Why your AI pilot is not paying back, and how to tell if it will

AI pays back when it is attached to a process with a measured baseline and a named owner. Most pilots have neither. They prove the model works on a clean sample, then stall because nobody can say how much time, error or revenue the process cost before, who will run it differently afterwards, or what the data preparation will cost to do properly.

The model is rarely the problem. The arithmetic around it usually is, and it can be checked before you spend another rupee, pound or dollar.

The same four patterns come up again and again, whatever the industry.

Nobody captured the baseline. The pilot shows the task now takes four minutes. Nobody measured how long it took before, how often it happened or how many errors it produced. Without that, the saving is an estimate, and finance discounts estimates.

The data preparation was never counted. The pilot ran on a sample someone cleaned by hand. Production runs on whatever the source systems hold that day: duplicate records, inconsistent fields, three years of unmapped activity. A model does not know the data is dirty. It learns the patterns it finds and repeats them at scale, so the cleaning becomes a project in its own right, with a cost nobody budgeted.

The value lands in someone else's budget. The cost sits with the technology team. The time saved shows up in customer service, the errors avoided in finance, the revenue lift in sales. Each team sees a small gain it did not ask for, and nobody adds them up, so the business case never closes.

Nobody owns it after the pilot. The enthusiast who ran the pilot moves on. The team that owns the process has a queue to clear and no reason to check the output. Usage drifts down, and the AI quietly becomes one more tool that people work around.

A pilot budget usually covers the licence and a few weeks of someone's time. The full cost of running AI in a process has two halves.

One-off costs

• Data preparation. Cleaning, de-duplicating and agreeing definitions for the data the use case needs. Often the largest single cost, and the one most often missing.
• Integration. Getting data in and results back into the systems people already use, so nobody has to copy and paste between screens.
• Change management. Training, revised procedures and the time it takes a team to trust the output.

Recurring costs

• Licences that scale with use. Per-seat, per-call or per-token pricing that looks small in a pilot and grows with adoption.
• Human review. The time people spend checking output, which never falls to zero for anything that matters.
• Governance. Keeping the use case in the AI use case register, monitoring quality and answering the questions customers, auditors and regulators ask. In Europe, some of this is now a legal duty: see the EU AI Act timeline.

The AI ROI calculator takes one use case at a time. You enter the baseline, what changes with AI, how confident you are in each kind of value, and the costs. It returns payback in months, net value over three years, the same figures at half and one and a half times the adoption you planned, and a recommendation, and emails you the results. Amounts can be in rupees, pounds or dollars.

A worked example. Illustrative only: the figures are invented to show the method, not taken from any business.

A customer service team handles 6,000 emails a month at five minutes each. AI drafts replies for 70% of them, cutting handling to two minutes, and misrouted emails fall from 4% to 1%. People cost ₹600 an hour and each error costs ₹400 to fix. The time saved is counted at 60% confidence and the errors avoided at 70%. No revenue lift is claimed.

The calculator applies three rules. They are deliberately strict, because a pilot that only pays back in the best case will not pay back.

• Scale when the use case pays back within 12 months at the adoption you planned, and still comes out ahead over three years at half that adoption.
• Fix when it pays back within 36 months at planned adoption but not at half. The case depends on people using it, so fix adoption, data or ownership before spending more.
• Stop when it does not pay back within 36 months even at planned adoption. Keep the lessons, record the decision in the register and put the money somewhere it works.

Stopping is a good outcome. A pilot that ends with a clear no has done its job, and it frees the budget for a use case that passes.