Measuring the return on AI: the numbers worth tracking

Measure each job against the number you wrote down before you started, and measure the business on four numbers a quarter. We can't say what AI did for the company named in the question – only they can, and they'd find out the same way.

By Matthew Steele4 min read28 October 2026

A quarterly review in a small meeting room, a printed one-page scorecard on the table with four boxes, two managers comparing it with last quarter's
Illustrative scene, not a customer's premises.

Per job: the before and the after

This is the clean measure, and we think it's the only one that is. Before the tool gets built, the job costs a known number: 11 hours a week across the team, 4 days from enquiry to quote, 1 in 8 quotes with an error in it. After a quarter, you measure the same number the same way. The difference is the return, in the job's own units, and the three lines of the budget are what it cost you.

If there's no before number, there's no return to measure – only a feeling, and feelings about new technology run optimistic. That's why the 90-day loop spends its first two weeks measuring.

The cheat to avoid

"We saved three hours a week per rep." Saved from what, and what happened to them? If the three hours became more calls, that shows up in calls. If they became an earlier finish, that's a benefit to the rep and not a return to the business, and that's fine as long as you call it what it is. Saved hours nobody redeployed aren't a return. They're capacity, and capacity is only worth something once someone decides what to do with it.

The business level: four numbers

Jobs roll up into teams, teams into the businessThree levels from bottom to top. The job: a before number and an after number in the job's own units, clean. The team: hours redeployed into output, mostly clean. The business: four numbers a quarter, profit per head, customer satisfaction, headcount and roles, staff satisfaction, with a note that the signal gets noisier going up and attribution must be honest.The jobEleven hours a week before, three after. Four days to quote before, one after. Clean, because it was measured before the build.The teamSaved hours become calls, quotes or an earlier finish. Say which. Capacity is not a return until someone decides what it is for.The businessProfit per head. Customer satisfaction. Headcount and roles. Staff satisfaction. Quarterly, one page.Noisier
Jobs roll up into teams, and teams into the business. Measure at every level, and expect the signal to get noisier the higher you go.
Number How to measure it simply What the tools can plausibly move
Profit per head Operating profit divided by full-time equivalents, each quarter. Up, if hours become output, or if the same output needs fewer hours and the business chose that.
Customer satisfaction A one-question score after delivery, or the complaint rate, or both. Up, through faster responses and fewer errors. Down, if a customer-facing agent is poor.
Headcount and roles The number, plus a one-line description of what each role actually does now. The roles change before the number does. Write the change down.
Staff satisfaction A three-question survey each quarter, and turnover. Up, when the hated tasks shrink. Down, when people fear for their jobs and nobody's said anything.

Four numbers, one page, every quarter, with last quarter alongside. It isn't sophisticated. It's enough to see a direction, which is about all anyone can claim with a straight face at the business level.

Working out what caused what

A good quarter has a lot of parents. The tools arrived, and so did a new sales manager, a price rise and a competitor's stumble. The fair statement is "profit per head rose while these three agents were running and the quoting process changed", not "AI increased profit". Keep the per-job numbers, which are clean, next to the business numbers, which aren't, and let whoever's reading see both.

On the company named in the question

We can't tell you what AI did to another company's profitability, customer satisfaction, headcount or staff satisfaction, and anyone who answers that on their behalf is guessing. If they measured the four numbers above before and after, they can answer it in a page. If they didn't, they're estimating – as most businesses are.