Copilot or ChatGPT Work: which does your firm actually need?

If your firm runs on Microsoft 365, start with the Copilot you already partly pay for. Add ChatGPT Work only where you need an agent that works beyond Microsoft's walls.

That is the short answer. The longer answer matters because both products changed this month, and almost every comparison on the web was written before they did.

What changed this month

On 1 July, Microsoft repackaged Microsoft 365. Copilot stopped being a promotional extra, Copilot Chat was folded into the core apps, and commercial prices rose about 16% on average. The new prices reach new and renewing customers first: an existing subscription keeps its price until its next renewal.

Eight days later, OpenAI shipped ChatGPT Work, an agent that takes a goal and hands back finished work (a spreadsheet, a deck, a document) rather than a chat reply. We explained what ChatGPT Work actually is when it launched. The same day, OpenAI announced that its new GPT-5.6 model had become the preferred model inside Microsoft 365 Copilot, across Word, Excel, PowerPoint, Chat and Cowork.

So the two obvious choices for a firm's everyday AI moved in the same fortnight. New bundles and prices arrived on one side, a new agent landed on the other, and increasingly the same engine sits under both. Here is how we would decide between them.

Why the model inside is no longer the deciding factor

Look at what has powered Copilot over the last four weeks. When Microsoft's agent, Copilot Cowork, arrived in the 365 apps in mid-June, it ran on Anthropic's Claude models. Since 9 July, OpenAI's GPT-5.6 has been the preferred model across the suite, Cowork included. That is one product running engines from two different companies inside a month, and you were not asked either time.

ChatGPT Work draws on the same GPT-5.6 family. A firm agonising over which product has the better brain is now often comparing the same brain in two boxes.

There is no point basing a buying decision on which model is smarter this week. The vendors will keep swapping engines, and mostly you will not notice.

What stays still long enough to decide on is everything around the model: where your data sits, what the agent may touch, how the billing works, and whether your people will use it. We call this choosing the wrapper rather than the engine, and it comes down to four questions.

The four questions that decide the choice

1. Where does your data already live? If your documents, mail and client files sit in SharePoint, OneDrive and Outlook, Copilot works inside the file permissions you already run. That cuts both ways. Out of the box, Copilot is exactly as safe as your existing file permissions, and no safer. A folder shared with the whole firm years ago, and forgotten since, is exactly what an assistant will surface.

ChatGPT Work reaches only what you choose to connect, which makes the boundary explicit and yours to draw. The ground rules for client material in these tools apply to both.

2. What is the agent allowed to touch? Copilot Cowork ships off by default, and an administrator decides who gets it and where it acts. ChatGPT Work connects app by app, account by account. Neither default is a policy. Decide which tasks an agent may finish without a person signing off, and write it down before you switch anything on.

3. How does the billing work? The seat price is no longer the whole story on either side. Copilot's agent, Cowork, needs the full Copilot licence and then bills its work in usage credits on top. ChatGPT Work's heavier tasks draw on your plan's included usage, and cost extra only if you add credits beyond it.

A flat seat fee is a budgeting comfort; a metered agent is not. Set a spending cap and an alert before the first ambitious job, not after the first surprising invoice.

4. Will your people actually use it? The assistant inside the apps your team already opens wins by default. A separate app only gets used if people remember to open it. A recruitment consultant drafting adverts in Word meets Copilot without changing a single habit. Asking the same consultant to carry the task into a new app is a behaviour change, and behaviour changes fail quietly.

The two products side by side

The differences that persist are reach, governance and billing, and the table lays them out.

Prices are US list prices. What a firm pays in the UK varies by reseller and agreement, so check your own renewal terms rather than budgeting from a table, including this one.

So which one should your firm buy?

For most firms in our audience, the sequence is plain. If the work is internal and your firm lives in Microsoft 365, start with the Copilot Chat that now comes bundled. Buy the full Copilot licence only for the people who need it inside the apps. Switch on Cowork last, for selected users, once low-risk internal tasks have shown what agent work is actually worth to you.

Reach for ChatGPT Work when the job sits mostly outside Office. Think of an agent that needs your project tool, your accounts package and the open web in one task, or of finished work that does not live in Office documents at all.

Copilot can reach some outside tools through connectors, so treat this as a guide rather than a strict rule. Trial ChatGPT Work on one task you already know how to judge.

Running both is rational for some firms, provided each has a distinct job and a distinct owner. An accountancy practice might sensibly give fee-earners Copilot inside Excel and Word, while the marketing manager runs ChatGPT Work for the monthly report that pulls from six tools. We would not buy both for everyone by default; that spreads the firm's attention across two subscriptions and makes it harder to prove the value of either.

And there is an honest last answer: neither, for now. If your SharePoint permissions are a mess, tidying them is duller than buying an AI and more valuable. Clean permissions are the precondition for the bundled option working safely. Working out which of a firm's real tasks justify which tool, and in what order, is much of what we do in AI Lessons for Leaders.

The skill that outlasts this choice

Whichever wrapper you pick, the return comes from the same place. In Management as AI superpower, Ethan Mollick argues that as capable AI becomes cheap, the scarce skill is delegation: "Now the 'talent' is abundant and cheap. What's scarce is knowing what to ask for." The people who thrive with these tools, he suggests, are the ones who know what good looks like and can explain it clearly.

That is a management skill, and it transfers. Licences do not. A firm that can specify a task and judge the output will get value from either product, and a firm that cannot will get noise from both.

So this week, pick one task your firm does often and knows how to judge. Run it through the tool you already pay for. Mark the result the way you would mark a capable junior's first attempt.

Then make the licence decision with evidence in hand rather than a feature list. If you want a second pair of eyes on it, book a discovery call.