Picture a pilot that went well. For three weeks the team ran AI-assisted triage of regression failures beside the manual process, scored the results against a written rubric and found a unit cost clearly below the baseline. Then the budget holder asked one more question: in which month does this pay for itself?
Nobody had an answer. The pilot produced a cost per accepted outcome, a snapshot of what one usable piece of work costs under pilot conditions. A payback month needs things the snapshot does not contain: the cost of putting the workflow into production, the cost of keeping it there, the volume that will actually pass through it, and the weeks in which the team is slower before it is faster. It also needs a rule for the uncomfortable case in which the answer is “never”.
This article adds that time dimension. It assumes the workflow already has an acceptance rubric and has met its quality bar; if not, start with the pilot design.
How the question looks from each role
Finance or budget holder. Sees a setup invoice, new subscriptions and a usage line that moves with volume. The question is rarely hostile: the State of FinOps 2026 press release states that many organisations report being asked to self-fund AI investments through efficiency gains. The fear is an open-ended commitment with no test for failure. The decision is how much to fund, for how long, and what evidence would end the funding.
CTO or Head of Engineering. Sees a successful pilot and a team that will be slower for some weeks while it changes how it works. The fear is twofold: being held to a date derived from pilot conditions, and keeping a workflow alive because nobody defined when to end it. The decision is which workflows have enough volume to justify setup.
QA lead or test manager. Sees the review queue. Every generated test or proposed failure classification has to be checked by someone who could otherwise be testing. The fear is that review effort stays at its early level, the saving never appears and the QA team is asked to explain why. The decision is how much review each kind of output needs and what to record each month so the forecast can be corrected.
Why the usual answers fall short
Multiplying the pilot saving by annual volume. It leaves out setup, the monthly cost of keeping the workflow running and the slower first months, and it assumes production will behave like a pilot run on a chosen sample by attentive reviewers.
Borrowing an industry figure. The World Quality Report 2025, a survey of more than 2,000 senior executives, found 89% of responding organisations piloting or deploying generative AI in quality engineering and only 15% doing so at enterprise scale. The average reported productivity boost was 19%, and one third had seen minimal gains. An average with that much spread beneath it cannot stand in for one workflow in one team. Even careful models carry a warning. According to InfoQ’s summary of DORA’s report on the ROI of AI-assisted software development, the worked example covers a 500-person engineering organisation and reaches payback in around eight months, and the authors ask readers to treat the calculations as “a high-uncertainty estimate meant to spark a conversation, rather than a rigid mathematical formula”.
Leaving out the first months. DORA introduces that report by referring to the initial “productivity dip” of a new rollout. InfoQ’s summary describes a J-curve, in which a temporary dip comes before longer-term gains, and names three causes: the learning curve, the “verification tax” of reviewing AI-generated code, and the need to adapt downstream processes such as testing and change approval. In a single QA workflow, we would expect the same dip to appear as a unit cost above the pilot figure for a while, and sometimes above the old baseline.
Treating the decision as one-way. Once setup is paid for, stopping feels like admitting a mistake. Without a rule written beforehand, low-volume and review-heavy workflows can carry on by default.
A method for the payback month
Build one spreadsheet, the payback sheet, with a row for each month. It takes six steps.
1. Fix the unit and the two unit costs
Use the same accepted outcome, rubric and accounting boundary as the pilot. Record the baseline cost per accepted outcome and the AI-assisted cost once the workflow has settled. The pilot’s unit cost is the starting estimate of that settled cost, to be replaced by measurement. The difference is the saving per accepted outcome. If it is zero or negative, there is nothing to pay back with and the exercise ends here.
2. Separate three kinds of cost
| Cost type | How it behaves | Typical lines |
|---|---|---|
| One-off setup | Paid once, before any saving | Integration with CI and the issue tracker, evaluation set, security review, training, parallel running |
| Recurring fixed | Paid every month whatever the volume | Licences or seats, monitoring, upkeep of prompts and context, a reserve for re-evaluation |
| Variable | Moves with volume and is already inside the unit cost | Model usage, review, correction, failed attempts |
Mark each line as cash or capacity; the distinction matters later.
The re-evaluation reserve deserves its own line. Providers retire models on published terms: OpenAI’s deprecation policy sets a minimum of six months’ notice for generally available models, and Anthropic’s policy gives at least 60 days’ notice for publicly released models and states that requests to retired models will fail. Each replacement means rerunning the evaluation set, comparing results and adjusting prompts. Estimate the cost of one such exercise, multiply it by the number you expect in a year and divide by twelve. The notice period sets only the deadline; take the number from the retirement dates published for your models, plus any planned upgrades.
3. Calculate the break-even volume
Break-even volume per month = recurring fixed cost per month / saving per accepted outcome.
Below this volume the workflow loses money every month and setup is never recovered. Compare it with the volume the workflow has actually handled in recent months, not the volume expected once adoption grows.
4. Put the dip on the sheet
Agree a ramp window, for example three months, in which the unit cost is expected to sit above its settled level. Reviewers check everything while trust is being established, prompts and context are still being tuned and the acceptance rate is lower. Estimate a unit cost for each month of the window. A workable basis is the pilot’s review minutes per outcome with every output checked, and its early acceptance rate. Replace each estimate with the measured figure as the month closes.
5. Find the payback month and the peak deficit
Monthly contribution = (saving per accepted outcome × accepted outcomes) − recurring fixed cost.
The cumulative position starts at minus the setup cost and adds each month’s contribution. The payback month is the first month in which it is no longer negative. Its lowest point is the peak deficit: the amount actually at risk, which exceeds the setup cost whenever the first months run at a loss.
An illustrative calculation. The figures below are hypothetical. They are not client results, benchmarks or a promised return. Suppose failure triage costs €9.00 per accepted outcome today and the pilot puts the settled AI-assisted cost at €5.00, a saving of €4.00. Volume is 500 accepted outcomes a month. Setup costs €12,000. Recurring fixed costs are €800 a month: €500 for licences and upkeep, plus a €300 reserve for an assumed two re-evaluations a year at €1,800 each. During a three-month ramp window the estimated unit cost is €10.00, then €8.00, then €6.00. The break-even volume is 800 / 4 = 200 accepted outcomes a month.
| Month | Unit cost | Saving per outcome | Monthly contribution | Cumulative position |
|---|---|---|---|---|
| Setup | – | – | – | −€12,000 |
| 1 | €10.00 | −€1.00 | −€1,300 | −€13,300 |
| 2 | €8.00 | €1.00 | −€300 | −€13,600 |
| 3 | €6.00 | €3.00 | €700 | −€12,900 |
| 4 | €5.00 | €4.00 | €1,200 | −€11,700 |
| 5 to 13 | €5.00 | €4.00 | €1,200 each month | −€900 by month 13 |
| 14 | €5.00 | €4.00 | €1,200 | €300 |
The same workflow gives three different answers. Dividing setup by the pilot saving alone, €2,000 a month, suggests six months. Including recurring fixed costs gives ten. Including the dip gives month 14, with a peak deficit of €13,600 in month 2.
6. Test the assumptions that move the answer
Change one assumption at a time. The table uses the same illustrative figures at the settled unit cost, before the dip.
| Scenario | Outcomes per month | Saving per outcome | Monthly contribution | Months to recover setup |
|---|---|---|---|---|
| As planned | 500 | €4.00 | €1,200 | 10 |
| Lower volume | 300 | €4.00 | €400 | 30 |
| Review stays heavy | 500 | €2.00 | €200 | 60 |
| Four re-evaluations a year | 500 | €4.00 | €900 | 13.3 (month 14) |
| Below break-even | 150 | €4.00 | −€200 | Never |
In this illustration a 40% fall in volume triples the payback period and a halved saving multiplies it by six, because fixed costs take a larger share of a smaller gross saving. Measure volume and review effort most carefully.
A financial stop rule, agreed before go-live
The stop rule says what happens when measurement disagrees with the forecast. Before go-live, the budget holder and the workflow owner agree three things in writing and review the sheet monthly, replacing forecast values with measured ones.
- A ramp window and an exposure cap. For the illustration above: three months and a peak deficit of no more than €14,500 (the forecast €13,600 plus an agreed €900 margin). Crossing the cap triggers a pause and a review.
- A horizon. The latest acceptable payback month, tied to something real such as the budget cycle or the period the workflow is expected to run without a major change.
- Two separate tests. Whether to continue is a forward-looking question: setup is spent whatever is decided, so what matters is whether the coming months’ contribution is positive. Whether to invest more depends on payback against the horizon.
| Measured after the ramp window | Decision |
|---|---|
| Contribution positive, payback inside the horizon | Continue; extending to similar work can be considered. |
| Contribution positive, payback beyond the horizon | Keep running, because each month still recovers part of the setup. No further investment. |
| Contribution negative for, say, two consecutive months, with a specific, fixable cause | Revise once, with a dated target. |
| Contribution negative, with volume below break-even or review effort not falling | Stop, return to the baseline process and record the result. |
Quality stop conditions apply whatever the financial position. A stop is also a result: keep the sheet, its assumptions and the measured values so the question can be reopened when volume, prices or models change.
Where this model stops being useful
It prices cost and nothing else. A workflow adopted for faster feedback or wider coverage may be worth running at a loss on this sheet. If so, say so explicitly and name the benefit and who values it, instead of adjusting assumptions until the sheet breaks even.
Capacity is not cash. The saving per outcome is usually people’s time. Unless contractor hours, overtime or other external spend actually fall, the cash position may never reach zero even when the capacity calculation does. Show both payback rows and agree which one the funding was approved against.
Small volumes and a moving baseline blur the result. With a few dozen outcomes a month, one difficult batch moves the unit cost noticeably, so use a rolling three-month figure. If the manual process improves or the suite shrinks, re-measure the baseline as well.
It does not price errors. A misclassified failure that hides a product defect costs more than its review time. That risk belongs to the quality bar.
A checklist before you give Finance a month
- The accepted outcome and rubric are the same for the pilot, the baseline and the payback sheet.
- Setup, recurring fixed and variable costs are separate, each labelled as measured, allocated or estimated and as cash or capacity.
- Recurring fixed costs include a re-evaluation reserve based on published retirement dates and planned upgrades.
- The break-even volume has been compared with actual recent volume.
- The sheet shows the ramp window month by month and the peak deficit.
- The payback month is given for at least three scenarios: as planned, lower volume and heavier review.
- A horizon, an exposure cap and a stop rule that separates “continue” from “invest more” are agreed in writing.
A payback month is a forecast that measurement is allowed to correct, and sometimes the honest outcome is to stop. Our QA & AI Value Engineering service starts with one recurring workflow that has enough volume to measure. It establishes a baseline with costs labelled as measured, allocated or estimated, compares realistic alternatives and leaves a before-and-after record your finance team can challenge.