Three categories. They are not about capability, they are about who carries the consequence. Copilot will attempt all three, confidently.


Numbers that have to reconcile

A budget, an invoice, two cost reports, a headcount that has to tie to payroll. A model returns a total that looks right and matches neither source. Use a formula, a query or a script, then use Copilot to explain the result to a human.

Judgment you will be held to

A hire, a go or no-go, a performance rating, a supplier you are about to drop. The draft is not the problem. The problem is that a fluent draft makes the decision feel already made. Write the call yourself, then ask Copilot what you missed.

Regulated data on the wrong surface

Client personal data, anything under a data-residency commitment, anything covered by an NDA, pasted into a surface whose route you have not confirmed. Confirm the route first. Once it is in, it is in.


The safety-critical line

<aside> ⚠️ AI prepares, humans decide. Nothing on this page is an argument for using Copilot to authorise safety work. A permit to work, a lockout, a confined-space entry, a job safety analysis, an incident classification, an inspection sign-off: a person signs those, every time, and Copilot is not in that loop.

</aside>

If your industry has a version of that list, write it down and add it here before you hand this page to anybody else. There is a page for that: Copilot decision log, one row per decision. The row is finished when somebody's name is in it.

Copilot decision log


Part of Copilot on One Page, by Mathieu Kessler, independent practitioner: kesslernity.com. Checked against Microsoft's own documentation in August 2026. Free to use and print, personally or inside your business. Keep the source credit. Full terms: kesslernity.com/license.