Calculate AI ROI Before You Build

The forecast that matters is not what AI can do. It is what changes in your accounts when it does it. Most AI business cases fail at that translation: they count hours saved, multiply by a wage, and present the product as a return. The hours are usually real. The money frequently is not, because nothing downstream of the saved hour ever changed.

This is a method for forecasting the return before you commit to a build, written for a company between $10M and $50M in revenue looking at one workflow. It is deliberately conservative, and it will sometimes tell you the answer is zero. That is the point: a zero you find in an afternoon is the cheapest result on this page.

Why do saved hours not equal saved money?

Because an hour freed from one task is almost always absorbed by another task on the same day, and payroll does not change. A dispatcher who spends ninety fewer minutes on quote paperwork does not go home ninety minutes early and does not cost less. The work expands. That is not a failure of the tool — it is what happens by default unless something else was arranged.

This is the gap the whole industry papers over, and the enterprise research is quietly clear about it. CIO’s December 2025 feature quotes practitioners making the same point from several directions: Movadex’s Salome Mikadze insists on tracking “task completion without human rescue, because an unused model has zero ROI,” while Source86’s Agustina Branz puts it flatly — “AI can definitely make work faster, but faster doesn’t mean ROI.” Deloitte’s executives described the same difficulty from the measurement side: AI is almost never deployed alone, so its specific contribution is hard to isolate after the fact.

The practical consequence is that the conversion mechanism has to be chosen deliberately, in advance, by a person with the authority to make it happen. If nobody has decided what will change when the hours appear, nothing will change, and the forecast should say so.

How do freed hours actually turn into money?

Exactly three ways: spend stops, a hire is avoided, or demand that was already waiting gets served. Pick one per workflow. Never count two.

Spend stops. A contractor invoice ends, an overtime line falls, a software seat is cancelled, an outsourced function comes back in-house or goes away. This is the only one of the three that shows up in the accounts on its own, without anybody deciding anything further. It is the strongest evidence and the easiest to verify afterwards — you can point at the line item that disappeared.

A hire is avoided. You were going to add a person at the next growth step and now you are not, or not yet. This is legitimate, and it is also the one most often abused. The test is whether the hire was actually planned before the AI conversation started: is it in a budget, a headcount plan, or a conversation someone can date? A hire invented retrospectively to justify a build is not a saving, it is a story.

Waiting demand gets served. There is work customers wanted that you were turning away or delaying, and now capacity exists to do it. This is the only mechanism that grows the top line, and it carries a hard precondition — the demand has to be already confirmed, not forecast. Quotes you declined, a waitlist, a backlog with dates on it. “We could probably sell more” is not this.

The one-per-workflow rule is where most business cases quietly double. If the freed dispatcher hours avoid a hire, those same hours cannot also serve new demand — the person is doing one thing or the other. Counting both is how a $60,000 case becomes a $140,000 case on a spreadsheet and a disappointment in month nine.

How should new output be valued?

At gross profit, never at revenue. If the system lets you take twenty more jobs a month at $2,000 each, the value is not $40,000. It is $40,000 less what those twenty jobs cost you to deliver. At a 30% gross margin, the honest figure is $12,000, and the difference between those two numbers is whether the project looks transformative or marginal.

This is the most common single error we see, and it is rarely deliberate. Revenue is the number everyone in the business already tracks, it is bigger, and it makes the case easier to approve. But incremental volume consumes materials, labour, fuel and support in proportion, and none of that is affected by the automation that created the opportunity. Only the margin is genuinely new.

The same discipline applies to a saved hour: value it at the fully loaded cost of the person doing it — wage plus payroll taxes plus benefits plus the overhead that person carries — not at their hourly wage and not at what you bill for their time. Billing rate is what you would earn if the hour were sold, which is the “waiting demand” mechanism, and counting it here would be counting the same hour twice.

What does the whole thing actually cost?

The model bill is the small line. The Thinking Company’s 2026 AI ROI guide breaks total investment down as technology 25–35%, data preparation 30–40%, talent 20–30%, change management 10–15%, and governance 5–10%. A business case built on API pricing is understating the investment by roughly a factor of three.

Then there is the recurring side, which is where forecasts age badly. A production system needs monitoring, periodic evaluation against real cases, retraining or prompt revision when the inputs drift, and human review time for anything consequential. None of that is optional and none of it stops. Put a monthly number on it and subtract it from the monthly benefit before you calculate anything; a project that pays back on gross benefit and loses on net benefit is a project that will quietly be switched off.

For the arithmetic itself, the standard forms are fine and you do not need a proprietary one. CIO reports the benchmark most companies settle on as ROI = (Δ revenue + Δ gross margin + avoided cost) − TCO, with practitioners targeting payback under two quarters for operations use cases. The formula was never the hard part. The inputs are.

How long should it take to pay back?

Assess a single mid-market workflow on a twelve-to-eighteen-month payback, and treat anything longer as a signal to pick a different workflow. The market averages are much worse than that, and it is worth understanding why before you borrow them.

Deloitte’s 2025 survey of 1,854 senior executives across Europe and the Middle East found most organisations reach satisfactory ROI on a typical AI use case in two to four years — against the seven to twelve months normally expected of technology investment. Only 6% reported payback inside a year; 27% took one to two years, 37% two to three, 25% three to four. Those are enterprise programmes spanning many workflows, entangled with data platform work and reorganisations, which is precisely why they take that long and why the attribution is so hard.

A $20M company pointing one system at one workflow is not running that programme and should not accept that timeline. The advantage of small scope is that the counterfactual stays clean: you know what the process cost before, you know what changed, and no parallel reorganisation is muddying the number. Use that advantage. If a well-chosen workflow cannot show a return inside eighteen months on conservative inputs, the problem is usually the workflow, not the ambition.

One more adjustment: year one is not the annual number. Builds ramp — there is a period where the system is live but adoption is partial and someone is still checking its work. Forecast the first year at a fraction of steady state and say which fraction you assumed, out loud, in the business case.

When is the honest answer zero?

When no mechanism applies. If the hours are freed but no spend stops, no planned hire is avoided, and no confirmed demand is waiting, the forecast is zero. Not small — zero. The tool may still work perfectly and the team may prefer their day. There is simply no financial return, and the business case should say so rather than manufacture one.

We say this in scoping conversations more often than is commercially convenient, and it is the most useful thing we do. A zero found before a build costs an afternoon. The same zero found after costs the build, the integration, the training, and the credibility of the next AI proposal anyone brings to that leadership team — which is the expensive part, because it is usually the next proposal that was the good one.

There are two honest reasons to proceed anyway, and both should be stated as what they are rather than smuggled in as ROI. The first is capability: some builds exist to learn how to build, and their return is the second project. The second is quality or risk — fewer errors, faster response, a service level you can actually promise — which is real value that does not show up as a dollar figure and should be argued on its own terms. What neither justifies is a spreadsheet.

Where to go next

The forecast is only as good as the workflow you point it at. What an AI readiness assessment measures and how to measure AI readiness cover choosing that workflow and judging the data behind it, and what an AI maturity model is for explains where a single project sits in a longer sequence. If you are pricing outside help to do the build, what AI consulting services include and how to choose an AI consulting firm cover the menu and the tests. Our own scope and prices are on the pricing page, and the free AI Opportunity Assessment maps one workflow end to end and produces exactly the inputs this method needs. More in the AI ROI hub.

Sources

• Deloitte, AI ROI: The paradox of rising investment and elusive returns, 22 October 2025 — the two-to-four-year payback finding, the 6% figure, the full payback distribution, and the attribution difficulty. Survey of 1,854 senior executives across Europe and the Middle East conducted 15 August to 5 September 2025, with 24 follow-up interviews.
• CIO, AI ROI: How to measure the true value of AI, by Josh Fruhlinger, 16 December 2025 — the quoted practitioners, the risk-adjusted ROI formula, and the payback targets.
• The Thinking Company, What Is AI ROI?, by Bartek Pucek, last updated 11 March 2026 — the total-cost-of-ownership breakdown quoted above.
• MIT Sloan Management Review, Three Approaches to Measuring and Managing AI ROI, by Mika Ruokonen and Paavo Ritala, 23 June 2026 — that two companies making near-identical AI investments routinely define success in entirely different ways, from interviews with more than 30 CEOs and senior leaders.

Method note. The three conversion mechanisms, the one-per-workflow rule, the gross-profit rule and the zero-is-legitimate stance are our own forecasting method, applied in client assessments and published here as principles rather than as a proprietary model — the operational detail behind them, including how we grade input quality and cap assumed capacity, is not published. Third-party figures are quoted from the named sources with their dates and are not our measurements. Deloitte’s survey covers Europe and the Middle East and skews larger than the $10M–$50M companies this article is written for, which is why it is used to characterise enterprise timelines rather than to set a mid-market benchmark.