An AI agent is now in your Office tools. Where do you let it act, and what does it cost?

The most consequential AI decision on your desk this month is one you did not go looking for. You did not evaluate a tool, sit through a demo or sign a new contract. An autonomous agent simply arrived inside the software your firm already runs on. On 16 June, Microsoft made Copilot Cowork generally available worldwide, an agent that takes a goal and works through a multi-step task across Outlook, Teams, Word and Excel, then hands back a finished result rather than a draft. Two decisions follow from that, and neither is technical. Where do you let it finish work on its own, and what is the new bill worth.

Start with what genuinely changed, because the noise around agents makes it easy to miss. Copilot Cowork is off by default; an administrator has to switch it on and decide who gets it. When it runs, in Microsoft's own words, it "runs end-to-end and returns a completed result, not just a draft or a recommendation." That last part is the shift. The Copilot you already know suggests a sentence, drafts an email, summarises a thread, and leaves you holding the pen. An agent that returns finished work is a different kind of colleague. It is also billed differently, which we will come to, and at general availability it runs on Anthropic's Opus and Sonnet models inside Microsoft's cloud. None of this requires you to become technical. It requires you to make two calls you would make about any new member of staff: what they are allowed to finish without you, and what they cost.

Decision one: where the human stays on the work

A draft invites a check. A finished result invites a rubber stamp. That is the quiet risk in an agent that completes a task end-to-end, and it is sharpest in exactly the work your firm is paid for. When a tool hands you a half-built thing, you naturally finish it, and in finishing it you read it. When it hands you something that looks done, the temptation is to glance and forward. The more capable the agent, the stronger that pull, and the more it matters that you decided in advance where it does not apply.

The useful move is to draw the line before you switch anything on, not after the first unchecked output reaches a client. Some work is fine for an agent to complete on its own: pulling figures from a set of spreadsheets into a summary table, comparing two versions of a long document and listing what changed, drafting an internal status note from a Teams thread. The cost of an error there is low and a person sees it before it travels. Other work is not, and the test is simple: anything that carries your firm's judgement, reaches a client, or touches compliance-sensitive ground keeps a human on it. A first-pass diligence summary in corporate finance, a note that interprets a regulation for a client, an advice email that goes out under a partner's name. The agent can do the gathering and the first pass. The judgement, and the sign-off, stay with the person whose name is on the work. We go into where that line sits, task by task, in what to keep off an AI's desk.

This is not caution for its own sake. An agent that runs while your laptop is off, across your real email and documents, is genuinely useful for the grunt work that clogs a week. The point is to be deliberate about which jobs those are, so the time you save on the gathering is not lost, with interest, to an error you waved through because it arrived looking finished.

Decision two: the cost shape just changed under you

The second decision is the one most leaders have not clocked yet, because it hides in the billing. A per-seat licence is a known number. You pay a fixed amount per person per month and you can put it in a budget and forget it. Copilot Cowork does not work that way. It is charged on usage, in what Microsoft calls Copilot Credits, on top of the existing Copilot licence, with the price of each task built from four things: how much the model works, how much context it pulls, how many tools it calls, and how long it runs. Pay-as-you-go is priced at a cent per credit. So a heavy task that reasons across many sources and produces several outputs costs more than a light one, and your monthly bill now rises and falls with how much work the agent does.

That is not a reason to avoid it. It is a reason to govern it. Microsoft has built the controls in: spending limits at the level of the whole organisation, a team or an individual, alerts when spend crosses a threshold you set, and reporting on where the credits went. Use them before you turn the agent on, not after a surprising invoice. Decide a budget, set a cap, and review the first month's usage against the work it produced. The honest question a usage bill forces is a good one: is this task worth what it costs to have the agent do it. For a lot of low-value busywork the answer will be yes and cheap. For some work it will turn out you were paying an agent to do something nobody needed done at all, which is worth finding out.

Both decisions point the same way

Notice that neither decision is about the model. They are about how your firm works. That is the pattern under almost every AI result worth having. A recent survey of 1,550 enterprise AI decision-makers found that 73 percent now use AI across most of their processes, while only 10 percent say it is core to how the business actually operates. The gap between those two numbers is not a tools gap. It is the distance between switching something on and changing how the work is done around it.

An agent in your Office tools makes that gap concrete. Switching it on is the easy part, and on its own it changes little. The value shows up only if you decide what it should do, where a person stays in the loop, and what it is worth spending. That is a leadership question, not a procurement one, and it is the work we do in AI Lessons for Leaders: turning a capability that just landed in your tools into a few clear decisions and a way of working your team will actually follow.

So before the agent runs its first real task for your firm, settle four things. Decide which jobs it may finish without a human signing off, and write the rest down as work that always comes back to a person. Set a spending cap and an alert so the bill cannot surprise you. Pick one or two real tasks to try first, on internal work where a mistake is cheap, rather than letting it loose across client files on day one. And put a date in the diary to look at what it actually did and what it actually cost, because the first month's usage will teach you more than any pricing table. If you want a second pair of eyes on where your line should sit, book a discovery call and we will work through it against your real client work.

Frequently asked questions

What is Copilot Cowork?

Copilot Cowork is an AI agent inside Microsoft 365 that carries out complex, multi-step tasks from a single instruction, working across apps like Outlook, Teams, Word and Excel and returning a finished result rather than a draft. Microsoft made it generally available worldwide on 16 June 2026. It is off by default, so an administrator has to enable it and choose who can use it.

What does Copilot Cowork cost?

It is billed on usage, in Copilot Credits, on top of the Microsoft 365 Copilot licence rather than as a flat per-seat fee. The price of each task depends on how much the model works, how much context it retrieves, how many tools it calls and how long it runs. Pay-as-you-go is priced at one cent per credit, so heavier tasks cost more, and the monthly total varies with how much the agent is used.

Should we turn it on?

Only after you have decided two things: which tasks the agent may complete without a person signing off, and what you are willing to spend. Set spending limits and alerts first, start with low-risk internal work, and review the first month's usage against the value it produced. Turning it on is reversible and cheap; letting it run unsupervised across client work is the part that carries real risk.

What should we never let an agent finish on its own?

Anything that carries your firm's judgement, goes to a client, or touches compliance-sensitive work. An agent can do the gathering and the first pass on that work, but the interpretation and the sign-off stay with the person whose name is on it. Keep the low-risk, internal, easily-checked tasks for the agent to complete, and keep a human on the rest.