How do you measure whether AI is working?

Measure three things, in this order: who is actually signing in, what happened to the specific work you defined at the start, and whether the money changed. Hours saved is the number everyone reaches for and the easiest to inflate, because an hour freed is only worth something if you can say what filled it. A defensible position at renewal looks like this: here are the licences we pay for, here is how many were used last month, here is the one process we measured before and after, and here is what we did with the difference.

1. Who is signing in

Every major tool reports this. It's the least flattering number available and the most useful, because it's the one your finance team can check without you. Read it honestly: a licence that logged in twice last quarter is a licence, not an adoption story.

2. What happened to the work you defined

This only exists if you wrote down a baseline before you started — how long the task took, how many rounds it went through, how often it came back. One process measured properly beats nine estimated. If no baseline was recorded, say the number is unavailable rather than reconstructing one from memory.

3. Hours versus dollars

Hours saved becomes money in three ways only: you spend less, you sell more, or the work that was queued gets done. Name which one applies. If none does, you have found capacity rather than savings — still worth having, and worth describing accurately, because a CFO who catches an inflated saving will discount every number that follows it.

The measure to stop using

Sentiment. Surveys after a workshop tell you the session went well, which is a fact about the session. Keep it for improving the training; keep it out of the business case.

How to make the numbers hard to argue with

• Agree the baseline in writing before anything changes, and have the person who will later question it countersign.
• Pick a small number of processes and measure them properly.
• Report the sign-in data unedited, including the months it looks bad.
• Say plainly which of spend, revenue or throughput moved.

Why this comes first in an engagement

An AI programme that cannot show its own numbers gets cut in the first tight quarter, regardless of merit. Getting the baseline recorded early is cheap, takes a day or two, and is the difference between a renewal conversation with evidence in it and one held on impressions.