The ROI on AI: how to actually measure it

Ask a room of directors what their AI programme returned last year and you get one of three answers. A number of licences. A number of hours saved. Or a slightly embarrassed silence.

None of those is a return. The first is a cost, the second is an input, and the third is at least honest.

Why hours saved is not a result

Say a tool saves each of your forty people four hours a week. That is 160 hours, roughly four full-time equivalents. It is a good slide and it changes nothing on its own.

Nothing moves until one of three things happens. The cost comes out, through a role you do not backfill or contractor spend that stops. The capacity gets sold, through more delivery from the same team. Or the hours get redirected into something that compounds, such as consistent follow-up on pipeline you already hold.

Most programmes stall precisely here. The tools work, the time appears, and it quietly gets absorbed. If you cannot say where the hours went, they went nowhere.

Three baselines, taken before you deploy anything

Measuring return afterwards is close to impossible if you did not measure anything before. Three baselines are enough, and all three are cheap to capture.

Adoption. What share of the people who could use this, do, weekly. Not licences issued. Weekly active humans.

Process. How long the target process takes today, end to end, and how many people touch it. One number and one count.

Outcome. The business metric the process feeds. Proposals sent. Days to invoice. Support tickets resolved without a human. Pick the one your board already asks about.

The third is the one people skip, because it is the one that can embarrass you. It is also the only one that is a return.

Tie every workflow to a driver before you build it

The discipline that makes this work is deciding, in advance, which line of the P&L a piece of AI is supposed to move. There are only seven candidates: lead volume, conversion, average deal size, sales mix, gross margin, overhead spend and overhead efficiency.

If a proposed workflow does not map to one of those, it is not a business case, it is an experiment. Experiments are fine, and they should be labelled as such rather than counted as return.

What good reporting looks like

A board pack that says: adoption is 68% weekly active, the proposal process went from six hours to ninety minutes, proposals sent per month rose from 22 to 31, average deal size held, and the resulting gross profit is worth £X against £Y of tooling and time.

That is a return. Everything else is activity.

The number most people miss

A permanent improvement is worth more than this year's saving, because it is valued at your EBITDA multiple. At 6x, £50,000 of structural EBITDA improvement is £300,000 of enterprise value.

Which means the question is not only what AI returned this year. It is what it did to what the business is worth.