3 of 5 caught without being asked.
For a lot of companies, Microsoft 365 Copilot isn’t one AI option among many. It’s the only one IT has approved. So it’s the one their people need to be good at.
We were preparing a hands-on AI training day for exactly that kind of team. Copilot is their only approved assistant, and their days run through Outlook, Teams, Word, Excel and PowerPoint. Before we taught anyone how to use it, we did what we’d tell any client to do with a new tool: we tested it on realistic work, and we hid mistakes in that work to see whether it would notice.
This is the full write-up: what Copilot caught, what it missed, what it now does on its own, how long it takes, and the eight habits we’d build in any team rolling it out. It’s how we review AI for knowledge work: hands-on, on the kind of tasks people actually do at their desks, with traps.
How we tested
- The setup. A brand-new Microsoft 365 business tenant with a Microsoft 365 Copilot licence. The app showed “M365 Copilot (Premium)”, version 2.20261001.61.0. Tested on 5 October 2026.
- The work. A fictional company we built for training, with the files a normal week produces: an event brief, a division update, rough FAQ notes, a 12-email customer thread, a meeting transcript, an eight-weekend sales workbook and a board report.
- The traps. Five mistakes, one per app, each checked to appear exactly once.
- The rules. Each task got the kind of prompt a busy employee would write, along the lines of “turn these notes into a one-page FAQ” or “recap this meeting with an action table”. Nobody asked Copilot to look for errors. We timed every answer.
- The limits. One tenant, one night, one run per task. This is a hands-on review, not a benchmark. Your admins may have switched features on or off, and Microsoft updates Copilot constantly.
Quick links
- 00The verdict at a glanceEvery finding in one table
- 01Five planted mistakesWhat it caught, what it missed
- 02No context, no answerAnd why grounding matters
- 03It acts nowDrafts, files and agents
- 04Speed and settingsTimings, models, new names
- 05The rollout playbookEight habits to build first
- ?FAQShort answers to common questions
00Summary
The verdict at a glance.
If you only have a minute, this is the whole review.
| What we tested | What happened | What it means for your team |
|---|---|---|
| Catching mistakes | Caught 3 of 5 planted mistakes without being asked. Missed 2. | The warnings are real, but each one is a single line in a long answer. Train people to read it. |
| No context | Refused to guess. Searched the web and found a different company with the same name. | No context means no answer, or the wrong one from the internet. |
| Your own files | Grounded answers were excellent, with every point cited. Fresh uploads took 10 to 15 minutes to show up in its search. | Point Copilot at the file. Don’t hope it finds it. |
| Taking action | Created two Outlook drafts and a designed PowerPoint deck from chat prompts. | It’s doing work now, not just answering. Check what it made. |
| Agents | New agents started with access to files, Outlook, Teams and the web. | Switch off the access an agent doesn’t need. |
| Speed | 20 to 50 seconds for most answers, 90 to 100 when it builds a file. | Send the main prompt first, then do something else. |
01Chapter · Accuracy
Five planted mistakes. Three caught.
Every app in the test got one deliberate mistake, the kind that slips through a real office every week. None of the prompts mentioned them. We wanted to know what Copilot does when nobody tells it to check.
| App | Result | The planted mistake | What Copilot did |
|---|---|---|---|
| Word | Caught | Rough FAQ notes citing a government returns directive that doesn’t exist | Wrote a clean one-page FAQ, left the fake directive out and said why |
| Excel | Caught | One row of sales entered twice | Found the duplicate, subtracted it in the formula and gave the corrected total and trend |
| Outlook | Caught | A reply in a 12-email thread promising customers a longer returns window than the written policy allows | Summarized the thread correctly and flagged the promise as unsupported |
| Teams | Missed | A meeting transcript naming a colleague who wasn’t actually in the meeting | Wrote a good recap and even flagged one unclear owner, but gave an action to the colleague who wasn’t there |
| PowerPoint | Missed | A board report labelling a last-12-months revenue figure as a calendar-year number | Built a designed six-slide deck and copied the wrong label onto a headline figure |
What the three catches have in common
Each one was a claim that didn’t hold up against a source. The directive had no source at all. The returns promise had nothing in the thread to back it up. The duplicate row broke the arithmetic. When Copilot has something to check a statement against, it checks, even unasked.
What the two misses have in common
Both were mistakes in the frame around the facts, not in the facts themselves: who was actually in the room, and which period a number covers. The names and numbers were real. They were just attached to the wrong thing. A colleague would catch that from context. An assistant reading the text didn’t.
Five traps is a small sample, so treat this as a pattern to watch, not a law. But it gives managers a useful rule of thumb: let Copilot check the content, and keep the frame check for a person.
The most skipped line on the screen
Here’s the part that matters most for training. In all three catches, the warning was a single line inside a long, confident, well-formatted answer. The FAQ looked finished. The sales summary looked finished. If you skim, and everyone skims, you take the finished-looking output and miss the one sentence telling you something was wrong.
Copilot caught three of these for you. Did you read the line where it told you?
That’s why, in our training, a catch only counts when a person can do three things: name the problem, show where the source proves it, and state the correct value. Copilot’s flag is where the check starts, not where it ends. In the Outlook task, for example, Copilot said the promise was unsupported but didn’t say what the policy actually allows. The customer reply still needs that number.
Would your team have caught the two Copilot missed?
We train teams on the AI tools their company already approves, using realistic work with mistakes hidden in it, so checking becomes a habit instead of a lecture.
02Chapter · Context
No context, no answer. Or the wrong one.
The fear most people bring to AI at work is that it makes things up. So we tried to make Copilot make something up. We asked a question that only makes sense with company context, roughly “are we ready for the weekend event, and how does it compare with last year?”, without giving it any files.
It didn’t invent an answer. We ran it twice, with Copilot’s Work IQ switch (whether it can use your mail, files and meetings) off and then on:
Work IQ off
- Didn’t guess.
- Searched the web and cited a real, unrelated company that happens to share our fictional company’s name.
- Asked us for details.
- Took about 20 seconds.
Work IQ on
- Searched mail, files, Teams and meetings in eight visible steps.
- Found nothing it could use. It didn’t look inside the Notebook where we’d put the files.
- Pulled in an unrelated place with the same name from the web.
- Concluded that readiness wasn’t evidenced. Took about 50 seconds.
That’s the right behaviour, and it’s better than inventing. But notice the new risk. An invented answer often looks wrong. A real company with the same name looks right. The danger with a well-behaved assistant isn’t fiction. It’s the wrong facts, confidently sourced.
No context means no answer, or the wrong answer from the internet.
Give it the files and it’s excellent
Then we did it properly. We created a Copilot Notebook, added the event brief and the division update, and asked the same question. The answer came back in about 25 seconds: five bullets, each citing the file it came from, every fact correct. That’s the version of Copilot worth paying for. Context is what turns it from a search engine into a colleague.
Point it at the file. Don’t hope it finds it.
One practical catch: files we had just uploaded didn’t show up in Copilot’s work search for roughly 10 to 15 minutes. If someone in a meeting says “Copilot, summarize the doc I just shared”, it may not find it yet. Ground it explicitly, every time:
- Type “/” in the prompt box and pick the file, person, meeting or email you mean.
- Attach it with + → Upload, or attach a cloud file.
- Work inside a Notebook when a task needs several files. Our uploads took about 40 seconds before chat unlocked.
One more thing people don’t expect: files you upload into Copilot Chat are copied into your OneDrive, in a folder called “Microsoft Copilot Chat Files”. They become part of your work data. Upload accordingly.
03Chapter · Action
Copilot doesn’t just answer anymore. It acts.
This was the biggest surprise of the test, and the one with the most consequences for how teams should use it.
- It wrote emails into Outlook. From one chat prompt about the customer thread, Copilot created two reply drafts directly in our Outlook. It didn’t send them, but they were sitting there, ready to go.
- It made real files. Copilot Chat’s + → Add capabilities menu now includes options like Design a presentation, Build a worksheet and Write a document. One prompt gave us a fully designed PowerPoint deck with headline figures and charts, saved straight to OneDrive. A couple of follow-ups later, we had three versions.
- Its agents start with wide access. In Agent Builder, a new agent’s work content (cloud files, Outlook, Teams and web search) was all switched on by default. We didn’t create one, but the defaults tell you where the risk sits.
Microsoft’s own documentation adds that Copilot in Outlook can now move or delete mail and create rules when you ask it to, with a confirmation step. Put together, that’s a shift from an assistant that suggests to one that does.
Remember the deck
The PowerPoint deck Copilot designed is the same one that carried the mislabelled revenue figure from Chapter 1, now on a polished headline tile. That’s the real risk of an assistant that acts. A mistake doesn’t stay in a draft paragraph you’d reread. It arrives as a finished-looking file, email or rule, and finished-looking things get forwarded.
Stop
- Approving confirmation cards without reading them
- Forwarding generated decks without checking every number
- Creating agents and leaving every default switched on
Start
- Checking your Drafts folder after any email task
- Treating every generated file as a first draft
- Switching off the access an agent doesn’t need
Paying for Copilot and three other AI tools?
The Marketing Stack Cost Optimizer totals your tool spend, spots overlapping tools and finds consolidation savings.
04Chapter · Speed & settings
Speed, limits and the settings that matter.
Copilot isn’t instant, and planning around that is half of using it well in a busy day. These are the times we measured:
| Task | Time in our test |
|---|---|
| A normal chat answer (summaries, recaps, rewrites) | About 20 to 50 seconds |
| Answering from a Notebook | About 25 seconds |
| Building something (an Excel analysis, a designed deck) | About 90 to 100 seconds |
| Uploading files into a Notebook | About 40 seconds before you can chat |
| Researcher, the deep-research agent | Under 5 minutes for simple questions and 10 to 45 for complex ones, according to Microsoft. Capped at 25 runs per person per month. |
The practical rule: send the main prompt first, then do something else. In a meeting, a 100-second wait feels like forever. And never let a deadline depend on Researcher finishing.
The model picker
Copilot now has a model picker. In our tenant it showed Auto, Quick response and Think deeper, plus models from Anthropic (Claude) and OpenAI (GPT). Quick response is handy for fast follow-ups. For everything else, stay on Auto unless you know why you’re switching. Microsoft runs some models and features on Copilot Credits, which stay off until your admin sets a spending policy.
Cowork: off until your admin says otherwise
Copilot Cowork showed only “Request access” in our tenant. Microsoft says usage-based features like Cowork stay off until an admin creates a spending policy. If yours “isn’t working”, that’s the likely reason, not a bug.
New names, old guides
If you’re following a Copilot guide from earlier in the year, some labels have moved:
- Prompt Gallery is now Prompt Lab. You can bookmark your own prompts and share a link to them.
- What used to be called agent mode now appears as Edit with Copilot, with an Allow editing switch in Word.
- “/” opens your context (files, people, meetings, emails, chats, channels and sites). “@” opens agents such as Researcher and Analyst.
Where would AI actually save your team time?
The Marketing Automation ROI Finder estimates the hours and money your team loses to repetitive busywork every year.
05Chapter · Rollout
How we’d roll Copilot out to a team.
Licences don’t change how people work. Habits do. If we were handing Copilot to a team tomorrow, these are the eight habits we’d build first, all taken straight from what the test showed.
- Ground every task. Type “/” and pick the file, attach it, or use a Notebook. Don’t rely on search to find what you just shared.
- Read the warning line. Copilot caught three of five mistakes on its own, and each time the warning was one line. Make “what did Copilot flag?” a standard review question.
- Check the frame yourself. Who was in the room, which period a number covers, which company a source is about. Both misses, and the wrong-company answer, lived there.
- Prove it before you fix it. When Copilot flags a problem, find the source and state the correct value. A flag isn’t an answer.
- Check what it made. Your Drafts folder, your OneDrive and every confirmation card. It acts now.
- Give agents the least access they need. Switch off Outlook, Teams and web search for an agent that only needs one document.
- Mind your uploads. They’re copied into OneDrive and become part of your searchable work data.
- Send the big prompt first. Plan for 20 to 100 seconds per answer, and keep Researcher off the critical path.
The best way to build these habits isn’t a slide deck about features. It’s realistic work with mistakes hidden in it, done against the clock, in teams. That’s how we design our AI training days now: on the tools your company has already approved, so people practise on the exact assistant they’ll open on Monday morning.
Related: how we trained 27 people to run their businesses on AI in a single day · AI Ops for growth teams: a practical playbook
So, is Microsoft 365 Copilot worth rolling out?
Yes, for teams that live in Outlook, Teams and Office. On one condition: you train people to ground it and to check it.
Copilot is now good enough to change how a team works, and good enough that people will be tempted to stop checking it. Building that checking habit is the real job. The licence is the easy part.
Your team already has the tool.
Do they know how to check it?
We run hands-on AI training days on the tools your company already approves, Copilot included: realistic work, planted mistakes, scored as a team. Book a short call and we’ll scope one for your team.
Build smarter with Kando.
?FAQ
Frequently asked questions.
Is Microsoft 365 Copilot accurate?
In our test it caught three of five planted mistakes without being asked: a fake regulation, a duplicated spreadsheet row and an unsupported promise in an email thread. It missed two: an action given to someone who wasn’t in the meeting, and a revenue figure labelled with the wrong period. Use it as a second pair of eyes, not the final check.
Does Microsoft 365 Copilot make things up?
With no context, it didn’t invent an answer. It searched the web and asked for details. The risk we saw was different: it pulled in a real but unrelated company with the same name. Grounding it in your own files, with “/”, an attachment or a Notebook, avoids that.
Can Copilot send emails or create files on its own?
In our test it created two Outlook reply drafts and saved designed PowerPoint files to OneDrive from chat prompts. It didn’t send anything. Microsoft shows confirmation steps for actions like moving mail or creating rules, so read them before you approve.
Why can’t Copilot find a file I just uploaded?
Freshly uploaded files took roughly 10 to 15 minutes to appear in Copilot’s work search in our test. Point Copilot at the file directly instead: type “/” and choose it, attach it, or work inside a Notebook.
How long does Microsoft 365 Copilot take to answer?
Most chat answers took 20 to 50 seconds in our test, and 90 to 100 seconds when Copilot built a file such as a deck or an Excel analysis. Microsoft says its Researcher agent takes under 5 minutes for simple questions and 10 to 45 minutes for complex ones, with 25 runs per person per month.
How should a company train employees to use Copilot?
Hands-on, on realistic work in the tools people already use, with mistakes hidden in the material. Build habits rather than feature knowledge: ground every task, read the warning lines, check who, when and which period yourself, review anything Copilot creates, and give agents only the access they need.