Skip to main content
Joshua Thompson

AI

Claude vs Copilot in Microsoft 365: how to test the value for your team

Claude and Copilot increasingly overlap. I would choose between them by the work a team finishes well, including the time and skill needed to check the results.

By Joshua Thompson

6 min read

A paper aeroplane folded from a spreadsheet, marked with orange circles and underlining, hovers above a stack of paper beside an upright pencil.

I would give Claude and Microsoft 365 Copilot the same awkward spreadsheet before giving either a bigger budget. A quick answer is useful only if somebody can use it. The time spent checking, correcting and getting it into the right place belongs in the comparison.

That sounds obvious until a demonstration makes a morning's work appear to disappear in seconds. The tempting purchase is the assistant that produces the most impressive first draft. The better purchase is the one that helps a team finish more worthwhile work to the required standard.

There is another complication: Claude and Copilot increasingly overlap. Choosing between the brands is less useful than understanding what each can do inside your particular business.

Claude and Copilot overlap

Microsoft's March 2026 announcement on Copilot and agents described bringing the technology behind Claude Cowork into Microsoft 365 Copilot. Buying a Microsoft product does not necessarily mean choosing against Anthropic's technology.

Claude can also connect to Microsoft 365, working with Outlook and OneDrive, plus SharePoint and Teams for work or school accounts. Access and the actions available depend on the account and administrator permissions.

That moves the buying decision towards the work itself. Can the assistant find the current proposal, distinguish it from last month's version, and produce something the next person can use? A persuasive answer drawn from an out-of-date file can be more trouble than no answer at all.

For a fair comparison, record the files, permissions and connected applications available to each tool. If one can search the relevant SharePoint folder and the other cannot, the result measures that difference as well as the quality of the answer. It may still be a commercially important advantage, but the team needs to understand what it is buying.

In Microsoft's own demonstration, Copilot Cowork assembles a customer brief and creates follow-up tasks in a customer-management system. It is a useful picture of the intended workflow. A workplace trial still has to establish how often that workflow succeeds and how much time it saves.

Video not loading? Watch on YouTube

A government Copilot trial: faster did not always mean better

A 2025 evaluation by the Department for Business and Trade gives a useful example of how results can vary by task. Its observed exercise involved 11 people and an earlier version of Copilot; it did not test Claude, and individual skills may explain some of the differences.

Copilot users produced report summaries in an average of 12 minutes 37 seconds, against 41 minutes 34 seconds for non-users, with higher accuracy scores. That is a gap of nearly 29 minutes on a task where the scored result was also better.

In PowerPoint, Copilot users finished almost eight minutes sooner, but averaged 1.5 out of 5 for accuracy against 5 for non-users. In Excel, Copilot users took about four and a half minutes longer and scored lower for accuracy. Emails were close to a tie, with a 13-second gap in completion time.

Faster output did not always score higher

DBT observed tasks, February to March 2025: Microsoft 365 Copilot users and non-users

  • Copilot users
  • Non-users

Average completion time

Minutes; lower is faster

Average accuracy

Score from 1 to 5; higher is better

Very small groups; 11 participants across four tasks. Not a Claude comparison.

Source: Department for Business and Trade, August 2025 evaluation, Table 6 and Annex C

Observed tasks involved 11 people, with two or three per group. Skills may have affected the results. This 2025 exercise tested an earlier Copilot, and did not test Claude.
View data
Average completion time (Minutes and seconds)
CategoryCopilot usersNon-users
Summarise reports12 min 37 sec41 min 34 sec
Write emails7 min 30 sec7 min 43 sec
Analyse Excel data25 min 1 sec20 min 33 sec
Create PowerPoint slides10 min 47 sec18 min 30 sec
Average accuracy (Score from 1 to 5; higher is better)
CategoryCopilot usersNon-users
Summarise reports4/52.5/5
Write emails4.3/54/5
Analyse Excel data1.5/52.7/5
Create PowerPoint slides1.5/55/5

The practical lesson is to buy for the jobs a team repeats. An assistant can be valuable for summarising documents and disappointing for a spreadsheet task. A company-wide answer to "Is Copilot worth it?" can hide both results.

I would take the report-summary finding seriously as a reason to try that use case. I would also want to know whether the team's own documents contain the awkward exceptions that make a summary difficult: contradictory figures, missing pages or a qualification buried near the end. A polished paragraph is a poor substitute for noticing that the source does not support a conclusion.

For slides, speed is only attractive if the material survives review.

The cost can move to the person checking

This is where an apparently cheap improvement can become expensive. A junior colleague might produce a draft more quickly while a manager spends longer checking its claims. The team has shifted the work, and possibly shifted it to somebody with less available time.

That does not make the assistant useless. It changes what needs to be measured. I would record who does the checking and whether it happens once or gets repeated further down the process. A report that saves its author half an hour but creates three rounds of clarification has not finished saving time yet.

Consider an illustrative task that normally takes 30 minutes. An assistant produces a draft in five, followed by 12 minutes of checking and eight of corrections. The finished job takes 25 minutes, saving five. Timing only the first draft would suggest a 25-minute saving, five times as much.

The distinction becomes important when a trial is scaled into a licence budget. Five minutes can be worthwhile on a frequent task. It needs to be counted honestly before being multiplied across a department.

Three jobs that would earn a place in my trial

I would start with familiar work whose quality the team can judge. The aim is to find out where either product earns its cost, including occasions when the existing process works better.

Summarise a report. Give both tools the same approved document and questions. A usable result needs supported claims, correct figures and the qualifications that change the conclusion.

Analyse a spreadsheet. Give both tools the same workbook and a question with a known answer. Check totals, formulas, units and the treatment of missing data.

Draft a client follow-up. Give both tools the same approved notes, offer and tone guidance. Names, commitments, prices and dates must be accurate, ready for the normal approval process.

Time the existing process too. Let people become familiar with both tools, try several examples and alternate which goes first. Where practical, have a colleague assess the outputs without knowing which product produced them.

Include an untidy but representative example, not just the clean file that makes a demonstration easy. Failed and abandoned attempts belong in the results. Otherwise the trial quietly awards the product a second chance that the team's usual process never gets.

Any small trial can be skewed by who takes part. I would use it to decide where to start, then check the results across more people and tasks before widening the rollout. Repeat the comparison when substantial product changes affect those jobs.

Speed is not the only reason to buy

A stopwatch can miss useful improvements. Suppose a colleague uses an assistant to write a clearer explanation in the same time, or to spot a mistake before a customer sees it. There may be value even when the minutes saved are zero.

The fair response is to score that benefit separately. Ask a reviewer whether the answer is clearer, more complete or more accurate. Keep those results alongside elapsed time instead of squeezing every benefit into a claim about productivity.

The opposite problem deserves the same attention. Creating extra presentations nobody needs is more output, but it has consumed time. A trial should ask whether the work was useful enough to have been worth doing in the first place.

A licence needs a job to justify it

For costs, use the actual licence quote and included features. Then ask what the recovered time enables: a shorter backlog, a quicker customer response, more completed work or less overtime. Those are different benefits and they need different evidence.

Multiplying minutes by a salary can estimate the value of capacity. It does not mean the same amount has disappeared from the wage bill. My guide to measuring the value of a Copilot rollout explains that distinction in more detail.

I would choose the tool that repeatedly produces reliable work with less total effort on the jobs the team does most. That could justify different tools for different roles, provided the benefit covers the added support and training. It could also justify buying only a few licences, or none yet.

The most convincing result would be a colleague finishing a familiar job, checking it properly and having time left for something useful. That is the demonstration I would pay attention to.

Comments

No comments yet - start the conversation.

Leave a Comment

Your email address will not be published. No links allowed - keep it kind.