This piece is for the CEO or CFO who has to say, a year after go-live, whether the AI platform paid its way.

Four measures, in business terms

Pick the ones that fit the job your AI agents take on. Each is stated in units your business already uses: hours, days, errors and rand.

  1. Time given back. Time per piece of work, times the volume. Count it only where the hours go somewhere: more work handled without new hires, or people moved to work that earns more. Hours that disappear into the week are a nicer week, not value.
  2. Revenue reached. Work that brings money in sooner or more often: quotes sent the same day, enquiries followed up, renewals chased on time. Count only the change you can link to the platform, ideally against a team or period that ran without it.
  3. Errors avoided. The error rate per hundred pieces of work, and what each error costs to fix: rework, credit notes, a customer lost.
  4. Software spend retired. Licences actually cancelled, and tools no longer renewed. A tool that "could be" cancelled counts once it is cancelled.

Five rules that keep the numbers honest

  1. Take the baseline before go-live. Count one full month of time, volume, errors and spend on the job before go-live. Without it, every later number is an estimate.
  2. Measure net. Subtract what it costs to run: the platform, and the time people spend checking what the agents do. A job that saves time but needs constant checking may save little.
  3. One owner signs each number. The person who runs the work confirms the figure, not the supplier who built the tool.
  4. Keep activity out of the value line. Tasks run and prompts sent show the platform is busy. They help you run it, not judge it.
  5. Review on a fixed cycle. Read the same measures the same way each quarter, so a trend means something.

A worked example

Take a mid-sized distributor that gives AI agents its supplier invoice capture. It illustrates the method; it is not a client.

Before go-live, the finance manager counts how long each invoice takes, how many arrive each month, how often one is captured wrongly, and which capture tool the team pays for. After a quarter, they count the same things again.

Time given back is the time per invoice saved, times the volume, less the time spent reviewing the agents' work. It counts only because the team took on month-end work that had sat with an outside bookkeeper. Errors avoided is the drop in wrong captures, times the cost of fixing one. The old tool counts as retired spend once its renewal is cancelled. The finance manager signs all three.

Where the method breaks

The baseline was never kept. If errors were never logged, start counting now, even if go-live waits a month.

The market moves. A good quarter can flatter revenue measures. Use a comparison group or period, and say plainly when you cannot separate the effect.

Spend moves rather than disappears. A licence cancelled in one team and bought in another is not retired spend.

What stays with people

The AI agents can count. People decide what counts. The owner of each job sets the measures, signs the figures and decides whether the agents earn more of the work. That is earned autonomy: scope grows against evidence, not enthusiasm.

Clear signal, not noise

An AI platform worth keeping shows its value in units your business already trusts. Set the baseline before it goes live, measure net, and let the person who runs the work sign the number.

At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets. If you want a second view on how to measure the value of your first AI platform, Start an Agentic conversation →.