What a Copilot seat costs, how many minutes it has to earn back, and how much of the published evidence you can put in front of a finance partner.
Microsoft Copilot ROI is easier to reason about as a break-even question than as a single percentage. At a $30 list price a seat needs somewhere between about one and three net minutes a working day, depending on the role. What decides the outcome is persistence: how many licence holders are still using it in month three. The evidence is uneven. The largest ROI figures come from studies Microsoft commissioned and are projections. A randomized experiment measured a smaller, email-specific saving, and a UK government evaluation found no robust evidence that saved time became higher productivity. Label each number by who paid for it and how time was measured, then judge your own seats with a Use, Persistence and Impact scorecard.
Say you’re the IT lead who rolled out 300 Microsoft 365 Copilot licences in the spring. Six months later a finance partner forwards you one line from the budget review: “Can you show me what we’re getting for the Copilot spend?” You have a licence count, a training attendance sheet and a handful of colleagues who say they love it. None of those is an answer, and you both know it.
The useful move is to stop treating that email as a request for one ROI percentage and treat it as two smaller questions. How many minutes does a single seat have to give back to cover its own price? And how many of your seats are still giving those minutes back in month three? The first has a smaller answer than most people expect. The second is where the money is won or lost, and a licence count can’t tell you.
That’s the argument of this article. At today’s prices a Copilot seat has to recover a modest amount of net time to break even, so Microsoft Copilot ROI depends less on the price than on how many people keep using it and whether the time they recover turns into better work. Our guide to measuring your AI ROI covers the general method, and the Copilot adoption benchmarks guide covers usage metrics. This piece stays on the money: what a seat costs, what it has to earn, and how much of the evidence you can show finance.
Every price and study below was read on its publisher’s own page in September 2026, and each number is marked as a vendor-commissioned projection, a measured result or a self-report.
Start with the price you’d actually be quoted. Microsoft lists Microsoft 365 Copilot at $30 per user per month on an annual commitment, and Microsoft 365 Copilot Business at $21, with a promotional $18 for the first year on offers taken between July 1 and December 31, 2026.[1][2] Both need a qualifying Microsoft 365 plan underneath. These are US dollar list prices, and your agreement and region can move them, so use the table as a starting point for your own quote.
| Option | List price (USD, per user per month) | Catches to price in |
|---|---|---|
| Microsoft 365 Copilot Chat | No additional cost with an eligible Microsoft 365 plan | Still depends on the plan you already pay for; the comparison to make is what a paid seat adds on top of it |
| Microsoft 365 Copilot Business | $21 on annual billing; $18 promotional; $25.20 on monthly billing | Promotion runs July 1 to December 31, 2026, needs an annual commitment and applies to the first year only. Needs a qualifying Microsoft 365 plan |
| Microsoft 365 Copilot | $30 on annual billing | Auto-renews yearly. Needs a qualifying Microsoft 365 plan |
List prices from Microsoft’s Copilot pricing pages, read in September 2026. Confirm against your own agreement before using them in a business case.
The licence line is only the first line of the cost side. When Forrester built its Microsoft-commissioned model of a composite organization with 25,000 employees, it also budgeted for implementation and management, and for training time: 10 hours of formal training for each new user, then 6 hours a year.[3] That’s a vendor-commissioned model, so read the amounts as assumptions rather than results. The categories are still the right ones, and leaving them out makes payback look faster than it will be.
Different roles feel different parts of that cost:
A business case gets sturdier once every cost has a named owner, because otherwise the only cost anyone has priced is the one on the invoice.
Turn the price into minutes and the argument changes shape. A $30 seat has to buy back thirty dollars of someone’s time each month, and what that means depends on what an hour of that person’s time costs your organization once salary, benefits and overhead are included. The list below uses three example loaded hourly rates, so swap in the figures your finance team already uses.
Those have to be net minutes, meaning what’s left after checking, correcting and re-prompting, and they have to be minutes that went into something finished rather than dissolving into a slightly calmer afternoon. Our guide to tracking your time savings from AI shows how to time one task. On paper, then, break-even looks within reach, and rollouts can still disappoint.
Price the seat in minutes first, then ask who will still be recovering them in month three.
The gap between paper and practice is persistence. Every licence you pay for owes those break-even minutes, including the licences nobody opens after the launch week. That’s the arithmetic the next section puts numbers on.
Here’s the arithmetic with every assumption visible. It’s illustrative, not client data: a 200-seat rollout at the $30 list price, a loaded cost of $45 an hour, six hours of training time per user, $12,000 a year for admin, champions and reporting, and 3 net hours a month recovered by each person still using Copilot in month three. Nobody else is assumed to recover anything.
| Line | 30% still using | 55% still using | 80% still using |
|---|---|---|---|
| Licences (200 x $30 x 12) | $72,000 | $72,000 | $72,000 |
| Training time (200 x 6 hours x $45) | $54,000 | $54,000 | $54,000 |
| Admin, champions, reporting (assumed) | $12,000 | $12,000 | $12,000 |
| Year-one cost | $138,000 | $138,000 | $138,000 |
| Persistent users | 60 | 110 | 160 |
| Time value (users x 3 hours x 12 x $45) | $97,200 | $178,200 | $259,200 |
| Year-one net | -$40,800 | +$40,200 | +$121,200 |
| Year-one ROI | -30% | +29% | +88% |
Illustrative arithmetic with the assumptions stated above, not client data. Time value is capacity recovered, not cash saved.
Read across the middle row. The price, the training and the hours per person are identical in all three columns, and the outcome swings from a loss to a gain of $121,200 purely on how many seats persist. In year one this example needs about 43 of every 100 seats still in use to break even. Once the one-off training cost is behind you, the year-two hurdle falls to about 26%, because the recurring cost is $84,000 rather than $138,000. And the net time matters as much as the seat count: halve it to 1.5 hours a month and the year-one hurdle roughly doubles to about 85% of seats.
Now a smaller scene. Say one of those seats belongs to a finance analyst who uses Copilot to draft the monthly variance commentary. The first draft takes four minutes instead of twenty-five. Checking it against the ledger and rewriting two sentences takes twelve. The net saving is nine minutes rather than twenty-one, and it only exists in the weeks the variance pack is due. That’s a real saving, and the kind of number that shrinks once someone measures it properly.
Judge Copilot by what each active seat costs, not what each licence costs.
At 55% persistence the $30 licence effectively costs about $55 per active seat, and at 30% it’s $100. Persistence is the middle step of the Use, Persistence and Impact lens we use for adoption work, and the percentages in the worksheet are exactly that figure. A licence count doesn’t move when people stop using the tool, so it can’t carry a business case on its own.
Is 3 net hours a month a realistic figure? The most useful measured evidence I found is a randomized field experiment: some knowledge workers were offered Microsoft 365 Copilot and comparable colleagues weren’t, across 66 large firms, between September 2023 and October 2024.[5] The authors come from Microsoft Research and Harvard Business School, so it’s Microsoft-affiliated work, but a comparison group is a very different design from a projection.
The comparison group lets the authors separate what Copilot changed from what would have happened anyway, which is the question a business case has to answer.
The clearest effect was on email. Across everyone offered access, email time fell by about 1.4 hours a week, and by about 2 hours a week among people who actually used Copilot. Time in Word and in Teams meetings didn’t change in any way the authors could tell from zero, and they found no shift in the mix of tasks people did.[5]
Roughly six hours a month of email time is well above the break-even minutes in the previous section, so an email-heavy role such as an account lead, sales manager or project coordinator is a sensible place for early seats. The flat results for documents, meetings and task mix are the better warning for a business case: saved email time doesn’t become a shipped deliverable unless a person, usually the team manager, decides where the hours go.
The UK Department for Business and Trade ran a three-month evaluation of 1,000 Microsoft 365 Copilot licences from October to December 2024 and published it in August 2025.[6] Participants’ own diary entries, adjusted for unused outputs, showed time saved on drafting, summarising, transcribing meetings and searching, and added time on scheduling and image generation.
Two limits apply before you read anything into it. It’s a snapshot of late 2024, and its diary study had a low response rate that the department itself flags as a possible source of bias. What it does give you is a test list of tasks to check before you assume they save time.
Within those limits, the department’s conclusion is the sentence a finance partner would underline: “We did not find robust evidence to suggest that time savings are leading to improved productivity.”[6] In the observed tasks, Copilot users completed Excel analysis more slowly and less accurately than non-users, and produced slides faster but at lower quality. A value-for-money assessment was still under way and wasn’t part of the publication, so the evaluation can’t say whether the licences paid for themselves. The practical rule is to compare Copilot’s output with what a person produces without it, especially in Excel, where a wrong answer looks as tidy as a right one.
Where Copilot takes on a real piece of work, spell out who owns what. This is the split I’d write down for three common tasks:
| What Copilot does | What the human still owns | How it gets checked |
|---|---|---|
| Drafts the monthly variance commentary from the figures | Which variances matter and what caused them | Analyst checks every number against the ledger before the pack leaves; finance manager samples one pack a month |
| Summarises a long email thread or meeting for a manager | The decision and any follow-up commitment | Manager opens the source thread for any item that involves money, a customer or a deadline |
| Builds a first-pass Excel analysis | Whether the method and the totals are right | A second person recomputes one total by hand; anything reaching a client or board gets a full review |
An example split for three tasks, written for a finance team. Adjust the reviewer and the sampling rate to your own risk.
You’ve now met four kinds of number: a vendor-commissioned projection, a randomized experiment, a government evaluation built on diaries, and a national trial built on perceptions. Headlines call all of them “Copilot ROI” or “time saved”, and they don’t answer the same question. Sorting them before comparing them saves a lot of confusion.
Before you quote a Copilot ROI figure, say who paid for it and whether the time was measured or remembered.
| Source | Who paid or ran it | How time was captured | What it can support |
|---|---|---|---|
| Forrester TEI, March 2025. Projected 116% three-year ROI and 10-month payback for a 25,000-employee composite [3] | Commissioned by Microsoft | Projected from 16 interviews and a 367-person survey, with risk adjustments | A structure of cost and benefit lines to copy. Forrester itself says it makes no assumptions about the ROI other organizations will receive |
| Forrester SMB TEI, October 2024. Projected three-year ROI of 132% to 353% [4] | Commissioned by Microsoft | Modelled low, medium and high scenarios for a composite small business (up to 300 employees) | How sensitive the answer is to assumptions. The gap between 132% and 353% is the lesson |
| Dillon et al., 2025. Email time down about 1.4 hours a week [5] | Microsoft Research and Harvard Business School authors | Randomized experiment with a comparison group | A measured, email-specific effect. Not evidence about documents, meetings or better work |
| UK Department for Business and Trade, August 2025. No robust evidence savings improved productivity [6] | UK government department | Self-reported diaries (32% response) plus observed tasks | A list of tasks to test, especially Excel. Not a verdict on today’s product |
| Australian Government trial. Perceived savings of around an hour a day on some tasks; one in three used it daily [7] | Australian Government | Participant perceptions; the evaluation says editing and verification cancelled part of the saving | The direction of the saving and the size of the verification tax. Not measured time |
Each row was read on its publisher’s own page in September 2026. Full sources are listed at the end of the article.
Sorted this way, a pattern shows. The largest figures come from projections Microsoft commissioned, and the smaller, more cautious ones come from designs with a comparison group or a public-sector evaluator. That doesn’t make the projections wrong, because a projection shows what happens if its assumptions hold. The useful thing to take from them is the structure, the cost lines and benefit categories, then to fill in your own numbers.
If you’re the CFO or finance partner, ask for the label first and the figure second. If you’re the IT admin or team manager presenting, put the label in the same sentence as the number. “Microsoft-commissioned projection” costs you nothing, and it makes the rest of your case believable when someone checks.
Studies can set your expectations, but only your own seats can settle the question. Instead of handing finance one ROI percentage, give them three numbers, one for each step of the Use, Persistence and Impact lens, with a threshold for each. The bands below are our own heuristic, not a published benchmark, and the persistence band is tied to the break-even share you calculated in the worksheet.
| Step | Number to pull | Green | Amber | Red |
|---|---|---|---|---|
| Use (IT admin) | Share of licence holders who took an intentional action in Copilot in the last 30 days | Above 70% | 40% to 70% | Below 40% |
| Persistence (team manager) | Share of all licence holders who were active in month one and are still active in month three | Above your year-one break-even share (43% in the example) | Within 10 points below it | More than 10 points below it |
| Impact (finance partner) | Net minutes saved on one named task, timed on 10 real outputs with and without Copilot, after checking time | At or above the break-even minutes for that role | Between half and all of the break-even minutes | Under half, or no before number exists |
A Future Factors heuristic for renewal decisions, not an externally published benchmark. Where to find usage data is covered in our Copilot adoption benchmarks guide.
Read it as a decision aid:
Training builds the capability, a named workflow gives it somewhere to land, and a manager who asks about it in month two turns it into a habit. The persistence column is where you find out whether that last step happened. It’s the part of ROI that depends on how people change the way they work, which we cover in what makes an AI-powered professional different.
You can run all of this yourselves. Expert-led, role-specific training tends to shorten the curve and to build shared capability across a team instead of a few enthusiasts, and it’s telling that Forrester’s model budgets hours for it. Our guides to Microsoft Copilot training and to the five questions executives ask about Copilot cover what that looks like in practice.
Here’s a first step for this week: pull the month-three active list for one team, look up the break-even minutes for that team’s loaded rate in the list above, and ask five people to time one named task with and without Copilot. That gives finance a number with a name and a label attached, which a licence count can’t.
Use the price on your own quote, on the billing term you’d really sign. As of September 2026 Microsoft lists Microsoft 365 Copilot at $30 per user per month on annual billing and Copilot Business at $21, with an $18 first-year promotion on offers taken between July 1 and December 31, 2026. Model the promotional price for year one only, then list price, so the case doesn’t assume a discount that has expired.
Use them for structure, not as a forecast. The Forrester studies were commissioned by Microsoft and are projections built on composite organizations. Take the cost lines and benefit categories, replace every assumption with your own, and say in the business case which parts came from a vendor-commissioned study.
It depends on persistence, and the licence and training costs land before the value does. In the worked example, if users stay active from month one, the up-front training cost is recovered in about 7 months at 55% persistence and about 4 months at 80%, while at 30% it takes close to four years. That ignores any ramp-up, so real timelines run longer. Forrester’s Microsoft-commissioned model projects a 10-month payback for its composite, a figure to compare your own month-by-month calculation against, not a promise.
The licence line is zero, but training time, admin effort and checking time still cost something. The ROI question becomes incremental: what does a paid seat let a specific person do that Copilot Chat doesn’t, and does that difference cover $21 or $30 a month? Run the break-even list for the people you’d upgrade, and upgrade only the ones whose work depends on it.
No. Time recovered is capacity, and it becomes cash only when it changes something you’d otherwise pay for, such as overtime, a contractor, a hire you no longer need or a deadline you no longer miss. More often it turns into capacity for other work, which is worth something but isn’t a saving on the budget line.
Prices come from Microsoft’s own Copilot pricing pages, read in September 2026 (US dollars, list). The Forrester Total Economic Impact study of Microsoft 365 Copilot (March 2025) was read in full on Forrester’s own site; the October 2024 small-business study is cited only through Microsoft’s own blog post about it, because the study itself was not read here. Both were commissioned by Microsoft and are projections. The randomized experiment was read on NBER and arXiv (version 4, November 2025), and the UK and Australian government evaluations were read on their publishers’ pages. The worked example is illustrative arithmetic with stated assumptions, not client data. Blog posts that quote these studies were deliberately not used as sources.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.