How do you measure Copilot's ROI without fooling yourself?

Measure one process at a time, with a before and an after: how long the work took, how long it takes now, how much volume the same people handle, and how much correction is needed. Logins and the time users say they save are not ROI. Set the net result against what the licences really cost.

Why isn’t the time users say they save enough to prove ROI?

Because it is an estimate from the person trying the tool, and it does not say what happened to that time. Researchers at the Federal Reserve Bank of St. Louis, with Vanderbilt and Harvard, asked American workers how much generative AI saves them: in the August 2025 survey the reported saving equals 1.6% of all hours worked, from which the authors derive a productivity gain of up to 1.3% (St. Louis Fed, 13 November 2025). These are self-reported answers about generative AI in general, not about Copilot.

In the pilot at the UK Department for Business and Trade (1,000 licences, October to December 2024) the savings are small, and the August 2025 evaluation finds no evidence of higher productivity, though it notes that this was not the main objective (DBT report). Once the evaluation removes outputs that went unused and tasks that only existed because the tool was there, the saving on presentations falls from 0.8 hours to zero and meeting scheduling turns into a loss of 0.6 hours. A neighbouring problem is rework: a Workday survey of 3,200 employees (November 2025, from a software vendor, and not only about Copilot) says it absorbs almost 40% of the time saved with AI (Workday, 14 January 2026).

Does the Copilot Dashboard measure ROI?

No: it measures usage and estimates value with a model. In the Microsoft Copilot Dashboard, “assisted hours” are users’ actions multiplied by coefficients drawn from Microsoft’s research, which Microsoft itself calls a “general estimate”. The money value is that number of hours times a default rate of $72 an hour, which you can change (Microsoft Learn, updated 29 September 2026). It is good for understanding how many employees really use Copilot.

How much weight should you give the Copilot ROI studies in circulation?

Read them as models, and the most quoted ones were paid for by Microsoft. Forrester’s Total Economic Impact study, commissioned by Microsoft, describes a composite organisation of 25,000 employees: $36.8 million in benefits over three years against $17.1 million in costs, which is an ROI of 116% (Forrester, March 2025). A little over half the benefits ($18.8 million) rest on two assumptions: 9 hours saved per user per month, half of them reused for other productive work. The study on small and medium-sized businesses, also commissioned, gives 132% to 353%, but it is a projection based on companies still in pilot or beta (Forrester, October 2024).

Microsoft itself has published a field measurement with usage data: a randomised experiment with more than 6,000 workers at 56 companies (Microsoft, April 2025). People with a licence spent on average 12 fewer minutes a week reading email, but only about 40% of the workers with access used Copilot regularly. The sum has to be done on every licence you pay for, not just on the people who use it.

Which metrics should you track in practice?

Three, all on a single process. First, cycle time, from the start to the end of the work, taken from system timestamps and not from an estimate. Second, the volume the same team handles in the same period: tickets, cases, bids. Third, rework, meaning the share of output corrected by hand, measured on a sample. The cost that used to leave the company, in overtime and outside suppliers, counts as well.

In our experience these numbers are only useful if the before is measured prior to assigning the licences, as in a well-run pilot, and if one person answers for them.

How many hours a month does it take to pay back a Copilot licence?

Less than an hour a month per user, if an hour of work costs the company more than $30. The list price for large organisations is $30 per user per month on an annual subscription (€26.00 in Italy, excluding VAT, per Microsoft Italy’s price page, accessed 3 October 2026): divide it by the full hourly cost, meaning pay plus employer on-costs. The hourly costs and the calculation are ours, so this is an example and not market data.

Hourly cost (assumption)	Hours a month to break even on the licence	Minutes a week	Hours a month if the licence is a third of costs
$30	1.00	14	3.0
$40	0.75	10	2.3
$50	0.60	8	1.8
$70	0.43	6	1.3
$90	0.33	4.6	1.0

In the last column the hours are tripled because in Forrester’s composite case the licences are worth about a third of the three-year costs, the rest being implementation, management and training; your share will be different. The Australian Treasury estimated that a mid-level official needs about 13 minutes a week to cover the licence, without having measured the time actually saved (Treasury, February 2025).

An example we made up. A bid office of 8 people prepares 20 bids a month at 6 hours each. With Copilot the draft takes 4 hours, plus 0.6 to recheck figures and references. The net saving is 2 gross hours minus 0.6 of rework, which is 1.4 per bid and 28 a month: at $50 an hour that is $1,400, against $240 of licences (8 × $30). The sturdier version uses those hours to produce about 6 more bids, if the market wants them. The structure is gross time minus rework, applied to the volume, set against a reference cost: if you put “users say they save time” where the gross figure should be, you have collected an opinion. It is the kind of number we look for, as at the consulting firm where a presentation went from four hours to under one.

When is it right to say “it’s not ROI yet”?

When the before is missing, when the process is too rare to produce data in a few weeks, or when the team is still learning and cycle time is, for now, longer. It is more honest than promoting the project on the dashboard or on a satisfaction survey. A project promoted on the wrong numbers gets none of the course correction it needs and, at the first budget review that looks at the profit and loss, has little to show.

If the pilot has already started and there is no before, start again from a single process: that is Map, the first move in every one of our adoption and change management projects. For the question upstream, see what Copilot really costs.