Measuring the return on an AI project

Many AI projects cannot show their value for a simple reason: nobody measured the process before the system went live. Without a starting point, every claim about improvement is a guess. We agree what success looks like before we build, and the same discipline works for any team.

Measure the baseline before you start

Before anything changes, record how the process performs today. Two to four weeks of data is usually enough to see a normal pattern.

• Volume: how many items arrive per day or week.
• Time per item: how long the work takes, from start to finish.
• Errors: how many items need correcting, and how long each correction takes.
• Response time: how long customers or colleagues wait for a reply or a result.

Use timestamps from your existing systems where you can. Where you cannot, time a sample of real cases. Write the figures down and agree them with the person who owns the process, so nobody disputes them later.

The measures that matter

• Time saved. Hours spent on the task before and after. Count it honestly: freed time is a benefit only when it goes into other useful work.
• Error rates. The share of items that need correcting, plus the cost of each mistake in rework or customer impact.
• Response times. Time to first reply and time to completion. For enquiries, speed often matters as much as cost.
• Adoption. The share of eligible work that actually goes through the new system, and how often people override it. Low adoption explains many disappointing results, and it is fixable with training and better design.

Then count the costs honestly. The return is the benefit minus the full cost, not just the build. Include the build itself, running and usage costs, integration work, ongoing maintenance and the staff time spent reviewing outputs and learning the new way of working.

An illustrative example

The figures below are invented to show the method. They are not from a client project.

Imagine a finance team that enters 400 supplier invoices a month by hand.

• Baseline: each invoice takes about 6 minutes, so 40 hours a month. One invoice in 25 needs correcting, and each correction takes 15 minutes, adding 4 hours. Total: about 44 hours.
• After automation: the system reads each invoice and a person reviews it in about 1 minute, which is under 7 hours. Corrections fall to around 6 a month, or 1.5 hours. Total: about 8 hours.
• Result: roughly 36 hours a month returned to the team.

Now apply adoption. If only three quarters of invoices go through the new system in the first month, the saving shrinks to about 27 hours. Multiply the hours by your own cost per hour, compare that with the full monthly cost, and you have a payback estimate you can defend.

Review on a schedule

Measure again at fixed points, for example after the first, second and third month, using exactly the same measures as the baseline. Each review should end in a decision: expand the system, adjust it, or stop.

Your next step

For a quick first estimate, try the Payback Forecast in our free tools. To set up a baseline and success measures for a specific workflow, book a free 30-minute consultation, or describe the challenge in the AI solution designer on our homepage.