Is your AI actually paying for itself?

Your firm is probably using AI in a dozen small ways, and if someone asked you which of them actually pays for itself, you would struggle to answer. Most leaders would. This is a way to fix that for one use, without a dashboard, a new hire or a consultant: a short payback check you can run on a single task this week, ending in a plain decision. Keep it, fix it, or drop it.

The reason the honest answer is usually "no idea" is not laziness. It is that AI use spread through most firms faster than any way of measuring it. In the Thomson Reuters 2026 AI in Professional Services Report, a survey of more than 1,500 professionals across 27 countries, only 18 per cent said their organisation tracks the return on its AI tools in any form, about the same share as a year earlier. Among the few who do measure, most track internal activity, seats used, prompts run, rather than anything the business cares about, like work won or client time saved. So the picture across the sector is a lot of AI in daily use and almost no idea whether it earns its keep. If that describes your firm, you are in the majority, not behind.

Why "it saves us hours" is not the answer

The instinct, when asked whether an AI tool pays off, is to reach for time saved. The team says the assistant saves them six hours a week, everyone nods, and the question feels answered. It is not, for two reasons that a proper check has to catch.

The first is that time saved is only the chance of value, not value itself. If those six hours quietly refill with more of the same low-value work, or with a longer lunch, nothing reached the business. The saving becomes real only when the freed time is pointed at something that shows up in the numbers. We have made that argument in full elsewhere, on the gap between AI productivity and business value and on why adoption keeps outrunning results. Here the point is narrower and practical: any check that stops at "hours saved" will overstate the return every time.

The second is that AI carries a cost most tallies ignore. The subscription is the obvious one and usually the smallest. The larger cost is the time your people spend checking and correcting what the tool produces, because a confident, wrong output that reaches a client costs far more than the minutes it saved. That verification work is real, it is skilled, and it belongs in the sum. Leave it out and the tool looks cheaper than it is.

The payback check: three numbers and a decision

Pick one use. One task, one tool, one team, something specific enough to reason about, like "drafting first-pass candidate summaries" or "summarising client calls". A vague "our AI" cannot be measured; a single use can. Then work out three numbers. Rough and defensible beats precise and imaginary.

One, what it saves. Time saved per instance, multiplied by how often the task happens, multiplied by a loaded hourly cost of the person who used to do it. Loaded means salary plus the on-costs, not the headline wage. This gives you a weekly or monthly gross saving.

Two, what it truly costs. The share of the licence or subscription for this use, plus the time spent checking and fixing the output, priced at the same loaded rate, plus any one-off setup. Subtract this from the gross saving. What remains is the net time the use actually frees.

Three, what the freed time does. This is the number most firms never write down, and it decides everything. Does the freed time convert into something the business measures, more matters handled, a faster turnaround a client will pay for, senior people moved onto higher-value work? Or does it evaporate into a fuller inbox? If it converts, your net saving is real value. If it evaporates, the use is running at a notional profit and a real cost.

Then the decision, in plain terms. Keep it if the net saving is clear and the freed time converts. Fix it if the saving is real but leaks, which usually means the tool is fine and the workflow around it needs a change, or the verification is heavier than it should be. Drop it if, once you have honestly priced the checking, there is little left, or nobody can say where the time went.

A worked example

Take a recruitment consultant using AI to draft first-pass candidate summaries. (An illustrative example, not a specific firm.) The tool saves about fifteen minutes a summary, and they produce forty a week. At a loaded cost of roughly forty-five pounds an hour, that is ten hours saved, about four hundred and fifty pounds of gross time a week.

Now the true cost. The licence is about twenty pounds a month, trivial here. The verification is not: every summary still needs a few minutes of reading and the occasional real correction, call it three hours a week, roughly one hundred and thirty pounds. So the net freed time is closer to three hundred and twenty pounds a week, not four hundred and fifty. Already the honest figure is a third smaller than the happy one.

The number that settles it is the third. If the consultant spends the freed hours on more candidate calls and fills even one extra role a quarter, the use has paid for itself many times over and you should keep it and do more of it. If the freed time simply absorbs into a busier day with the same output, that three hundred and twenty pounds is notional, the verification cost is not, and you have a fix on your hands: decide, deliberately, what the recovered time is for. Same tool, same saving, opposite verdict, and only the third number tells them apart.

The one most firms forget

If you run this check on three or four uses, a pattern tends to show up. The tools are rarely the problem. The saving is usually real at the task level and then leaks somewhere between the task and the business, either because the workflow swallowed it or because nobody decided where the freed capacity should go. That is also why buying another tool so often changes nothing: it adds to the pile of task-level savings while the leak sits one step higher up.

Deciding where the recovered time goes is a leadership call, not a software setting, and it is closely tied to how you price work once AI has cut the hours it takes. The firms getting a return are not the ones with the most licences. They are the ones who picked a use, measured it honestly including the checking, and then made a deliberate decision about the time it freed.

That decision, run across your real work rather than in the abstract, is the core of AI Lessons for Leaders: one-to-one coaching that takes the tasks you already handle, works out which AI uses genuinely pay, and turns the good ones into repeatable workflows with the judgement kept in your hands. If you want to stop paying for AI on faith, start with one use and a clear head. Book a one-to-one discovery call and bring the use you are least sure about.

Questions leaders ask

How can a small firm measure AI ROI without a big analytics project?

You do not need one. Take a single use and estimate three things you can defend: the time it saves at a loaded hourly cost, the true cost including the time spent checking its output, and whether the freed time converts into something the business measures. That gives you a keep, fix or drop decision for that use in an afternoon. Do it for your two or three biggest uses before you think about any tooling to automate the tracking.

Why include the time spent checking AI output as a cost?

Because it is real work and it is often the largest cost after the tool itself. An AI draft that still needs a skilled person to read, correct and sometimes rewrite it has not saved as much as the raw time suggests, and a wrong output that slips through unchecked can cost far more than it ever saved. Pricing the verification honestly is what separates a genuine saving from an optimistic one.

What if the time saved does not turn into anything measurable?

Then the use is running at a notional profit and a real cost, and that is useful to know. It usually means one of two things: the workflow around the task needs to change so the saving is not immediately swallowed, or you have not decided what the freed time is for. Neither is a reason to buy a different tool. Fix the leak first, and only drop the use if, once the checking is priced in, there is little genuine saving left.