How to tell whether your AI investment is actually working

I've sat in a lot of boardrooms over the last eighteen months where the same conversation plays out. The business bought AI tools six or nine months ago. The team is using them, or says it is. The monthly invoices keep arriving. And the owner has a nagging feeling that sits somewhere between "this is probably working" and "I have no idea."

That feeling is the right one to pay attention to. If you can't say in one sentence what your AI spending has produced, the problem is almost always the measurement, not the tools.

This piece is about how to measure it properly. Not with dashboards or KPI frameworks borrowed from enterprise, but with the kind of questions a 20 to 50 person business can answer with what it already knows.

Why is AI ROI so hard to measure for SMEs?

Three reasons, and they compound each other.

The first is that most AI tools produce diffuse value. A Copilot licence doesn't generate revenue. It saves seven minutes on an email, twenty minutes on a report, an hour on a proposal. Those savings are real. They're also invisible unless you go looking for them, because they don't show up on any line of your P&L.

The second is that the baseline was never recorded. You can't measure "time saved" if you never measured the time it took before. Most SMEs adopted AI tools without benchmarking the tasks those tools were meant to improve. The result is that six months later, everyone agrees things feel faster but nobody can put a number on it.

The third is that AI spending is spread across several budget lines. The Copilot licence is in IT. The training came out of HR. The consultant who set up the workflows was a one-off project cost. The meeting transcription tool sits on someone's corporate credit card. Nobody has added it all up, which means nobody can divide the total by the return.

If you recognise all three of these in your own business, you're in the majority.

What's the wrong way to measure?

Two approaches that look sensible and tell you nothing useful.

The first is usage metrics. "80% of our team logged into Copilot this month" is a number that makes everyone feel better and measures nothing that matters. Usage is an input, not an outcome. A team member who opens Copilot once a day to summarise an email is "using" it. That doesn't mean it's producing value worth the £25 a month you're paying for their licence.

The second is satisfaction surveys. "On a scale of 1 to 5, how useful do you find AI tools?" tells you how people feel about the tools. It doesn't tell you what changed in the business. People will rate a tool highly because it's novel, because they think you want them to, or because the alternative is admitting they haven't learned to use it. None of those correlates with business value.

Both of these measures are common because they're easy to collect. The things worth measuring are harder to get at.

What should I measure instead?

Four questions. Answer them with a straight face and you'll know more about your AI ROI than most businesses ten times your size.

Question 1: What specific task takes less time than it did before, and by how much?

Pick the three tasks you expected AI to improve most. Estimate how long each one took before. Time them now. That difference is your productivity return, and the most concrete number you'll get.

I worked with an engineering firm last year that tendered 160 to 220 times a year, with each response consuming four to five days of effort. Hours went into locating historical project data, cost references, and technical precedents buried in a 30-year archive. The diagnostic showed that data locked in unreadable formats was the real constraint, not the tools. Once the business ran a structured data audit and built a system to make that archive searchable, the tender response time dropped. That drop, multiplied by the number of tenders per year, produced a measurable productivity return in the range of £340,000 to £470,000 annually across commercial, design, and operations.

That number didn't come from a dashboard. It came from knowing how long the task took before, knowing how long it takes now, and doing the arithmetic.

Question 2: What can the business do now that it couldn't do six months ago?

This is where the value often hides. Not "we do the same things faster" but "we do things we didn't do before." A 15-person accountancy firm I spoke to last quarter started using AI to draft advisory notes for clients on regulatory changes. Before AI, they didn't offer that service because the research time made it unprofitable. Now they charge for it. That's new revenue, not saved time, and it's far easier to measure.

If nothing comes to mind when you ask this question, your AI adoption has been about efficiency rather than capability. Efficiency gains are real but they plateau. Capability gains compound.

Question 3: What's the ratio of AI spending to identifiable return?

Add up every pound you've spent on AI in the last twelve months. Licences, training, consulting, data work, internal time. Put a number on it. Then add up every identifiable return: hours saved (valued at cost), new revenue generated, errors avoided, work brought in-house that was previously outsourced.

For most SMEs doing this properly, the ratio in year one lands somewhere between 1:1.5 and 1:4. If your ratio is below 1:1, you're either measuring too early (give it six more months) or you've spent money on the wrong things. If you can't calculate the ratio at all because you don't know what you've spent or what you've gained, that's the first thing to fix.

Question 4: What would you lose if you cancelled everything tomorrow?

This is the question that cuts through the noise. If you switched off every AI tool and cancelled every licence in the morning, what would hurt by Friday?

If the answer is "not much," your AI adoption hasn't reached the point where it's producing real value yet. If the answer involves specific people who'd lose specific capabilities they now depend on, you've got something worth protecting and investing in further.

The engineering firm I mentioned earlier had Copilot licences sitting idle for months before the diagnostic. If they'd cancelled those licences, nobody would have noticed. After the structured enablement programme, with role-mapped training and specific use cases per function, cancellation would have meant the commercial team losing a tool they'd built their weekly workflow around. Same tool. Different adoption. The measurement changed because the value changed.

When should I cut an AI tool, and when should I invest more?

Two signals for each.

Cut when a tool has been live for more than three months and you can't name a single specific outcome it's produced. Not "it helps with things" but a named task that takes less time or produces a better result. If nobody on the team can give you that sentence after three months, you've got a use case problem. More training won't fix that.

Cut when the cost of a tool exceeds the value of the time it saves by more than 50%, and there's no capability gain to offset it. A £300-a-month tool that saves two hours of £30-an-hour work is losing you money. That calculation is worth doing for every tool on the list.

Invest more when a tool has produced a measurable return and you're using a fraction of what it can do. The pattern I see most often in SMEs is a tool that's delivering well in one function while the rest of the business ignores it. Expanding adoption to the next team is almost always cheaper than buying a new tool for them.

Invest more when the constraint on value isn't the tool but the data feeding it. This was the pattern with the engineering firm. The tools were fine. The data was locked in formats the tools couldn't read. The right investment was a data audit and a system to make 30 years of project history searchable, not another software licence. That investment, in the range of £110,000 to £220,000, was designed to self-fund from first-year savings. The tool spending that followed it produced returns the earlier tool spending never could.

What should I do this week?

If you've been running AI tools for more than three months and haven't measured the return, three things.

First, add up what you've spent. Every licence, every training cost, every hour of internal time someone spent configuring or learning a tool. Most owners I ask have never done this. The number is usually 20 to 40% higher than they expected, because nobody counted the time.

Second, pick the two tasks where AI was supposed to make the biggest difference. Ask the person doing each task to estimate the before and after. You don't need a time-and-motion study. You need an honest estimate from the person who does the work. If they can't give you one, that tells you something too.

Third, answer the cancellation question for each tool. If you'd lose nothing by switching it off, you've got a subscription problem. If you'd lose something specific, you've got a tool worth expanding.

The businesses getting real value from AI in 2026 can describe that value in a sentence. The ones not getting value can't, and they don't know why. The difference between those two groups is rarely the tools. It's whether anyone stopped to measure what the tools produced.