What return on AI were you expecting?

You already know what AI tools are costing you, but are you measuring the value that you're getting back from them?

The short version (TLDR)

• Twice as many law and accountancy firms were using AI last year as the year before. The number checking what they got back stayed at 18%.
• Working out a figure is harder than it sounds. If you bill by the hour, work getting faster shows up as a smaller invoice.
• The first few months of any change also look worse than they really are.
• What you can find out this week is who is using AI and whose account they are on.
• Paying by the head is why most of your employees are not on a company account.
• Some questions worth asking your team at the bottom.

The metrics gap that opened last year

Thomson Reuters asked more than 1,500 people across 27 countries in legal, tax and accounting whether anyone at their organisation kept track of what they got back from AI. Eighteen per cent said yes. That was about the same as a year earlier, even though in those twelve months the number of organisations using AI went from 22% to 40%. 

Something similar happened across the rest of the economy. The Office for National Statistics asked businesses directly whether they use AI, and among those with ten or more staff, the answer has gone from about 12% in late 2023 to about 35% now. 

Plenty of small businesses do know what their AI is worth to them and had no reason to write about it anywhere, so the lack of published examples proves very little.

The story on Uber's AI budget

Uber gave its engineers Claude Code, Anthropic's coding tool, in December 2025. By March this year more than eight in ten engineers were using it, and around 70% of the code the company shipped came from an AI tool. Four months into the year the whole AI budget for 2026 had gone.

In June the company put a limit of fifteen hundred dollars per engineer per month. Most of the reporting made that a story about AI costing too much. Uber's chief operating officer had said something more specific in May. Andrew Macdonald's point was that the money going out had not yet turned into anything a customer would notice. He could see one of those two things and not the other, so he capped the one he could see.

Uber could move that quickly because of some dull admin. People knew which tools they were allowed to use. The work sat in a shared system the company ran, so anyone could go back to it later. Somebody signed off the output before it reached a customer. Spending had a ceiling on it.

That is admin rather than software, and none of it needs a big company behind it. A Remi workspace already has the shared system and a hard ceiling, so the worst month you can have is a number you picked.

How does it work in smaller businesses

A tool charged by usage, plus one person running something heavy, can produce a bill that a business of twenty feels in a way Uber never would. The person paying those bills is often the owner. There is no spare month to soak it up and nobody further up the chain to do it for you, which is why the ceiling for AI usage matters more here.

The less dramatic version costs more in the end. Nothing arrives in the post to tell you the AI produced very little, so nobody has a reason to go and check. A cost that never causes a fuss gets renewed every year, until somebody eventually decides the whole thing did not work.

Measuring it properly runs into something odd. If you bill by the hour, work getting faster arrives as a smaller invoice. An hour saved is an hour off the bill unless somebody repriced the job, so a real improvement can read as lost revenue.

Timing is against you as well. Three economists, Erik Brynjolfsson, Daniel Rock and Chad Syverson, working from national productivity figures, wrote about what they call a productivity J-curve. New technology needs work around it, redesigning a job or retraining the people doing it, and that gets paid for straight away while the benefit arrives later. The numbers dip first and rise afterwards, so the worst-looking month can be the one where it starts paying. 

You can skip all of that and still find out who is using AI and whose account they are on. Harmonic Security looked at 1.9 million minutes of AI sessions across six of the main tools over seven weeks to April 2026. On personal accounts, 64.5% of it was work rather than personal use. That is the market rather than your business, but it does mean a licence count tells you very little about how much AI is going on.

Pricing explains a lot of it. Buy AI by the head, and you end up covering the three or four people who asked, while everyone else does the same work on their own login. Remi charges one price for the whole workspace, so bringing in the seventeenth person costs nothing.

Questions worth asking

Every one of these can be answered with what you already have.

Who is actually using AI, and whose account are they on? Ask everyone anonymously what they used for work last month, and whether it was a company login or their own. Anonymity matters, because a signed answer gives you the reply people think you want. Then hold the total up against your headcount rather than against the licences you bought.

How does a new tool get adopted here? Somebody found the AI, somebody paid for it, and at some point other people started using it. Following that chain backwards tells you more about how your business takes on anything new than it does about AI. In a business of twenty there is often no step where anybody decided.

Have we written anything down, and does it match what people do? Plenty of businesses have nothing written, which is a normal place to start from. If there is something, read it next to the answers from the first question. A page describing four approved users, in a business where fifteen people use AI daily, will not help anyone on the day it is needed.

What should AI be doing for us at all? The hardest one, and the one you are least likely to answer from inside. The useful version usually comes from somebody who has watched it go well and badly at a dozen businesses your size.

If you want a quick read on any of it, turn the AI tool off on Monday and see who complains. Silence is an answer, and a cheaper one than waiting for the renewal. If the same two people complain about the same job, that is the bit that works, and that is where the next change needs to be made.

For anyone in a regulated practice, the first question got less optional over the summer. The Solicitors Regulation Authority said in August that solicitors

"remain accountable for all work and advice delivered to clients... whether or not AI has been used in its preparation"

Accountancy's own bodies (CCAB) published draft guidance in June telling members to put AI use in the engagement letter. Neither asks you to prove a return. Both need you to know what your people are using. 

Remi AI can help answer the first question for free for businesses of twenty to sixty people. It contains anonymous questions for your team, giving you a page back that tells which tools are in use and whose account each one sits on. This requires no systems access, no data leaving your side, and nothing for anyone to run a security review over.

If per-seat AI pricing is holding you back from using AI to its fullest, then that is a pricing decision rather than a verdict on AI tool value. It is also the decision Remi exists to remove.