I spent six weeks pointing Microsoft Copilot at real work instead of demo tasks: 43 documents summarised, around 120 email threads triaged, a dozen spreadsheet formulas built, and a long list of questions I already knew the answers to. The goal was simple. Find out how often Copilot is right, and whether it is obvious when it is wrong.
Accuracy turned out to be less about the model and more about the setup. The same assistant that quoted a report correctly one minute produced a confident, entirely invented statistic the next, and its writing style barely changed between the two.
What people actually mean when they say Copilot is accurate
Two different things get scored as one. Factual accuracy asks whether a claim is true in the world. Fidelity asks whether the answer reflects the document, thread or spreadsheet you handed over. Copilot is noticeably stronger on fidelity than on facts, and most of the complaints I hear come from people who expected the first and got the second.
There is a third layer too: completeness. An answer can be technically correct and still useless because it dropped the one constraint that mattered, like a budget cap or a hard deadline.
Where Copilot’s accuracy held up
Summarising material you supply
This is the strongest case by a distance. Across 43 summaries of meeting notes, contracts and long email chains, I found three that misrepresented a key point. Two of those were documents with contradictory sections, where a careful human would have flagged the conflict rather than quietly resolving it. Roughly 93% usable is a fair characterisation, and the failures were visible if you read the source alongside the output.
Rewriting and tone shifts
Ask it to turn bullets into a paragraph, soften a message to a client, or cut a 900-word draft to 400, and it holds the meaning well. In about 30 rewrites it changed the substance once, converting “we cannot commit to a date” into “we expect to confirm a date soon”. That is the kind of drift worth watching, because it is a small edit with real consequences.
Formula building and structured tasks
Eleven of twelve spreadsheet formulas worked first time, including a nested XLOOKUP and a SUMIFS with three conditions. The failure involved a date range where it assumed a US date format. Boring, predictable, easy to spot with a ten-second sanity check.
Where accuracy falls apart
Numbers, dates and citations
Ask for a statistic and you enter the danger zone. Copilot gave me a plausible-looking market size figure attributed to a report that does not contain it. The structure was right, the source name was real, the number was fiction. Anything that carries a figure needs checking at the origin, every single time.
Anything published recently
Free Copilot with web access does better here, but it still leans heavily on whatever the search layer returns. Asked about a funding round announced ten days earlier, it named the right company and the wrong amount, lifted from an older article. In a chat window, that reads like a clean answer.
The confidence gap
Copilot does not hedge. Wrong answers arrive with the same tidy structure and pleasant phrasing as right ones. No wobble in tone, no warning sign. That is the biggest accuracy problem in practice, bigger than any individual hallucination, because it means you cannot triage by feel. You have to verify by default.
Grounding explains most of the variance
Copilot is an umbrella label on top of several very different configurations, and the reliability gap between them is larger than the gap between Copilot and its rivals.
- No grounding: the model answers from training data. Fastest, and the least reliable for anything specific.
- Web grounding: it searches and cites. Better for current events, still exposed to weak or outdated sources.
- Work grounding in Microsoft 365 Copilot: it reads your files, mail and chats through Microsoft Graph. The strongest fidelity of the three, but only as good as your permissions and file hygiene.
If your team is weighing options, a direct comparison of Copilot against ChatGPT, Gemini and Claude is worth the read before you commit seats.
Free Copilot versus the paid Microsoft 365 version
The free tier is a useful general assistant with a web connection. The $30-per-user-per-month Microsoft 365 version is a different product: it can ground answers in your own tenant, point to the specific file it drew from, and follow up on a thread you were copied into last Tuesday.
That grounding is what buys accuracy. It is also why the per-seat price deserves scrutiny rather than a blanket yes, which is the angle taken in this value-for-money review of business AI tools. If you mostly write and research, the free version plus good habits goes a long way.
Five habits that measurably cut the error rate
None of these require a settings change. They are prompt discipline, and in my testing they moved usable output from roughly 70% to around 90% on knowledge-heavy tasks.
- Paste the source material rather than describing it. “Here is the contract, find the termination clause” beats “summarise our contract” every time.
- Ask for verbatim quotes with the paragraph or section they came from, so you can trace each claim back.
- Split compound requests. One task per prompt, then combine the pieces yourself.
- Tell it to say when it does not know, and to name which parts it is unsure about.
- Check every number, date and proper noun at the original source. No exceptions, however clean the answer looks.
Once a prompt pattern works, save it. The step-by-step guide to automating a first Microsoft AI workflow shows how to chain those reliable prompts into something that runs without you retyping them.
Where the accuracy is trending
Citation quality has improved noticeably. Answers that used to hand you a bare claim now often carry a source link you can open, and the Researcher-style agents are built around evidence gathering rather than confident recall. The direction of travel is toward grounding almost everything, which is the right fix for the failure mode I hit most.
Scepticism is still warranted. A broader look at what works and what does not in Copilot tracks the same pattern: strong on your own data, shaky on the open web, improving in public but not there yet.
A benchmark you can run in an afternoon
Rather than trust my numbers, build your own. Pick twenty tasks that resemble your actual week. Include three that hinge on a specific figure, two that require reading a long document you know well, and five where you can check the answer against a source in under a minute.
Run them, then count failures by category rather than in total. In my sample, factual errors clustered in statistics and recent events, while fidelity errors clustered in documents with internal contradictions. That split tells you exactly where to place your review effort, and it tells you whether the paid tier’s grounding is worth the seat cost for your team. A tool that is 90% accurate on summaries and 60% accurate on figures is not a 75% tool. It is two different tools wearing the same name, and the sooner you map which one you are relying on, the less time you will spend correcting work you assumed was finished.

