Measure AI ROI on one workflow. Time it before and after with the same method, price the hours at a loaded rate, count only results a person accepted, and put every running cost in the after column. Then report capacity and cash on separate lines, because payback can be months on one and years on the other.
That gives you numbers on a piece of paper the board can check. Ask 3 people what an AI project returned and you’ll get 3 figures. The vendor counts every task the system touched, the team counts the hours they think they got back, and finance can’t find a saving on the ledger.
The method fits operations and finance work, and the supplier invoice example can go straight into a spreadsheet.
Why AI ROI figures disagree
The gap between what AI teams report and what finance sees is wide. In PwC’s 2026 survey of 4,454 CEOs, 56% said AI had brought no significant financial benefit so far, and 12% said it had delivered both higher revenue and lower costs.
The UK government’s AI adoption survey for 2026 found 24% of businesses using AI had seen operating costs fall, and 64% had seen no immediate financial effect.
Those surveys don’t say why. In the projects we’ve measured, the gap comes from 4 habits:
- Hours estimated, never timed. The before figure is a guess in a workshop, and the after figure is another guess.
- Capacity counted as cash. Released hours get multiplied by a salary and called a saving, while payroll stays the same.
- Running costs left out. The build is in the business case, but the monthly model bill, hosting, maintenance and review time aren’t.
- Outputs counted instead of results. The system drafted 5,000 replies, and nobody checked how many were sent unchanged.
Each habit inflates the number in a different way, and each step of the method closes one.
Start with a baseline you timed
A baseline is the current cost of one workflow, measured over a period long enough to include its peaks. Pick a workflow with a clear unit: an invoice posted, a request routed, a client onboarded, a report delivered. Then measure 5 things.
| Measure | How to get it | Why it matters |
|---|---|---|
| Volume per month | Count from the system of record | Every other figure scales with it |
| Handling time per item | Time a sample of 30 to 50 items, or use system timestamps | The largest line in most business cases |
| Rework rate and time | Count items corrected, reversed or reopened | Errors cost twice: the fix and the original work |
| Waiting time | Timestamps from arrival to completion | Shows delays the hours figure hides |
| Who does it | Roles and loaded rates | Turns hours into money |
Time the work. Don’t ask people to estimate it. Self-reports drift, even from careful people. In METR’s randomized study of 16 experienced developers on 246 real tasks, they expected AI to speed them up by 24%. Afterwards they believed it had sped them up by 20%. Measured, it had made them 19% slower.
System timestamps are the cheapest reliable source. When the system doesn’t record a start time, sit with the people doing the work and time a sample. 4 weeks is the minimum, so the baseline includes a month-end close, a payroll run or whatever peak the work has.
Write down the method next to the number. The after measurement has to use the same one, or the comparison means nothing.
Turn hours into money
Price an hour at the loaded rate: salary plus employer taxes, pension, benefits and a share of overhead. Finance usually has this figure. If they don’t, choose an assumption, such as salary plus 30%, and print it next to every number that uses it.
Keep the people separate when their rates differ. A finance manager approving exceptions costs more per hour than a clerk entering lines, and the automation usually shifts work between them. A business case that moves 50 clerk hours into 10 manager hours has a different value from one that removes 60 clerk hours.
Cost per accepted result
Divide the total cost of a period by the number of results a person accepted in that period. That’s the unit you compare before and after.
The acceptance rule decides the number, so write it before you start counting:
- An invoice counts when it’s posted and not reversed within 30 days.
- A support reply counts when it’s sent and the customer doesn’t come back about the same issue within 7 days.
- A drafted report counts when the owner signs it off without rewriting a section.
Outputs the automation produced but a person rejected still cost money, and they stay in the numerator. A draft that was thrown away and rewritten from scratch is the expensive case: you paid for the draft and for the person.
For support automation, we found the gap between counted and accepted results was large. A vendor’s dashboard showed 53% of conversations resolved, and grading every conversation against agreed rules gave 36% verified. Our guide to automated resolution rates covers that measurement in detail.
Verified outcomes
Somebody has to check that the accepted results were right. Approval in the moment catches some errors and misses others, so add a sample check after the fact.
- Draw a random sample of completed items each week. 20 to 50 is usually enough to see a trend.
- Have someone who didn’t approve them check each one against the source documents.
- Record each error with its cause: wrong reading, wrong match, wrong rule, or a missing rule.
- Count the time the check takes. It’s a running cost of the automation.
The sample gives you an error rate, and the error rate gives you a rework cost. Both go in the after column. The causes tell the team what to fix next.
What the automation costs to run
The build is a one-off cost. Everything else repeats, and it goes in the monthly after column so the comparison is like for like.
| Cost line | Typical content | Often missed? |
|---|---|---|
| Build | Fixed price or days of work, plus your team’s time in workshops and testing | Your team’s time, often |
| Model usage | Per-token or per-request charges from the AI provider | Rarely missed, often overestimated |
| Hosting and monitoring | Servers, logs, alerts, storage | Often |
| Maintenance | Fixes, supplier format changes, model updates | Often |
| Licences | Platform seats, connectors, extra modules | Sometimes |
| Review and approvals | People time spent on what the automation prepares | Very often |
| Rework | Correcting the automation’s mistakes, found later | Very often |
Model usage is usually the smallest of these for back-office work. Our article on AI cost management shows a case where review time was 57% of the cost of each accepted result and provider usage was 17%.
Payback and ROI
2 formulas cover most of what a finance director will ask for.
Payback period is the build cost divided by the monthly net benefit. It answers how long until the project has returned what it cost.
ROI over a period is the total net benefit over the period, minus the build cost, divided by the build cost. Pick the period before you calculate it. 12 months after go-live is a common choice for operations work.
Both formulas depend on what you call a benefit. A benefit is cash, avoided cost or capacity, and each needs different evidence.
- Cash. A cost line went down: a temp contract ended, overtime fell, a supplier fee was cancelled. Finance can see it on the ledger.
- Avoided cost. A cost that was in the approved plan is no longer needed. The usual case is a hire budgeted for volume growth. It counts only if the plan existed before the project.
- Capacity. Hours were released, and the same people are still paid. It’s real value if the time goes to something useful, and it belongs on its own line.
A worked example
The workflow is supplier invoices at a company with an accounts payable team of 3: reading each invoice, matching it to the purchase order and delivery note, entering it and chasing anything that doesn’t match.
Baseline, timed over 4 weeks
| Line | Calculation | Monthly figure |
|---|---|---|
| Invoices | From the finance system | 2,000 |
| Handling | 2,000 at 6 minutes | 200 hours |
| Rework | 4% corrected, 80 at 15 minutes | 20 hours |
| Total hours | 220 hours | |
| People cost | 220 hours at a loaded $40 | $8,800 |
| Cost per invoice | $8,800 divided by 2,000 | $4.40 |
40 of the 220 hours were done by a month-end temp, billed at the same $40 an hour, or $1,600 a month.
Month 3 after go-live, timed the same way
The automation reads each invoice, matches it to the order and delivery, and posts it for one-click approval when everything agrees within the tolerance the team set. Anything else goes to a person with the difference shown.
| Line | Calculation | Monthly figure |
|---|---|---|
| Approvals, straight-through invoices | 1,600 at 45 seconds | 20 hours |
| Exceptions handled by a person | 400 at 12 minutes | 80 hours |
| Weekly sample checks | 2 hours a week | 8 hours |
| Rework from automation errors | 1% of 1,600, 16 at 30 minutes | 8 hours |
| Total hours | 116 hours | |
| People cost | 116 hours at $40 | $4,640 |
| Model usage, hosting, maintenance | Monthly agreement and bills | $1,200 |
| Total monthly cost | $5,840 | |
| Cost per invoice | $5,840 divided by 2,000 | $2.92 |
Exceptions take longer per item than the old average, 12 minutes against 6. The easy invoices no longer reach a person, so that’s expected. 104 hours a month are released.
The build was a fixed $24,000, plus 60 hours of the finance team’s time in mapping and testing at $40, so $26,400 in total.
The automation is identical in all 3 lines. What changes is whether released hours are valued as capacity, whether a cost line went down, and whether a hire in the approved plan was no longer needed.
Payback on capacity, cash and avoided cost
| View | Monthly net benefit | Payback |
|---|---|---|
| Capacity: 104 hours at $40, minus $1,200 running cost | $2,960 | 8.9 months |
| Cash: temp contract ended ($1,600), minus $1,200 running cost | $400 | 66 months |
| Cash plus avoided cost: as above, plus a half-time hire budgeted for next year’s 25% volume growth ($3,200) | $3,600 | 7.3 months |
On the capacity view, and assuming the month 3 figures hold for the whole year, first-year ROI is 34.5%. 12 months at $2,960 is $35,520, less the $26,400 build, divided by $26,400.
On cash alone, the project takes 5.5 years to pay back. Both numbers are correct. The decision depends on what the team does with 104 hours, and whether the growth plan holds.
In this example, the avoided hire is the largest line. It’s also the one finance will question first, so attach the approved budget to the business case. Without that document, report it as capacity.
What to leave out
These numbers make a business case look stronger and fall apart under review. Leave them out, or report them separately with a label.
- Time savings from a survey. Use surveys to choose what to time, and then time it.
- Hires that were never in the plan. “We would have needed two more people” counts only if someone approved the budget for them.
- Pilot volumes. A pilot on 200 clean invoices says little about 2,000 invoices from 140 suppliers.
- The vendor’s own dashboard. Use your acceptance rule and your sample checks. The vendor counts what its product touched.
- Soft benefits converted to money. Faster answers and fewer errors matter. Report them as faster answers and fewer errors.
- The first 2 weeks after launch. Volumes, exception rates and behaviour settle after a few weeks. Measure at steady state.
Leaving these out makes the number smaller, and it holds when finance checks it.
What to report each month
Once the workflow is live, the same figures go to the owner every month, against the baseline.
| Line | Baseline | This month | Last month |
|---|---|---|---|
| Volume | 2,000 | ||
| Cost per accepted result | $4.40 | ||
| Straight-through share | 0% | ||
| Error rate from sample checks | 4% | ||
| Hours released | 0 | ||
| Where released hours went | |||
| Cash lines changed | |||
| Running cost | $0 |
The row for where released hours went is the one leadership reads first. If the answer is “nowhere in particular”, the capacity view is weak, and the next decision is about redeploying time before building more.
Where to start
- Choose one workflow with steady volume, a clear unit of work and an owner who can sign off a result.
- Time the baseline for 4 weeks, and write the method down.
- Agree the acceptance rule and the sample check before the build starts.
- List every running cost the automation will add, including review time.
- After a month at full volume, measure again the same way and report all 3 views.
The AI savings calculator gives a first estimate for a role or department in a few minutes, which helps you choose what to time. When you need timed numbers for the business case, our 2-week AI readiness assessment produces the baseline, the acceptance rule and a fixed-price build estimate for one area. This guide to running one explains what it should contain.
Next step: See the 2-week AI readiness assessment