How to measure the ROI of AI in a professional services firm (2026)

The simplest way to measure the return on AI is to compare what you're paying with what's actually changing in the work. I'd look at four things: cost (what you pay against what gets used), time (hands-on time against the total time it takes to finish), quality (how senior people score the work on real tasks) and risk (whether flawed AI output gets caught before a client sees it). Then measure them against people who haven't been trained yet, and do it again each quarter.

Very few firms are doing this today. In the Thomson Reuters Institute's 2026 AI in Professional Services Report, only 18% of respondents said they knew their organisation was tracking the return on its AI tools, about the same as a year earlier. Another 40% didn't know either way. Over the same year, the share whose organisation uses AI across the business rose from 22% to 40%. In the UK, 38% of businesses with ten or more staff now use AI, according to the Office for National Statistics.

Why is AI so hard to measure?

I think part of the reason is that a lot of firms bought AI to keep up, before they really knew what it was for or how they'd know if it was working. So the question of measurement tends to come up after the money has already been spent.

The other reason is that the easy numbers are the wrong ones. In the same Thomson Reuters report, among organisations that do measure the return, the most common measures are internal cost savings (77%) and employee usage (64%). Few measure anything outward-facing, such as client satisfaction, revenue or new business won. Usage tells you people opened a tool. It doesn't tell you whether the work got any better. I wrote about this in Why usage data won't tell you if AI is paying off.

The wider evidence points the same way. A March 2026 study of almost 6,000 executives in the US, UK, Germany and Australia, published as a Federal Reserve Bank of Atlanta working paper, found more than 80% of firms reported no impact from AI on either productivity or employment over the previous three years. BCG's September 2025 survey of 1,250 executives found 60% of companies weren't getting material value from AI at all.

What should you measure?

I'd focus on four things, as each one answers a question a managing partner is likely to ask.

1. Cost: what are we paying for, and what gets used?

Start by listing every AI tool the firm pays for, with the number of seats and the monthly cost, and then compare the seats with the people actually using them. Microsoft 365 Copilot Business is £16.10 per user a month on an annual plan, before VAT and any promotional discount. So a firm with 60 seats and 44 regular users is paying about £258 a month for nothing.

Then add the tools people are using that the firm doesn't pay for. These carry a different kind of cost, which is client data going somewhere you haven't approved. I've covered both gaps in Paying for AI nobody uses.

2. Time: is the work finished sooner?

Self-reported time savings are a good starting point, but I'd treat them with some caution. In the UK government's cross-department Copilot experiment, involving 20,000 civil servants, participants reported saving an average of 26 minutes a day. In HMRC's own trial, published in July 2026, participants reported saving about an hour a week. And the Department for Business and Trade's evaluation said: "We did not find robust evidence to suggest that time savings are leading to improved productivity."

Two things help here. First, time real tasks rather than asking people to estimate. Second, separate hands-on time from the total time to finish. AI speeds up the part where someone is actually working on a task, but it does nothing for the days a draft spends waiting for a review or for the client. If most of a job's time is spent waiting, faster drafting won't change when the client gets it.

3. Quality: is the work better, or just faster?

Being faster isn't a return if the work gets worse. When the law firm Ashurst tested generative AI with 411 of its people in 2023 and 2024, AI cut the time to produce first drafts by between 45% and 80%, depending on the task. But on one task, summarising UK cases, a blind panel of expert lawyers scored the AI drafts 3 out of 5 for accuracy, against 3.5 for drafts written by people. Ashurst itself warned the results couldn't simply be extrapolated.

So I'd score real work the way your senior people already judge it, against a short, fixed set of criteria, by reviewers who don't know who used AI.

4. Risk: are flawed outputs caught before a client sees them?

One AI error that reaches a client can be costly, and you can't blame the tool. In February 2024 a Canadian tribunal ordered Air Canada to compensate a passenger after the chatbot on its website gave him wrong information about bereavement fares. Air Canada argued it couldn't be held liable for information from its chatbot. The tribunal called this "a remarkable submission", and said: "It should be obvious to Air Canada that it is responsible for all the information on its website." It's worth measuring how often your people catch flawed AI output before it leaves the firm. My checklist for checking AI output is a good place to start.

How do you know AI caused the change?

Firms change all the time. People join and leave, a big client lands, or a quiet quarter frees up time. If you only compare before with after, you can't separate the effect of AI from everything else going on. The answer is a comparison group: people doing the same work who haven't had the training or the tool yet, measured the same way over the same period. I explain how to set one up in How to prove AI training worked.

How do you turn it into pounds?

My advice is to only convert what you've actually measured. Hours saved multiplied by a cost rate gives you a figure you can defend, and seats cut gives you another. Quality and risk are harder to price, so I'd report them as measured, as scores and as the share of flaws caught, and let the partners judge their value rather than inventing a number. A smaller figure you can stand behind is worth more than a big one nobody believes.

It also helps to keep the two kinds of number apart: what you measured, and what you worked out using your own rates.

Where should you start?

1. List every AI tool, its seats and its monthly cost.
2. Compare seats with active users, and ask people which tools they actually use.
3. Pick one or two jobs that come up every week, and time them.
4. Ask senior people to score a sample of the output against fixed criteria.
5. Agree a comparison group before anyone is trained.
6. Measure again after a quarter, and report the difference.

This is the approach I'm taking with Osava: test people on real work from their sector, compare the results with people not yet trained, and show spend against use in one report. Whether you use us or do it yourself, the principle is the same. Measure the work, not just the logins.