The ROI of AI discovery sprint: building a business case that survives review

The return on an AI discovery sprint is mostly avoided cost. For $3,250 or ₹2,00,000 you either confirm a build worth $26,500 and upwards, or you cancel it in week two. On a single avoided wrong build the ratio is roughly eight to one, and it lands immediately rather than after a year of waiting for adoption.

That is the headline. A business case has to survive a finance review, so this article sets out the three return lines that hold up under questioning, a worked model with its assumptions labelled, the costs that get left out, and the situations where the ROI case honestly does not stand.

Why the standard ROI formula misfires here

Benefit divided by cost works when the benefit is a flow. A discovery sprint does not produce a flow. It produces information, and the value of information is the change it causes in a later decision. Model it as a revenue line and a reviewer will ask which revenue, and you will have no answer.

Decision theory has a better name for it: the expected value of information. The sprint is worth what it changes. If the recommendation is identical to what you would have done anyway, the return is zero no matter how good the document is. If it reorders the roadmap, kills a build or halves an estimate, the return is the size of that change.

This is also why sprint ROI is asymmetric. A sprint that confirms your plan returns confidence, which has modest value. A sprint that stops a wrong build returns the entire build budget, which is large. You are buying an option on being wrong, and most roadmaps are wrong somewhere.

There is a second-order return that reviewers accept once it is named. A discovery sprint produces an evaluation set and a costed architecture that any vendor can be held to, including an in-house team. That artefact travels. It makes the next quote comparable, it makes the first release gateable, and it survives a change of supplier, which a slide deck does not.

The three return lines a reviewer will accept

Avoided build cost

The largest and most defensible line. If discovery removes one candidate from the roadmap, you avoid the build price, the integration work around it, and the run cost for as long as it would have lived. An AI-accelerated MVP starts at $26,500 or ₹17,60,000, and a multi-agent system starts at $24,500 or ₹16,00,000, so a single correct cancellation pays for several sprints.

Compressed decision time

Most organisations do this work anyway, slowly, through committee. Six to twelve weeks of part-time internal debate consumes senior salary and delays whatever is right. Ten working days to a written recommendation converts that into a small, visible cost. Price it at the loaded cost of the people currently in those meetings, and it is usually the second largest line.

De-risked estimates

An estimate built after someone opened the real tables is materially tighter than one built from a schema diagram. Tighter estimates mean less contingency held against the programme, and less rework when data quality turns out to be worse than assumed. Claim this conservatively; a reviewer will believe a modest number and challenge a heroic one.

A worked model, with the assumptions labelled

The figures below are illustrative arithmetic, not measured results, and the point is the structure rather than the numbers. Replace each assumption with yours before the paper goes anywhere near a committee.

Read the table as two scenarios rather than one sum. In the cancel scenario, the return is the avoided build and its run cost, and payback is immediate. In the proceed scenario, the return is the compressed decision time plus the released contingency, and the sprint fee is credited against the build, so the net cash cost is your internal time only.

One discipline keeps the model honest: put the cancel scenario first. Papers that lead with the upside of proceeding read as advocacy, and reviewers discount them. A paper that opens by stating what the organisation avoids if the answer is no reads as analysis, and the same numbers land differently.

What is the payback period on an AI discovery sprint?

Immediate in the cancel scenario, and inside the first build in the proceed scenario. There is no adoption curve to wait for, because the value is realised at the moment the decision is made. That is unusual among AI investments and it is the strongest argument in the paper: this is the one line item whose return does not depend on users changing their behaviour.

Compare it with the build itself, where payback genuinely does depend on adoption, deflection rates and handling time. Our AI agent ROI calculator models that second question, and the two should be presented separately rather than blended into one number.

The costs people leave out

A business case loses credibility through omissions more often than through optimism. Five costs belong in the paper.

• Internal time, priced properly. A data owner at one day a week for two weeks, plus four to six interviewees for two hours each, is a real number. Leaving it out invites the obvious question.
• Access and extraction work. If the data cannot be read without building an export path, that engineering is separate and is often the largest surprise in the whole exercise.
• The opportunity cost of the ten days. Those people are not doing something else. Name what slips.
• Follow-on run cost on whatever you do build. A care plan starts at $1,000 or ₹68,000 per month, with a $750 or ₹40,000 AI add-on for evals and cost monitoring, as total cost of ownership for AI systems sets out.
• Model usage during and after the sprint. You pay providers through your own accounts, so it never appears on our invoice and always appears on yours.

When the ROI case does not stand up

Three situations, and pretending otherwise damages the next business case you bring.

First, when the build it precedes is small. If the thing you are considering costs $7,000, spending $3,250 to decide about it is disproportionate. Run a fortnight's trial instead and accept the risk of being wrong.

Second, when the decision is already made. If the board has committed and the budget is approved, a sprint that concludes differently will be overruled, and the return is zero by construction. Say so and move the money to the build, or be honest that what you want is a second opinion rather than a decision.

Third, when nobody will act on a negative result. The entire asymmetry of the return depends on your organisation being willing to cancel. If cancellation is politically impossible, you are buying confirmation and should price it as such. This failure pattern is the subject of why AI pilots never reach production.

What this looked like on a real programme

A university approached us to replace a fifteen-year-old ERP. Discovery found the core database and business logic were sound and the pain was at the edges: interfaces, integrations and reporting. The recommendation was to modernise around the system rather than replace it, and the university ERP modernisation case study describes what followed.

The return there was not a percentage. It was a replacement programme removed from the roadmap and a phased sequence the finance committee could approve one stage at a time. No adoption assumption was needed to justify the discovery spend, which is exactly the property that makes these cases easy to defend.

The wider lesson is that the discovery fee bought a sequencing decision rather than a technical one, and sequencing decisions are where most of the money in a programme actually sits. Choosing the right third phase saves more than optimising the first one.

Presenting it: a checklist for the paper

• State the decision the sprint unblocks, in one sentence, on the first page
• Show cancel and proceed as two scenarios, never as a blended expected value
• Price your own team's time and put it in the cost column
• Label every assumption as measured, estimated or illustrative
• Name the build the sprint is deciding about, with its published entry price
• Say explicitly that the fee is credited against that build if you proceed
• Include the first-year run cost of the thing you might build, not just its build price
• Name who will cancel, if the answer is no

Related reading

How to rank AI use cases by ROI, not excitement gives the scoring method that turns a shortlist into a sequence, and AI discovery sprint pricing sets out what the ten days include, with every starting figure on the pricing page. The underlying idea, that the cheapest experiment is the one that tells you to stop, is the validated learning principle set out by The Lean Startup, and it applies to AI roadmaps more sharply than to most software.

Build the case on avoided cost, show the cancel scenario first, and the sprint will survive a review that a benefits-led paper would not.