sit underneath it.
