Copilot vs Claude vs ChatGPT for government teams: choosing what to train on

Before anything else: my bias

I train teams on Claude. Fautons is a partner in Anthropic's Claude Partner Network. If you are about to read a tool comparison written by someone who makes money from one of the tools, you should know that before the first opinion lands, not in a footnote at the bottom of the page.

So hold me to a test. If you run a department that has already committed to Microsoft Copilot, you should finish this piece and think it was fair to you. If I can't clear that bar, the comparison isn't worth writing, because a comparison that hides where it's coming from is just an advert with headings.

And here is the short version of what I actually think, which is probably not what a Claude specialist is supposed to say: for most public sector teams the right answer is not one tool. The first move is almost always to train people properly on the one you already have.

The classic mistake: picking a tool before defining the work

Nearly every conversation I have with a public sector team starts in the wrong place. The question arrives as "which AI should we roll out?" before anyone has written down what they want it to do. That order guarantees you'll be sold to, because with no definition of the work, every demo looks impressive and none of them are comparable.

The Cabinet Office made its 2025 "One Big Thing" campaign about exactly this gap, themed "AI for All", with the aim of every civil servant becoming a confident, responsible AI user. It cited that only 28% of civil servants felt confident using AI at work. Read that number carefully. It is not a statement about which model is better. It is a statement about people not knowing how to use any of them well.

So before you compare anything, spend an hour with your team and write down the tasks that actually eat your week. In government the list usually looks something like this:

• Drafting a submission or briefing note from scattered inputs and half-finished notes.
• Getting through a long consultation response, an inspection report or a policy annex without losing the thread.
• Rewriting a service letter into plain English at a reading level your users actually have.
• Pulling structure out of a messy spreadsheet, or out of a meeting you sat through and half-remember.
• Working through an analysis where the reasoning matters more than the final paragraph.

Once that list exists, most of the tool question answers itself, and you can tell within two minutes whether a demo you're being shown is relevant to your work or just good theatre.

Copilot: it is already in the building

Many UK government departments already have Microsoft Copilot through their existing Microsoft 365 estate. That makes it the tool already sitting in front of most staff, whether or not anyone has trained them on it.

People in my line of work under-rate that, and they under-rate it because it's commercially inconvenient. Your documents are already in SharePoint. Your mail is in Outlook. Your meetings are in Teams. A tool that lives where the work already lives, that is already licensed, already through your assurance process and already on the taskbar, has a head start that no amount of model quality erases for everyday tasks. Access beats elegance more often than vendors want to say out loud.

So here is the advice that costs me money: if your department already has Copilot and your staff are not using it well, training them on it is very likely the highest-return thing you can do this quarter. No new licence, no procurement, no change programme. Just people learning to use the thing that is already on their screen.

Where I see teams get stuck with it is expectation, not the tool. Someone types one lazy sentence, gets back something bland and safe, and concludes that AI is overhyped. That would happen in any of these tools. It is a skills problem wearing a technology costume.

ChatGPT: the one they have already touched

ChatGPT is the one most of your staff have already used, usually on their own phone, usually before anyone official told them anything about it. That is both an asset and a liability, and it is worth being clear-eyed about both halves.

The asset is that the fear barrier is gone. People who have already asked it to write a best man's speech or explain their mortgage are not frightened of a blank prompt box. In a room where only a minority feel confident with AI at work, that head start is real and you should use it.

The liability is that habits formed on a personal account do not carry your department's data rules with them. Someone who has been pasting whatever they like into a consumer tool at home will, without a moment's malice, do the same thing at their desk. That is not a ChatGPT problem. It is a policy and training problem, and it will follow you into whichever tool you choose.

Worth knowing: the Government Campus AI prospectus sorts AI learning into Foundation, Working and Expert tiers, and its published course listings include both Microsoft Copilot and ChatGPT-oriented courses. Whichever of the two you land on, you are not going off-piste. There is an official ladder you can map your training onto.

Claude: where I find the quality lives

This is the section where my bias is loudest, so I'll keep it narrow and I'll keep it honest.

The pattern I see, across the teams I actually train, is that for careful drafting, for long-document work and for reasoning through an analysis where you want to follow the thinking rather than just receive an answer, Claude is where the quality lives. That is my judgement from doing this work most days. It is not a benchmark result and I am not going to dress it up as one, quote a score at you, or tell you it wins by a percentage. If I did, you should stop reading.

What I would rather you did is test it, because you can, cheaply, this week. Take one real piece of work, a genuine briefing or a real annex with the sensitive material stripped out, and run it through whatever you already have and through Claude. Then judge the two outputs the way you would judge a draft from a new starter. Which one would you send back, and how many times?

And here is what I will not claim: Claude does not beat Copilot at the thing Copilot is genuinely best at, which is already being inside your documents, your calendar and your mail. For a large slice of ordinary daily work, that proximity matters more than anything I can tell you about drafting quality. If you want to see what deeper, hands-on Claude and AI training in 2026 looks like inside a government team, I've written that up separately.

What actually transfers: judgement, and a policy people can recite

Strip away the vendor noise and there are two things worth training, and neither of them is a tool.

The first is judgement about which tool suits which task. That sounds soft. It isn't. It is the difference between a person who uses the assistant in front of them for a summary and knows when to take the careful drafting somewhere else, and a person who bludgeons every job with the first icon they see. Judgement is the only part of this that survives the next release, the next licence change and the next reorganisation. Models will change under your feet. The habit of asking "what kind of task is this, and what is it worth checking?" will not.

The second is data handling, and I want to be blunt about where that responsibility sits. Whatever tool you choose, the thing likely to cause you an actual incident is not the model. It is a well-meaning person pasting in something they should not have. Your staff need to know, in plain words, what can and cannot go into the tool. Not the licence tier. What information, in what circumstances.

That is a question for your department, for your data protection people, your security team and your SRO. It is not a question for a vendor, and any vendor who tells you their product makes it go away is selling. The practical test is whether the rule fits in a sentence your staff can repeat from memory. If your policy needs a slide deck to explain, it will not survive a busy Tuesday afternoon. If you want a plain-English version of the underlying principles, I've set out what responsible AI use actually means for a civil servant in practice.

So, the honest recommendation. Train people on what they already have, because that is where the compounding starts. Teach them the second tool for the work the first one handles badly. Put a data rule in front of them that they can actually remember. Then, and only then, argue about which model is best. If you want help shaping that, I run AI and Claude training for public sector teams built around the work you actually do, and I will tell you when the answer is to use what you've got.