Insights
MeasurementNovember 14, 20264 min read

How to measure AI ROI in a portfolio company.

A baseline before the build, four numbers tracked weekly, and a conversion to money the CFO agreed in advance.

AI return on investment in a portfolio company is measured the same way as any other operating improvement: a baseline taken before the change, a small set of numbers tracked weekly after it, and a conversion to money that the CFO agreed in advance. Most AI programmes cannot show a return because they skipped the first and third steps, not because the technology failed.

The test is simple. If the CFO would not put the number in the board pack, it is not a return yet.

Take the baseline before anything is built.

For the process being changed, record four things over at least four normal weeks:

  • Volume. Items handled per week.
  • Effort. Minutes of staff time per item, measured by sampling, not by asking.
  • Quality. Error and rework rate, and what an error costs to fix.
  • Cost. Fully loaded cost of the people and any outsourced work.

A baseline taken after the build is a guess, and everyone in the room will know it.

Track four numbers, in this order.

1. Adoption. The share of real volume going through the new process. This comes first because nothing else counts without it. A tool that handles a small share of the work has not changed the cost base.

2. Hours saved. Volume through the new process multiplied by the minutes saved per item, net of the time people now spend reviewing the output. Leaving out review time is the most common way these figures get inflated.

3. Quality. Error rate against the baseline. If quality has dropped, the hours saved are borrowed from future rework.

4. Running cost. Model usage, software licences, and the engineering time needed to keep it working. This is small at pilot scale and worth watching as volume grows.

Convert hours to money honestly.

Hours saved are not EBITDA until something else changes. There are three legitimate conversions, and the business case should say which one applies:

  • Capacity. The team handles more volume with the same headcount. The value is the hiring that did not happen, or the revenue the extra capacity supports. This is the usual first result.
  • Cost removed. Outsourced work, overtime or contractor spend falls. This is the easiest to see in the P&L.
  • Headcount. Roles are not backfilled or are removed. This is real, slower than people expect, and should be agreed with management before it is counted.

Report the three separately. A single blended figure invites a challenge that the whole number then fails.

Payback.

Payback is the one-off cost of the build divided by the monthly net benefit. Include internal engineering time in the build cost, and subtract running cost from the benefit. State the figure as a range while the use case is still scaling.

Rolling it up for the fund.

An operating partner needs the same four numbers from every company on one page: use case, adoption, annualised net benefit by type, and payback. Two rules keep the roll-up credible. Count only use cases that are live with real volume. And keep capacity, cost removed and headcount in separate columns all the way to the total.

Five ways the number goes wrong.

  • Counting licences bought or staff trained as progress.
  • Using survey answers about time saved in place of measured minutes.
  • Ignoring the time spent checking the output.
  • Annualising a good pilot week across the whole company.
  • Claiming the same saving in two initiatives.

What a good monthly report looks like.

One line per live use case, with the baseline, the four current numbers, the benefit by type, and a one-sentence note on what changed. It should take a board member less than a minute to read and should match the finance system when someone checks.

Measurement is the fourth step in how I run this work, after assess, prove and ship. The Approach page sets out the full sequence.