The Four Microsoft Copilots

There are four Copilots, not one

Here's the whole mess in one place. Read this section once and most Copilot confusion disappears permanently.

1. Free Copilot

The assistant at copilot.com, in the Windows 11 taskbar, in Edge, and in the phone apps. Sign in with a personal Microsoft account and it costs nothing. You get web-grounded chat with citations, image generation, file uploads, voice conversation, memory, and a few workspace features called Pages and Notebooks.

Microsoft no longer publishes usage limits for the free tier. The exact wording is that you can chat, create images, and upload files for free "subject to available capacity and limits." Every specific number you'll find online, the "300 messages a day" figure and its cousins, traces back to July 2024 and is obsolete. Nobody outside Microsoft knows the current caps, and that's worth knowing before you build a workflow on top of them.

2. Microsoft 365 consumer subscriptions

This is where most solo operators and creators actually meet Copilot, usually by upgrading a subscription they already had.

Two things people consistently get wrong here, and the second one is expensive.

The first: Family does not give your family Copilot. Microsoft's own footnote says AI features are available only to the subscription owner and cannot be shared. Five other people on the plan get Office and storage. They do not get the AI.

The second is the big one. These consumer plans do not include the work-data grounding that all the Copilot marketing is about. You get Copilot living inside your Office apps, which is useful. You do not get the version that reads across your whole organisation's email, files, and chats. Buy Microsoft 365 Personal expecting Copilot to know everything about your business, and you'll be disappointed, and you'll probably blame the AI when the real problem was the subscription tier.

3. Microsoft 365 Copilot, the work product

This is the one Microsoft means on an earnings call. It's the version with the actual differentiator: it queries your organisation's mail, files, chats, meetings, and calendar under your existing permissions.

Prices are per user per month. The $18 Copilot Business price is a promotion running through 31 December 2026.

Every one of these is an add-on. Microsoft's own line: "A separate license for a qualifying Microsoft 365 plan is required." The cheapest realistic base plan is Microsoft 365 Business Basic at $7 a user per month. So if you're a one-person business and you want the real work-grounded Copilot, your floor is roughly $25 a month, or $300 a year. Hold that number, we'll come back to it.

There's also no trial. Microsoft states plainly that no trial is available for the paid Microsoft 365 Copilot plan.

4. Copilot Studio

The platform for building your own agents. It's free to licence and charged on consumption. We'll deal with it in Section 06, where the answer for solo operators is short and it's a no.

What changed in 2026

Copilot moved more in the past nine months than in the two years before it. If your mental model is older than this summer, it's stale.

The apps merged. In August 2026 Microsoft folded the consumer Copilot app and the Microsoft 365 Copilot app into one app, with an account switcher separating personal from work. Chat history and files migrated automatically. Mobile and web landed in mid-August. The Windows and Mac desktop versions are rolling out this month, so if your desktop app looks different from a screenshot you saw last week, that's why.

Five features were killed on 18 August. Deep Research, Podcasts, Group Chat, Copilot Labs, and the Mico character in voice mode. Deep Research had a successor called Researcher, but it moved to the $19.99 Microsoft 365 Premium tier, skipping Personal and Family entirely. Podcast content and group chat threads were not migrated and there was no export option, so anyone who had made things there lost them.

Microsoft pulled Copilot back out of Windows. In March 2026 it removed Copilot from Photos, Widgets, Notepad, and Snipping Tool, and shelved planned integrations in Settings and File Explorer. The Windows executive who announced it wrote that it was "critical that we remove Copilot from places where it doesn't live up to its promise," then deleted the post. In May, Microsoft conceded that the floating Copilot button in Word, Excel, and PowerPoint "was a mistake" and shipped an option to hide it. A company walking back its own AI features is unusual and worth noticing.

The Excel formula died. A worksheet function called COPILOT() was announced, previewed, and then cancelled before it ever shipped properly. It retires on 14 September 2026. Microsoft's statement: "We have decided not to move forward with this feature." More on why in Section 05.

How Copilot actually differs from ChatGPT and Claude

The usual answer is "it's Microsoft's ChatGPT." That's wrong in a way that leads people to buy the wrong thing, so let's be precise. There are two real differences and one of them is a genuine advantage.

Copilot is not a model. It's a router in front of several.

When you ask ChatGPT a question, you know which model answered, because you picked it. Copilot doesn't work that way. Microsoft runs a router that sends your prompt to whichever model its own systems choose, and as of September 2026 that pool spans three different companies:

• OpenAI. GPT-5.6 was announced in July 2026 as the "preferred model" for Microsoft 365 Copilot.
• Anthropic. Claude became an official Microsoft subprocessor in January 2026 and Claude models are now selectable across Copilot, Researcher, and the Microsoft 365 apps. They're on by default outside the EU, the UK, and EFTA countries, and off by default inside them.
• Microsoft's own MAI models. Announced at Build in June 2026. Bloomberg reported in July that Microsoft had started routing Copilot traffic to them inside Excel and Outlook, and the stated reason was cost.

Microsoft does not tell you which model answered your question. In fact, the model picker in Microsoft 365 Copilot lost its model names in July 2026. Where it used to say GPT-5.5, it now says Auto, Quick response, Think deeper, and Opus. The framing shifted from which model to how much thinking.

This matters practically, not philosophically. It means a Copilot feature that behaved one way in spring can behave differently in summer without anyone telling you. It means any quality test you run has a shelf life of a few weeks. And it's worth knowing that on the public SWE-bench Pro coding benchmark, Microsoft's own MAI models score at the bottom of the pool it draws from, around 52 percent, against roughly 80 percent for the strongest Claude model. Copilot is increasingly being routed to the weakest models in its own lineup, for reasons that are about Microsoft's costs rather than your output.

Copilot reads your work data. That's the actual product.

This is the real advantage and it's a genuine one. Through Microsoft Graph, and a newer layer called Work IQ, Copilot can query your organisation's mail, files, chats, meetings, and calendar. It does so under your existing permissions. Microsoft's commitment is explicit: Copilot "only surfaces organizational data to which individual users have at least view permissions."

What that buys you is a category of question no chat window can answer. "What did we agree with this client back in March, and what have we shipped since" is answerable in work-grounded Copilot. It is not answerable in a tool that has never seen your files. That's a real difference, not a marketing one.

Microsoft's data commitments are also strong and worth stating fairly, because they're Copilot's best argument. Prompts, responses, and data pulled through Microsoft Graph are not used to train the underlying models. Copilot respects sensitivity labels. It's covered by GDPR, ISO 27001, and HIPAA commitments.

There's one live exception that matters if you have European clients. Microsoft's own documentation states that Anthropic models running inside Copilot "are currently excluded from the EU Data Boundary." So the data residency guarantee, which is the single thing most often cited as Copilot's enterprise advantage, is conditional on which models your organisation allows. That's a real detail and almost nobody is writing about it.

How to actually use it

This is the practical core, and it rests on a single rule that everything else follows from.

Outlook: this is the strongest case

If you take one thing from this article, take this one. Email is the only Copilot benefit that survives objective measurement in a properly designed trial, and we'll look at that evidence in the next section. If your inbox is a meaningful part of your job, Copilot in Outlook earns its keep.

What to use it for:

• Thread summaries. Open a long chain and ask for the summary before you read any of it. This is the highest-value thirty seconds in the product.
• Draft replies with a specified tone. Don't just say "reply." Say "reply declining, warm but firm, mention I'm available in October."
• Coaching on a draft you wrote. It reviews tone and clarity before you send. Genuinely useful on anything difficult.
• Meeting prep. Ask what you need to know before a call with a particular person, and it pulls from prior threads.
• Natural-language inbox management. Added in August 2026. You can ask it to organise messages, apply categories, and create rules by describing what you want.

Word: first drafts and catch-up

Copilot in Word drafts, rewrites, summarises, and answers questions about the document you have open. The 2026 additions are more interesting than the originals. Copilot Catchup summarises what changed since you last opened a document, which is quietly excellent for anything with collaborators. It also builds and maintains document structure, so tables of contents, headers, and footnotes stay correct as you edit. And it can act on edits requested in review comments, which saves a lot of tedious clicking.

One hard limit to know. Microsoft's stated guideline is that Copilot can handle documents up to 1.5 million words. In practice, users have documented a 65,000-word file being silently truncated, and a Microsoft moderator confirmed the guideline is "not a guarantee." You get no warning. So if you ask for a summary of something long and the answer feels thin, assume it only read part of it. For anything past roughly 20 to 30 pages, check that the summary actually covers the end of the document.

PowerPoint: better than you'd expect

This one has improved sharply and gets less credit than it deserves. Copilot builds decks from a prompt, from an existing document, or from a folder in OneDrive. That last one is genuinely useful: point it at the Word document you already wrote and let it draft the deck.

The 2026 additions matter if you deliver work to clients. There's an admin-approved Brand Kit, a strict brand adherence mode that locks Copilot to approved templates and blocks layout changes, and "note steering," where you write plain-language instructions in the slide notes and Copilot follows them slide by slide. That last one is the closest thing to directing it properly.

Teams: recaps and transcript search

Meeting recaps are consistently among the highest-adopted Copilot features in every study, and for good reason. Ask questions of a transcript after the fact, get action items captured automatically by the Facilitator agent, and search the last thirty days of meetings in one place.

Worth knowing: after user complaints, Microsoft added a mid-meeting toggle in July 2026 so organisers and presenters can switch meeting AI off during a call. If you record client calls, tell people the AI is running. It's a courtesy that costs nothing and prevents an awkward conversation later.

Excel: use it carefully

We'll cover why in the next section, but the short version is that Excel is Copilot's weakest area by a wide margin, and it has been measured as such. The newer Agent Mode is materially better than the older in-app Copilot, and it can now use Python for analysis. But it scores 57 percent on the standard spreadsheet benchmark against a human baseline of 71 percent, which means it's meaningfully worse than a competent person.

Use it for formula generation and explanation, for conditional formatting, and for describing what a workbook contains. Check every number it produces. Never paste a Copilot-generated analysis into something a client sees without verifying the arithmetic yourself.

How to prompt it well

Copilot rewards specificity more than most tools, because it has to decide what to go and fetch before it answers. Three habits do most of the work:

What the evidence actually says

Here's the unusual thing about Copilot: the evidence base is excellent. Several governments rolled it out to thousands of staff and published the results. Microsoft's own economists ran randomised trials measuring actual behaviour rather than asking people how they felt. Almost no other AI tool has been examined this carefully.

The results are more complicated than either the marketing or the backlash suggests, and knowing them makes you better at using the tool.

The finding that shapes everything else

Nearly every "time saved" number you've seen about Copilot is self-reported. Somebody was asked how much time they thought they saved, and they answered. The handful of studies that measured what people actually did, using software telemetry rather than surveys, found much smaller effects and in several cases none at all.

The best-designed study in existence is a randomised field experiment across 66 companies and 7,137 knowledge workers, published as a US National Bureau of Economic Research working paper and revised in November 2025. It read outcomes from Microsoft 365 telemetry rather than asking anyone.

It found one clear win: two hours a week less time in Outlook among people who actually used Copilot, along with 10 percent fewer emails read and about two extra hours of uninterrupted, email-free work time.

It found no statistically significant change in Teams meeting time, meetings attended, document completion time, number of documents completed, or collaborative document production. In the authors' words: "Apart from these individual time savings, we do not detect shifts in the quantity or composition of workers' tasks."

The government trials

The satisfaction paradox

Look at those two right-hand columns together and you'll see the most interesting pattern in the whole corpus. Satisfaction is consistently high. Measured productivity is consistently absent. 95 percent satisfied at DWP with no output effect. 72 percent satisfied at DBT with no robust productivity evidence. In Australia, 86 percent wanted to keep it while 61 percent of their own managers couldn't tell which work it had touched.

People genuinely like Copilot. That's real and it shouldn't be dismissed, because a tool that makes a tedious job less tedious has value even if a spreadsheet can't find it. It's just not the same claim as making people more productive, and the entire sales pitch depends on treating them as one thing.

The Excel problem is real and specific

This is the most thoroughly evidenced weakness, and it comes from independent sources that agree with each other.

The UK Department for Business and Trade measured it directly: Copilot users "completed Excel data analysis more slowly and to a worse quality and accuracy than non-users." That is a measured negative effect, not a bad review. The Australian evaluation named "poor Excel functionality" as a specific barrier. Excel adoption topped out at 23 percent in the largest UK trial, the lowest of any core app. And the COPILOT() worksheet function was cancelled before general availability, with Microsoft itself conceding that results "may therefore change over time even when the formula receives the same arguments."

Independent testing of that function found it omitting airports from a list of airports, sorting US presidents by the wrong column, and returning 194 UN member states instead of 193.

None of this means avoid Excel Copilot. It means treat its output as a draft that a human checks, which is exactly how you should be treating all of it anyway.

The number that answers the buying question

Recon Analytics surveyed more than 150,000 US paid AI subscribers between July 2025 and January 2026 and asked a question nobody else thought to ask: what do people choose when they have a choice?

The same study found that of people given workplace access to each tool, 83 percent actually use ChatGPT regularly and 36 percent actually use Copilot. Roughly two in three Copilot seats never convert into real use.

Recon's own conclusion: "as worker choice expands, Copilot adoption collapses." For a corporate buyer that's a warning about wasted licences. For you, as somebody with unlimited choice, it's close to a direct answer.

Should you pay for it?

Now the money. This splits cleanly depending on your situation, so pick your path.

The argument nobody makes, and should

Copilot's genuine advantage is grounding in an organisation's data. The value of that grounding is a function of how large and how unknowable the pile is. At 5,000 people with fifteen years of SharePoint, being able to ask "what did we tell this customer in 2019" is worth $30 a month without argument.

At one person, you are the pile. You know where things are. You wrote them. The advantage that justifies the whole product mostly evaporates when the organisation is a single human being.

What's left for a solo operator is Copilot as an in-app writing and summarising assistant. That's a real product. It's just an ordinary one, competing directly against tools that measurably do that part better.

What each path actually costs you

Buy it if

Skip it if

What about building agents with Copilot Studio?

People ask about this constantly, so here's the honest answer. Copilot Studio requires a work or school account. Personal Microsoft accounts are rejected outright. So before you build anything, you buy a business plan, make yourself a Global Administrator, create security groups, link an Azure subscription, configure billing policies, and approve your own agent into the store.

Then there's the money. The default way to build agents since August 2026 is a newer engine, and Microsoft publishes no price list for it. Their own documentation says only that credit costs "vary." It charges you from the moment you start building, not from when you publish, and since 1 September 2026 it charges inside trial and development environments too. The 30-day free trial won't let you publish anything at all.

You cannot forecast what it will cost, and the meter runs while you're still learning. For a business whose only real budget is its own time, that's disqualifying. If you want to build agents as a solo operator, there are tools with published prices.

One genuinely interesting detail: Copilot Studio's Skills use the same SKILL.md format Anthropic developed, and you can upload existing skill files directly. If you've already built skills for Claude, they're portable.

Getting more out of it

If you're going to use Copilot, these five habits come out of the evidence rather than the marketing.

Run a two-tool stack, not a one-tool stack

This is the shape the research supports and it isn't a compromise. Use Copilot for the Microsoft-native surfaces: inbox triage, meeting recaps, editing inside a document you already have open, building a deck from a document you already wrote. Use Claude or ChatGPT for thinking, drafting from scratch, analysis, and anything where you need to know what produced the output.

People who try to make one tool do everything end up disappointed with whichever one they picked. People who split the work by where the data lives get value from both.

Budget for verification, and mean it

60 percent of outputs in the Australian government trial needed moderate to significant edits. 22 percent of UK users reported seeing hallucinations. Some participants found that checking Copilot's work took longer than writing from scratch would have.

Copilot is a first-draft machine. If you price your time as though it produces final drafts, you'll conclude it saved you nothing, and you'll be right, but the mistake was yours.

Never trust it on numbers without checking

Worth repeating because it's the single most evidenced failure. Any figure Copilot produces in Excel gets verified before it goes anywhere near a client. This is not general caution. It's a specific, measured finding from a government evaluation.

Check the length before you trust a summary

Long documents get silently truncated with no warning. Anything past roughly 20 to 30 pages, ask a question about the final section specifically. If it can't answer, it never read that part.

Know your actual limits

On Microsoft 365 Personal and Family you get 60 AI credits a month for in-app AI, 30 minutes a day of voice, and 10 minutes a day of vision. Premium doubles voice to 60 minutes and vision to 15. The free tier's limits are unpublished. Knowing these stops you building a workflow that breaks on the 20th of the month.

What to do with all this

Microsoft Copilot is a good product being sold badly. The marketing promises a general intelligence for your business, which it isn't, and buries the thing it genuinely does better than anything else, which is knowing what's in your email.

The honest position, on everything above, is narrow and specific. If you already work inside Microsoft 365, Copilot at consumer pricing is a reasonable buy and the AI is close to a free extra on software you were purchasing anyway. If your employer has already bought you a seat, use it, because someone paid and two thirds of those seats sit idle. If you're on Google Workspace, or you already pay for Claude or ChatGPT and you're wondering whether to add a second subscription, the evidence says don't.

And notice what the good version of this looks like. Nobody who gets real value from Copilot is having long conversations with it. They're getting a summary before opening a thread, catching up on a document a collaborator changed, and pulling a deck out of a report they already wrote. Small, boring, repeated. That's what two hours a week actually looks like when someone measures it properly.

Here's your next step, and it takes about twenty minutes. Open whichever Copilot you already have. Go into Outlook, find the longest unread thread you've got, and ask for a summary before reading a word of it. Then do the same thing tomorrow. If that one habit sticks, you'll have captured most of the value this product has been proven to deliver, and you'll know from your own experience whether the rest of it is worth paying for.

That's a better test than any review, including this one.