How do you measure the ROI of AI?
Measure AI ROI against a baseline captured before rollout, combining time, quality, revenue and risk measures, and count full costs including training.
Why time saved can mislead you
Time saved is the easiest number to collect, so most AI business cases lead with it. On its own it is also the most misleading. Suppose a team saves an hour a day, but that hour drifts into more meetings or slower work elsewhere. The business has paid for a tool and gained nothing measurable.
Time saved only becomes a return when you can show where the time went. That could be more client work billed, faster turnaround, a backlog cleared, or a hire you no longer need to make.
Capture a baseline before you start
You cannot prove a change without knowing the starting point. Before any rollout, record these for each task in scope:
Volume: How often the task happens per week or month.
Time per task: How long it takes now, measured on real examples, not estimates.
Quality: Error rates, rework, complaints or review comments.
Cost: Who does the work and at what internal rate.
If you have already rolled out AI without a baseline, measure a team or task that has not yet adopted it and use that as a comparison. It is imperfect, but far better than relying on memory.
Use a balanced set of measures
A credible ROI case combines four types of measure, because each catches what the others miss:
Efficiency: Time per task, throughput and turnaround. Useful, but only meaningful alongside where the saved time goes.
Quality: Fewer errors, more consistent outputs, better client feedback. Clients often value quality gains more than speed.
Revenue: More proposals sent, faster response to leads, extra capacity for billable work. These are harder to attribute but closest to what a board cares about.
Risk: Fewer compliance slips, better documentation, less reliance on one person's knowledge. Risk reduction rarely shows in a single quarter but can be the largest long-term value.
Count the full cost
Licence fees are the visible cost and usually the smallest. A fair ROI calculation also includes:
Training time: Hours spent learning the tools, including the productivity dip while people adjust.
Management time: Time leaders and champions spend setting rules, reviewing outputs and answering questions.
Integration and setup: Any configuration, data preparation or external support.
Ongoing review: Human checking of AI outputs, which continues after the trial ends.
Leaving these out flatters early results. Later reviews then look like failure when the true picture finally appears. For a breakdown of where costs tend to fall, see how much an SME should budget for AI.
Report honestly and on a rhythm
Review results at set points, such as after the first month and again after a quarter. Early figures often look better than reality because enthusiastic early adopters do most of the work. Check whether the gains hold as use spreads to the wider team.
Be prepared to report what did not work. A tool that saves little time but noticeably reduces errors may still be worth keeping. A tool with high usage but no change in outcomes may not be. Many stalled initiatives trace back to weak measurement, a pattern covered in why AI pilots fail to scale.
How you value time, risk and capacity depends on your business model. A professional services firm will weight these measures differently from a manufacturer. Treat these figures as management information, and take advice from your accountant on how AI costs and benefits should be treated in your own accounts.