One Copilot license, two very different risks

Microsoft Copilot only surfaces organizational data that the individual user already has at least view permission to see. That one design choice means the license does not decide what Copilot touches. The person using it does, and so does the job they point it at.

Picture two people in the same company with the same Copilot license on the same day. A sales rep asks it to draft a follow-up email to a prospect. An HR manager asks it to summarize the notes from an open workplace investigation. Same tool, same contract, same line on the software inventory. These are two very different risks, and a governance program that scores the tool will score them the same.

The tool is the wrong unit of risk

Most companies start AI governance with a tool list. Which assistants are approved, which are blocked, who has a license. That list is worth having. I have written about why it goes stale fast. But even a perfect tool list answers the wrong question for a board.

A board does not need to know that Copilot is approved. It needs to know what the company is using AI to do, which of those uses could hurt someone, and who answers for each one. A tool approval says nothing about that. Microsoft describes Copilot as an engine that combines a language model with content in Microsoft Graph, such as emails, chats, and documents that you have permission to access. The model is the same for everyone. The data, and the consequences, change with every use.

So the unit of risk is the business use. One row per thing the company does with AI.

Four ways the two uses differ

Put the sales email and the HR summary side by side and four differences show up. These are the four I would ask any leadership team to look at first.

The data involved. The sales rep works with a prospect's name, a product sheet, and maybe the last email thread. The HR manager works with allegations, witness statements, and personal details about employees who never agreed to have their words fed into anything. Because Copilot follows existing permissions, the HR manager's prompt can reach everything in the investigation folder they can open. Microsoft tells organizations to use the permission models available in Microsoft 365 services, such as SharePoint so the right people have the right access. If those permissions are loose, the use inherits that looseness.

The people affected. A bad sales draft embarrasses the rep and maybe annoys a prospect. A bad investigation summary can shape a decision about someone's job. The person most affected is not in the room, did not write the prompt, and will probably never know a summary existed. That asymmetry is the heart of it.

How much the AI acts on its own. Both of these examples are drafts a human reviews, and Microsoft's own documentation says output is not guaranteed to be 100% factual and should be reviewed. Review is a weaker control than it sounds. A rep rereads an email before sending it. A busy HR manager may lift a summary straight into a findings memo. And once Copilot is connected to agents that it can query on the user's behalf, the question shifts from "what does it write" to "what can it do without a person in between."

Who owns the outcome. The sales use has an obvious owner: the head of sales. The HR use usually has no owner at all. IT owns the license, HR owns the process, legal owns the risk, and nobody owns the point where the three meet. When something goes wrong there, the first meeting is spent working out whose problem it is.

Why this matters at board level

None of this needs a technical reading to understand. It is the same logic a board already applies everywhere else. Nobody rates the risk of "spreadsheets." They rate the risk of the spreadsheet that calculates payroll.

AI readiness at leadership level means being able to answer three questions without a long pause. What are we using AI to do? Which of those uses touch people's livelihoods, money, or private data? Who is accountable for each one? A company that can answer those is ready to expand AI use with some confidence. A company that can only show a license count is not, no matter how many licenses it has.

The other benefit is speed. When risk is scored per use, the low-risk uses stop waiting behind the high-risk ones. The sales team can keep drafting emails while HR's use gets a closer look. Scoring the tool forces one answer for both, and that answer is usually either "block everything" or "allow everything."

The AI use register

The working artifact here is simple: an AI use register. One row per business use. Five columns are enough to start.

1. Owner: the named person who answers for this use, by role and name.
2. Data: what the AI can read or write in this use, including what it can reach through the user's permissions.
3. Who it affects: the people on the receiving end of the output, especially people who are not users.
4. Can it act on its own: whether output goes straight out, or a person reviews it, and whether any agent can take actions.
5. Unknowns: what you have not confirmed yet. This column is the one boards learn the most from.

Here is a small example for the two uses in this post, plus one more.

| Business use | Owner | Data | Who it affects | Acts on its own? | Unknowns | |---|---|---|---|---|---| | Sales follow-up email drafts | Head of sales | Prospect names, product sheets, prior email threads | Prospects, the rep | No. Rep reviews and sends | Whether reps paste pricing or contract terms into prompts | | HR investigation summaries | Not yet assigned | Allegations, witness statements, employee records in the case folder | Employees under investigation, witnesses | No, but summaries may be copied into findings unchanged | Who can open the case folders; whether summaries are retained, and where | | Meeting recap for leadership team | Chief of staff | Meeting transcripts, attached decks | Attendees, people discussed in the meeting | No. Recap is edited before it is sent | Whether board-sensitive meetings are excluded |

Notice that the HR row has no owner. That is normal on a first pass, and finding it is the point. An empty owner cell is a decision leadership has not made yet.

Keep the register short and keep it current. Add a row when a team starts a new use, not when a new tool is bought. Review the rows with blank cells first. The register does not replace a tool list. It sits on top of one and tells you which licenses carry which risks.

Where to start this week

Pick one licensed AI tool, ask three department heads what their teams actually use it for, and write each answer as a row. If any row comes back with no owner or a long unknowns column, start there. A free AI Visibility Check is on this site. The AI Governance Framework is the structure for the rows, and the AI Use Register is the written next step.