ChatGPT Won the AI Head-to-Head. Here's Why I'd Still Choose Microsoft Copilot.

I watched a head-to-head comparison recently: Copilot, Claude and ChatGPT, run through three ordinary office jobs. Pull an action plan out of a messy project folder. Build a forecast in a spreadsheet. Write a difficult client email. Same prompts, same scoring, three runs each.

ChatGPT won. It was the quickest, and it was the only one that handled the awkward client email cleanly. Claude had a rough time in the spreadsheet. Copilot landed in the middle — it went and found a reference file in SharePoint that nobody had told it about, then turned around and softened a fixed deadline and put the wrong name against a task.

It's a fair test. It's just not the question you're actually buying an answer to.

Four things the head-to-head test didn't measure

• The copy-paste tax. For Claude or ChatGPT to touch your real work, the work has to leave your systems first. Every task starts with a copy, a paste, and a small judgement call about what you're comfortable sending.
• The plumbing cost. Yes, you can wire those tools into your files with connectors — on the right plans, per person, set up by someone who knows how. That's an afternoon for a team of five. It's a project for a team of fifty.
• Where the data lands. Drop a client contract into a browser chat and it's sitting on someone else's servers, quite possibly in another country. In health, legal or government work that isn't a preference — it's a line you're not allowed to cross.
• How small the gap actually was. All three got the job close and none of them got it finished. Every one of them invented something. You were going to check the output either way.

Why the margin between first and third place almost doesn't matter

That last point is what I keep coming back to. The distance between first place and third was small enough that it vanishes the moment you count what it costs to get the work in front of the tool.

Because Copilot is already sitting there. Slide the pane open in Excel, in Word, in Outlook, and it's looking at the file that's already on your screen. Nothing travels. Nothing gets pasted. Nobody has to be granted anything. For that to be worth giving up, the alternative would have to be dramatically better — and it wasn't. It was a bit better, sometimes.

The "which AI is smartest" question is becoming less relevant

Here's the part most of these comparisons skate past: the intelligence underneath Copilot is largely the same frontier intelligence being tested against it. Microsoft routes to whichever model suits the job. So "which one is smartest" is quietly turning into a setting rather than a purchase — and it'll keep updating every few months whether you pay attention or not.

Even the sharpest hit in the whole test has a use-by date on it. Copilot didn't think to go looking for a signed contract sitting in SharePoint — but the moment it was pointed at the thing, it found it and reached exactly the conclusion I'd have reached myself. That's a "not yet", not a "no". The layer that lets it reason across your own files and messages is new, and it is getting better on Microsoft's clock, not yours.

The tool that wins the test rarely wins the job

The tool that wins the test and the tool that wins the job are almost never the same tool. A test scores the answer. Your business pays for everything around the answer — getting the work there, wiring the accounts up, keeping track of who has access to what, and knowing where the file ended up. At five people you can absorb all that yourself. At fifty it becomes somebody's actual job.

Which leaves Copilot with the opposite problem: being everywhere makes it confusing. Do I use the one in the browser, or the one in Word? Is the Excel one different? That's not a tool failure, it's a training gap, and it's the cheapest gap you'll ever close. The instruction your team needs is one sentence long: whatever app you're in, use the Copilot that's in it.

And set the expectation honestly while you're there. That report someone was going to spend two hours building by hand? Copilot drafts it in five minutes and they spend forty putting it right. That's still well over an hour back — on time you'd already written off.

Copilot isn't trying to be the frontier. It's trying to be the conduit.

The real question: how far does your work have to travel before AI can help?

If the honest answer is "out of our systems, into a browser tab, and back again", that's not an AI problem — it's a plumbing problem, and it's usually the thing quietly capping how much your team bothers using any of this. Think about where your work actually lives before you choose which tool to bet on.

If you're running Microsoft 365 and want to close that gap, a good place to start is understanding what a proper M365 setup actually unlocks for your team — or if you want structured help getting your people using Copilot day-to-day, take a look at the AI Operator programme.