How to Measure the ROI of an AI Project Before and After You Build It
November 11, 2026 · 4 min read · Etzal Group Team
Business cases for AI are often written with a single impressive percentage and very little arithmetic behind it. That makes them easy to approve and hard to check. Measuring AI project ROI properly takes four numbers before the project and the same four numbers afterwards. None of them requires a data scientist.
The four numbers
For the workflow you plan to change, find:
- Volume. How many cases per month. Count from system records where possible.
- Time per case. How long a person spends on one case today. Time ten real cases with a stopwatch. Estimates from memory are usually too high for dull tasks and too low for interrupted ones.
- Cost of errors. How often the task goes wrong, and what one error costs in rework, refunds or lost customers.
- Running cost. What the new system will cost each month: model usage, hosting, licences, maintenance and the time of people who review its output.
Everything else in the business case is built from these.
Building the estimate
Take a simple example. A team answers 3,000 customer enquiries a month. Each takes six minutes. That is 300 hours a month.
Suppose a pilot suggests that an AI drafting assistant will let staff handle a case in three minutes, because they approve or edit a draft instead of writing from nothing. The saving is 150 hours a month. Multiply by the loaded hourly cost of the staff involved to get a monthly value.
Now subtract the running cost: model usage for 3,000 drafts, the platform fee, and a few hours of maintenance. The difference is the monthly net benefit. Divide the build cost by that figure and you have the payback period in months.
Three rules keep this estimate honest:
- Use the pilot's measured time, not the vendor's. If there is no pilot yet, use a conservative guess and say that it is a guess.
- Apply an adoption rate. Not every case will go through the new path in the first months. Sixty to seventy percent is a more realistic starting assumption than one hundred.
- Show a low, expected and high case. A single number hides the uncertainty that a decision-maker needs to see.
Benefits that are not hours
Time saved is the easiest benefit to count and not always the largest. Consider:
- Speed. A quotation sent in an hour wins work that a quotation sent tomorrow loses. If you can measure conversion by response time, include it.
- Error reduction. Fewer wrong entries mean less rework and fewer credit notes.
- Coverage. Enquiries answered outside office hours that used to wait until morning.
- Consistency. The same question getting the same correct answer.
Put a number on these only where you have data. Where you do not, list them separately as unquantified benefits. Mixing guesses into the main calculation weakens the parts that are solid.
The costs people forget
- Review time. In the first months a person checks every output. That time is a cost of the project.
- Integration. Connecting to the CRM, ERP or document store is often more work than the AI part.
- Change management. Training, new procedures and a period of slower work while people adjust.
- Growth in usage fees. If volume doubles, model costs roughly double.
- Maintenance. Prompts, knowledge sources and integrations need upkeep as the business changes.
Measuring after launch
A business case is a prediction. Three months after launch, measure the same four numbers again, in the same way:
- Volume through the new path, and volume still handled the old way.
- Time per case, timed on real cases.
- Error rate, including errors the AI introduced.
- Actual running cost from invoices.
Compare with the baseline and with the estimate. If the result is better, say by how much. If it is worse, find out which of the four numbers moved the wrong way. That tells you whether the fix is adoption, quality or cost.
When the numbers say no
Sometimes the arithmetic shows that a workflow is too small to be worth automating, or that review time cancels the saving. That is a useful result. It is far cheaper to learn it from a spreadsheet and a two-week pilot than from a finished system.
A baseline, a costed estimate and a review plan are standard outputs of our AI Consulting engagements. If you already have the four numbers for a workflow, send them to us and we will tell you whether the case holds.
Summary
Count volume, time per case, error cost and running cost before you start. Build a low, expected and high estimate with a realistic adoption rate. Measure the same numbers after launch. An AI project that is measured this way can be defended, improved or stopped on evidence.
Frequently asked questions
What is a reasonable payback period for an AI project?
It depends on the size of the project, but a first project is easier to defend when it pays back within a year. If the estimate needs three years of perfect adoption to break even, the scope is probably too large for a first step.
How do we count time saved if nobody is made redundant?
Count the hours as capacity and say what the capacity is used for: more enquiries handled, faster turnaround, less overtime, or hiring that can be postponed. Hours that are saved but not redirected do not appear in any financial statement.
What costs are most often left out of the estimate?
The time staff spend reviewing AI output, integration with existing systems, model usage fees as volume grows, and ongoing maintenance. A business case that lists only the build cost will look better than the project turns out to be.
Related service
WhatsApp AI Chatbot
We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.
