“The AI saves us a lot of time.” How much? Compared to what? And what did it cost?
Most AI projects can’t answer these questions, and that hurts them twice. Good projects don’t get the budget to expand, because nobody can show they work. Weak ones keep running, because nobody can show they don’t.
Measuring the return on investment (ROI) of an AI project isn’t harder than for any other change to how work gets done. You need a baseline before you start, an honest count of all the costs, and a clear answer to one awkward question: will the time you save actually be used? This post walks through each step and ends with a fully worked example.
The process in five steps
Measuring an AI project
- Baseline Measure the work as it is today
- Pilot Same work, with AI and a person checking
- Measure Time, quality, usage, cost
- Decide Roll out, fix or stop
- Keep tracking Monthly, after launch
Step 1: measure the baseline before you build
Once the new tool is in place, nobody remembers how long the old way took. So measure first. For the process you’re changing, record:
- Volume: cases per month (emails, invoices, tickets, orders). Usually it’s already in a system.
- Time per case: minutes of hands-on work. Time 30–50 real cases, or use timestamps from your helpdesk or ERP. Don’t rely on guesses; they tend to be off in both directions.
- Error rate: the share of cases that need correcting later, and what a correction costs.
- Cycle time: how long a case waits from arrival to done. Your customer cares about this more than about your minutes.
- Cost of an hour: salary plus employer contributions and overhead. Your finance team can give you the figure.
A week of tally sheets is often enough. For a support team, a month of helpdesk data is better, because volumes swing from week to week.
Step 2: decide what counts as a benefit
- Time saved per case, after review. If a case took 6 minutes and now takes 2 minutes including the check, you saved 4 minutes, not 6.
- Throughput. More cases handled by the same people. This matters if you have a backlog or you’re growing.
- Cycle time. Quotes out the same day instead of in three days. If you claim this wins business, compare won deals before and after.
- Quality. Fewer errors, fewer corrections, fewer complaints.
- Revenue effects. Faster replies to enquiries, fewer missed leads. They’re real but the hardest to attribute, so count them only with evidence.
Gains are rarely even across a team. In a well-known study of more than 5,000 customer support agents, an AI assistant raised the number of issues resolved per hour by 15% on average. Newer and less-skilled agents gained the most, while the most experienced saw little effect (Brynjolfsson, Li and Raymond, “Generative AI at Work”, Quarterly Journal of Economics, 2025). Look at the numbers by group, not just the average.
Leading and lagging metrics
Lagging metrics, such as cost per case or cycle time, tell you whether the project paid off, but they take months to move. Leading metrics tell you within weeks whether it’s on track.
What to watch, and when
| Leading: check weekly | Lagging: check monthly | |
|---|---|---|
| Usage | Share of cases going through the new process | Cases handled per person |
| Quality | Share of drafts the reviewer changes | Errors found later, complaints |
| Speed | Review time per case | Cycle time from arrival to done |
| Money | Model cost per case | Cost per case, overtime, temp staff |
If usage drops in week three, you’ll see it in the leading numbers long before the quarterly ones.
Step 3: count all the costs
- Build (one-off): the vendor or developer cost, plus your own people’s time in workshops, testing and training.
- Run (monthly): model usage, hosting, software licences. Model usage is billed per token, a chunk of text a little shorter than a word on average; tokens and context windows explains how it’s counted. For text workflows it’s usually small. Voice and agents that take many steps cost more.
- Maintain (monthly): updating instructions when inputs change, retesting when a provider updates a model, monitoring, support.
- Review time: people checking the AI’s output. It belongs in the “after” time per case. Leaving it out is the most common way to inflate a business case.
What an AI project really costs goes through each bucket in more detail.
A worked example
This example is illustrative. The company and numbers are made up to show the method; replace every assumption with your own.
A wholesale distributor of building materials receives customer orders by email. Some are in the email text, some in PDF or Excel attachments, in Slovak, Czech and English. Five people in the sales office read each one and type it into the ERP.
They build an AI workflow. The model reads the email and attachments and fills in a draft order in the ERP. Rules check prices and stock. A person checks the draft and confirms it. Orders the AI isn’t sure about are flagged and handled by hand, as before.
Assumptions:
- 2,000 order emails a month.
- Today: 6 minutes per order, timed on a sample of 50.
- After: 90% of orders arrive pre-filled and take 1.5 minutes to check. The other 10% are flagged and still take 6 minutes. That averages about 2 minutes per order, review included.
- An hour of office time costs €25, including employer contributions and overhead.
- Build: €12,000 fixed price, connected to the mailbox and the ERP, with a review screen.
- Internal time: 40 hours of workshops and testing plus 10 hours of training, so 50 hours × €25 = €1,250.
- Running: model usage of about €0.05 per order (€100 a month) plus €100 a month for hosting and monitoring.
- Support and maintenance: €300 a month.
Benefit. 4 minutes saved × 2,000 orders = 8,000 minutes, about 133 hours a month. At €25 an hour, that’s about €3,333 a month.
Costs. One-off: €12,000 + €1,250 = €13,250. Monthly: €100 + €100 + €300 = €500.
Net monthly benefit: €3,333 − €500 = €2,833.
Payback: €13,250 ÷ €2,833 ≈ 4.7 months.
First-year ROI. Benefits: 12 × €3,333 = €40,000. Costs: €13,250 + 12 × €500 = €19,250. ROI = (€40,000 − €19,250) ÷ €19,250 ≈ 108%.
The worked example
- 133 h saved per month
- €2,833 net benefit per month
- 4.7 months to pay back
- 108% return in the first year
Look at where the money goes in the first year:
First-year costs
Model usage is about 6% of the total. In text workflows like this one, it’s rarely the big number. The build and people’s time are.
The reality check: is the saved time used?
133 hours a month is a bit under one full-time job, assuming about 160 working hours a month. Nobody will be let go over that, and it shouldn’t be the goal anyway. So where does the €3,333 a month actually show up? Only in one of these ways:
- the same team handles more orders as the business grows, without hiring
- less overtime or fewer temporary staff in the busy season
- people spend the time on work that earns money, such as calling back customers about open quotes
If none of that happens, the saving exists only on paper. Run the same numbers assuming that only half of the freed time turns into useful work:
- Benefit: €1,667 a month. Net: €1,167 a month.
- Payback: €13,250 ÷ €1,167 ≈ 11.4 months.
- First-year ROI: (€20,000 − €19,250) ÷ €19,250 ≈ 4%.
Same system, same costs. Payback in under five months or in almost a year, depending on a management decision, not on the AI. Decide before launch what the freed time is for, and write it into the business case.
Two things the example leaves out on purpose. Fewer typing errors would add to the benefit, but count that only if you measured the error rate before. And in the first month or two, fewer orders go through the new process while people get used to it, which pushes payback out a little.
Pitfalls
- Counting time saved that isn’t reused. See the reality check above.
- Leaving out review time. If the check takes 2 minutes, you saved 4 minutes, not 6.
- Comparing to memory. “It used to take ages” isn’t a baseline.
- Measuring only the easy cases. The pilot handled the clean emails, while the messy ones still go the old way. Measure all cases.
- Counting licences as adoption. Fifty seats bought doesn’t mean fifty people changed how they work. Measure usage.
- Forgetting running costs. The build is paid once. Support, hosting and model usage are paid every month.
- Giving AI credit for everything. If volume dropped because of the season, hours dropped too. Compare like with like, or keep one team without the tool for a few weeks as a comparison.
Checklist
An ROI case you can defend
- Baseline measured, not guessed Volume, minutes per case, error rate, cycle time
- Cost of an hour from finance Salary plus contributions and overhead
- After-time includes review And the cases still handled by hand
- All costs counted Build, internal time, running, maintenance
- A plan for the freed time Growth, less overtime, or other work
- Leading metrics weekly Usage, edits, review time
- Lagging metrics monthly Cost per case, cycle time, errors
- A pessimistic scenario Half the benefit: does it still pay?
What to do next
Pick the process you’re considering and measure its baseline this month, before anyone builds anything. If you haven’t chosen one yet, how to find the first AI use cases worth building helps. Then put the numbers into the same simple model: benefit per month, one-off and monthly costs, payback, and a pessimistic case. When the pilot is done, from AI pilot to production covers what comes next. Our guide to bringing generative AI into your company puts all the steps in order.
We agree a fixed price after a workshop where we map the process with the people who run it. The build cost in your model is then a known number, not a guess. If you’d like to run the numbers on your own process, get in touch.
