Here is the first page of a pilot. We built it to show the method, so the company is not real, but the two gates on it are the part worth copying: the result that earns a second team, and the result that ends the project. Both were agreed with finance before a single invoice went near the agent.
The case below is constructed. It is not a client, not a past project and not a measurement. Every number in it was picked so the arithmetic is easy to follow, and you should replace them with your own. What is real is the method, and the point it makes about agentic AI ROI: a pilot is a decision with a deadline, and the return percentage comes second.
The workflow is invoice matching in accounts payable. The agent reads each supplier invoice, matches it to its purchase order, passes the clean ones and hands the mismatches to a person.
Week zero: the page
| Item | Written down |
|---|---|
| Workflow | Match supplier invoices to purchase orders; hand mismatches to the accounts payable lead |
| Baseline | 1,000 invoices a month. 6 minutes of handling each, plus 4% needing rework at 25 minutes. Effective cost: 7.0 minutes per invoice. Measured over a full cycle including a month end |
| Outcome metric and owner | Effective minutes per invoice, owned by the accounts payable lead and not by whoever built the agent |
| Gate to scale | At or below 4.5 minutes, with rework no higher than baseline |
| Gate to stop | At or above 6.0 minutes, or rework above 5%, after one permitted fix |
| In between | One fix, then re-read |
| What freed time becomes | Temporary month-end cover is not rehired |
| Pilot cost and run cost | Separate lines. Integration and data cleanup are one-off. Model usage and human review recur |
Two rows do the work. The baseline had to cover a month end, because a quiet fortnight would have flattered the agent or embarrassed it depending on the calendar. And the freed-time row exists because "saved 58 hours" is not a financial result unless something changes. If the hours vanish into the same working day, a finance director is right to count them as zero.
The effective-minutes figure also counts the errors, so the pilot cannot look better than production will.
Week four: between the gates
The first reading: the agent passes 45% of invoices cleanly and hands 55% to a person. Clean ones need a one-minute spot check. Handed-off ones take the usual 6 minutes. Rework runs at 6%.
That is 0.45 x 1 + 0.55 x 6 = 3.75 minutes, plus 6% x 25 = 1.5 minutes of rework: 5.25 minutes per invoice.
It is better than the 7.0 baseline and it passes neither gate. It is above 4.5 and below 6.0, so the rule from week zero applies: one fix, then re-read. Nobody argues about whether 5.25 is "promising".
The handoff rate points at the cause. More than half the cases going to a person means the agent is not recognizing the vendors. Duplicate supplier records in the master data are splitting one vendor's invoices across two names, and the agent cannot tell which is right. The fix is not a better prompt. It is merging the duplicates for the suppliers that account for most of the volume, a slice and not the whole database.
Week eight: the reading
After the cleanup the agent passes 65% cleanly and hands off 35%. Rework falls to 3%.
0.65 x 1 + 0.35 x 6 = 2.75 minutes, plus 3% x 25 = 0.75 minutes: 3.5 minutes per invoice, against a baseline of 7.0.
That clears the scale gate, and the rework is below baseline too. At 1,000 invoices a month, effective handling falls from about 117 hours to about 58. Those 58 hours count because week zero named their use: the temporary cover is not rehired at month end. Run cost, the model usage and the review time, goes on its own line and is compared with those hours, not blended with the one-off cleanup.
Now run it the other way. Suppose the cleanup had not helped and week eight had read 5.25 minutes with rework still at 6%. The one fix is spent and rework is above the 5% line, so the stop gate applies: the pilot ends, or restarts with a new design and new gates. The stop condition does its job in the bad quarter. The argument about it took place on calm terms, in week zero, with finance in the room.
The wider numbers explain why this is worth the discipline. McKinsey's State of AI research reports 23% of organizations scaling an agentic AI system somewhere in the business and another 39% experimenting, with no more than 10% of respondents scaling agents in any single function. Gartner's forecast is that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The middle item is the one a page like the one above addresses. A project cannot show value nobody defined, and cancellation arrives when a budget holder asks for evidence in month nine and is shown a demo.
The language of ROI measurement itself has little audience: "AI ROI measurement", "agentic commerce" and "AI agent observability" returned no measurable volume in Google Ads (DataForSEO, US, English, 11 October 2026). "AI in ERP" drew 1,000 searches a month at $21.44 a click in the same measurement. That is one tool and one country. We read it as buyers wanting AI where their numbers already live, and it has a practical consequence here: when the workflow runs in your ERP, the before and after both come from timestamped records in one place with the same definitions. When the agent works beside the system, the review argues about whose numbers are right instead of what they mean. For what happens when a pilot that read well underperforms live, see AI agent observability and the production gap.
If you have a pilot that has run for months with no way to prove or disprove it, write the page now, even late. Our approach begins with the baseline before any automation goes live, and the services team can help you set gates your finance team will trust. Get in touch and bring the workflow.
Sources: McKinsey, The State of AI | Gartner, 25 June 2025 | Search volumes measured via Google Ads (DataForSEO), US, English, 11 October 2026. The pilot figures are constructed for illustration.
