Microsoft has rebuilt Copilot. Not the one you hate.

Four people, five Copilots.

Adaeze runs a West African restaurant and takeaway in Birmingham with six staff. She has Microsoft 365 at home for the laptop she does the VAT on, and every time she opens Excel a Copilot button asks her to try something. She has not. Sanjay runs three homewares shops in Leicester plus a Shopify site, 14 staff, stock in a spreadsheet. ChatGPT writes his product descriptions; the Copilot in his Microsoft 365 Business apps has never earned its place. Hannah is a senior HR business partner at a housing association in Gateshead. Her employer pays for the work Copilot, she is the only person in the building who uses it well, and she has become the unofficial helpdesk for colleagues who pressed the button once, got a flat answer and went back to ChatGPT on their phones. Craig is an ex-plumber in Cornwall who builds automations for tradespeople and small firms, and his rule for any tool is simple: if he cannot plug into it, it is dead to him.

All four have met something called Copilot and none of them was impressed. The trouble is that they have each met a different one. Microsoft uses the name for at least five products, most people have only ever used the bottom two, and on 25 September it rebuilt the one near the top. So before the story, the table.

Which Copilot is which.

Read the table once and half the argument is in it. The free one is what most people have and dislike. The home one is the free one with more buttons. The rebuilt one is two rungs up, and even there the clever parts need a second payment. Studio is for Craig, not Adaeze.

Why everyone thinks it is rubbish.

Ask a room of office workers what Copilot is like and you will get an opinion formed in 2023. The name arrived on Bing’s chatbot, then on a Windows sidebar, then on a button in every Office app, and the first paid work version launched at $30 a head to reviews that ranged from polite to puzzled. Since then the free one has been pushed at people from the taskbar, the Start menu and the Snipping Tool, and the paid one has sat behind a licence most employers did not buy. Almost everyone’s idea of Copilot is built on the bottom two rows of that table and on that history. It is a preconception, and it is out of date.

I say that from my own desk, not from a press release. I run a licensed work Copilot inside an organisation that lives in Microsoft: Teams, Outlook, Excel, SharePoint, the lot. On the default setting it is mediocre, and I would not blame anyone for walking away after ten minutes. Pick GPT-5.6 or Claude Sonnet 5 from the model list, switch thinking on, and it becomes the same brain you would pay OpenAI or Anthropic for, with one difference the others cannot match: it can read the Teams thread, the mailbox and the spreadsheet you are asking about without you pasting anything in. The models are eons apart. The button looks the same.

Nobody hates Copilot. They hate the free one, and they have never met the other four. The preconception, in one sentence

What was rebuilt, and whether it is better.

On 25 September Microsoft announced “the new Copilot”, built around three things. Home is the front door: quick chat on one side, and on the other Cowork, where you hand over a whole job (a tender response, a month-end pack) and it goes off and does it, with Word, Excel and PowerPoint built into the app. Code lets someone who cannot programme describe a tracker or an automation in plain English and have it built and hosted inside the organisation’s own Microsoft environment. Autopilot is an agent you give a name, a role and a goal. It keeps working when you have gone home and turns up in Teams and Outlook like a colleague.

Is it better? On three counts, yes, and one of them matters more than the features. First, the documents are real: ask for a budget and an editable Excel workbook appears, shared with the team, showing what changed and why. Second, Cowork now lives in the same window as chat, so handing over a job is no longer a separate product. Third, and this is the one to tell the sceptics, the model list is the point. Inside the one Copilot window you can pick GPT-5.6 or Claude Sonnet 5 on the licence, Opus 5 is coming with limits, and Claude Fable and GPT-6 Astra, the two strongest models on sale, are there on credits. Microsoft says more labs and open models follow. You are no longer choosing between Copilot and the frontier. You are choosing which frontier model reads your inbox.

Who gets it: the Enterprise row first, through the Frontier programme, which an organisation’s IT team has to opt into. Home and Code reach Frontier users “in the coming weeks”, which Microsoft’s admin message centre pins to October. Autopilot goes to private preview at the end of September. The Business row follows. The home row gets a version of Code on the £18.99 Premium plan “later in 2026”. The free row gets nothing; Bloomberg’s headline on the day was that Microsoft had abandoned the personal chatbot race, and earlier this year it started pulling Copilot back from Photos, the Snipping Tool and Notepad after complaints.

Where does your prompt go?

This is the question that matters most if your organisation has rules about data leaving the country, and the marketing does not answer it. Microsoft’s own documentation does, and the answer is different for every brain in the list. Two things surprised me. The UK is not inside Microsoft’s EU Data Boundary, so a UK organisation gets fewer promises than a French one. And since July, some GPT models in Copilot are run by OpenAI rather than by Microsoft, on by default, with no in-country promise.

The plain-English version. GPT run by Microsoft never leaves Microsoft, but for a UK organisation it may be processed abroad at busy times. GPT run by OpenAI is handled by OpenAI under Microsoft’s contract, with nothing kept, and no promise about country outside the EU. Claude is handled by Anthropic under Microsoft’s contract, outside the EU Data Boundary, off by default for UK and EU organisations, and Copilot puts a marker on screen when a Claude brain is answering. Claude with data retention, the Fable 5.0 and Mythos 5 class, is Anthropic’s own contract, stored by Anthropic for up to 30 days, and off everywhere until an admin opts in twice. None of it is used to train models. And the comparison people actually ask about: picking Sonnet inside Copilot and opening the Claude app are the same brain on two different contracts. Inside Copilot, Anthropic sees your prompt under Microsoft’s terms, keeps nothing, and the record stays in your Microsoft 365. In the Claude app, or in ChatGPT, the prompt goes straight to a US company under its own terms, is kept as your chat history, and on a personal account may be used for training unless you say otherwise. Neither route is “kept in the UK”. The Copilot route is the one with a contract, an audit log and an off switch your admin controls.

Same brains as Claude and ChatGPT, two bills.

So “can it compete with Claude or ChatGPT” is partly the wrong question. It is Claude and ChatGPT, wearing a lanyard. Cowork itself was built with Anthropic, on the same foundation as Claude Cowork. What you pay Microsoft for is where the brain sits: inside your organisation’s files and permissions, rather than in a browser tab with your data pasted in. And you pay for that twice: the licence, then the meter.

The everyday work is on a fixed licence. The interesting work, the agents that run for hours and the best models, is on a meter that is off by default until IT switches it on. Analyst Patrick Moorhead, whose firm counts Microsoft as a client, calls it “margin repair and market expansion, not a price cut”, and notes the awkward consequence: the best-scoring models in Copilot are the metered ones. Against that, Claude Pro and ChatGPT Plus are one bill of about £16, and Claude Cowork comes with it.

What the people with early access say.

Cowork has been in the Frontier programme since March, so some people have run it for six months, and their verdict beats the launch video. An IT administrator who has run it in a production tenant since it opened calls it the most operationally interesting thing Microsoft has shipped inside Microsoft 365. Then the caveat.

That pattern repeats everywhere the launch was discussed. On the Hacker News thread, the people who like Copilot at work are the ones who picked a specific model and a thinking mode; the people who loathe it are the ones left on the default “Auto” setting, which picks the cheapest. One commenter’s summary of colleagues: they stick to their very first bad experience. Hannah has watched exactly that happen across three floors of her building. Autopilot has fewer witnesses because it has barely left the building; Microsoft’s own executive in charge of Copilot told a customer event that the autonomy which makes a long-running agent useful is the same thing that makes it “absolutely terrifying to an IT admin”.

How to win the room.

If you are the Hannah in your organisation, the person who knows the paid Copilot is good and cannot get anyone to believe it, the evidence above tells you exactly how to make the case. Five moves.

• Never demo the free one. The preconception was built on the bottom two rows. Every demo happens on a licensed account with a named model and thinking on, or it confirms what people already think.
• Win the first ten minutes. People keep their first bad experience. Have IT set the default model for the pilot group, so nobody meets Auto by accident.
• Show the thing the rivals cannot do. “Summarise what was agreed in this Teams thread and draft the reply in Outlook.” “What changed in this workbook since Tuesday, and who changed it?” “Pull the last three meeting recaps with this supplier.” ChatGPT in a browser tab cannot see any of that. This is Copilot’s own advantage, and it only exists because the brain is standing inside the building.
• Sell the fence to security, not to users. The same models, inside the organisation’s permissions and audit log, is the answer to the shadow ChatGPT already on staff phones. Security will back it before users do.
• Measure two numbers. Minutes saved per run and credits used per run. A pilot with those two figures is a business case. A pilot without them is a testimonial, and testimonials are how Copilot got its reputation in the first place.

Everyday AI is a seat. Delegated AI is a meter. The brains worth having are on the meter. Issue 023 · 29 September 2026

What Adaeze, Sanjay, Hannah and Craig should do.

Adaeze should ignore it, with a clear conscience. Nothing in this announcement reaches her home Copilot this year except a Code preview on the £18.99 Premium plan, and that is more than a ChatGPT or Claude subscription. The free Copilot will write a menu description or a reply to a review about as well as the free version of anything else. When a rep says “Copilot”, her first question is which row of the table they mean. It is usually the bottom one.

Sanjay should wait, which will suit him. Copilot Business at £13.80 a month would be about £193 a month for his 14 staff before the meter, a fifteenth subscription for a business already drowning in them. The part that would actually pay, Code building the tool that stops his staff rekeying stock between the tills and Shopify, is on the meter and not yet in his hands. ChatGPT at about £16 does his product descriptions; Issue 022 showed a free route to building the tracker. Revisit when Code is generally available and somebody has published what it costs to run for a quarter.

Hannah already has the good one and needs IT to let her prove it. Three items make the case: a spending limit set in the admin centre with an alert (until it exists, Cowork, Code and Autopilot are visible but switched off); a pilot of one small group, hers, enrolled in Frontier with the model pinned to GPT-5.6 or Sonnet 5 with thinking on; and one job that lives entirely in Microsoft 365, the monthly people report or the onboarding checklist, measured in minutes and credits. Then run the five moves above.

Craig should build one thing and keep it portable. Copilot Studio is Microsoft’s answer to the tools he already uses: it builds an agent and hands it to a website or an app, and for a client who already pays for Copilot it is included for in-house use. The new runtime even has a developer kit, so his flows could run inside a client’s Microsoft environment with the client’s sign-in and audit trail, which is what those clients will ask for. His risk is the meter: £153.80 a month in credits is a conversation to have before he builds, not after. Keep the logic in n8n, use Studio as the front door, and put a spending limit on it before it goes live.

Before you switch anything on.

Three things for anyone in the UK who has the admin rights, or has to persuade whoever does.

• Know your routes before you pick a brain. The UK is not inside the EU Data Boundary, OpenAI-run GPT models are on by default with no in-country promise, and Anthropic’s models are outside the boundary and switched off by default for UK and EU organisations until an admin enables them. Each of those is a data-protection decision, not a tick box: write them down, and ask where a Code app or an Autopilot keeps what it touches, because Microsoft has not yet said.
• Set the budget before the demo. Nothing on the meter bills until a spending limit exists. Set one for the whole organisation with an alert, then decide who gets credits. New metered features join an existing limit automatically unless you turn that off.
• The public sector has already jumped, in England. NHS England is rolling Copilot to 505,000 staff by October, and HMRC has 32,000 licences rising to 50,000. Those are licence deals, the Enterprise row without the meter. Whether Cowork and Autopilot follow, and on whose credits, is the question trusts and boards will be putting to their account managers this autumn.

Adaeze, Sanjay, Hannah, Craig: the Copilot you hate is not the one that was rebuilt. The one that was rebuilt is two rungs up the table, it runs the brains you already trust, and if your organisation lives in Microsoft it can see things no browser tab can. Find your row, pick a model, then judge it by the bill, not the demo.