Will This AI Project Pay for Itself?

Every AI vendor will show you a return-on-investment calculator. Every one of them is designed to produce a good answer. Here is how to do the sum yourself, in a way that survives contact with your accountant.

Start with hours, not percentages

Ignore claims like "40% more productive". They are unfalsifiable in a small business. Count hours instead.

Take the specific task. How many people do it, how often, for how long? A quoting process that occupies two estimators for six hours a week is 624 hours a year. That is a number you can test a claim against.

Then price those hours at the fully loaded cost — salary plus employer's National Insurance, pension and overhead — not the headline salary. Most SMEs under-price their own time by around a third, which makes every business case look worse than it is.

Then ask what actually happens to the saved time

This is where most business cases quietly fall apart. If the tool saves your estimator four hours a week, one of three things happens:

They quote more work — real revenue, and the strongest case
They do other valuable work that was being neglected — real but harder to measure
They finish earlier and the time evaporates — no financial return at all

Be honest about which one applies. In a business that is quote-constrained, option one is genuinely transformative. In a business that is demand-constrained, the same tool saves time that has nowhere to go — and the money would be better spent on the demand problem.

Count the costs vendors leave out

The subscription is the smallest number in the calculation. The real costs:

Setup and integration — connecting it to your ERP, CRM or drawing store, usually the largest line
Data preparation — the part nobody quotes, because nobody knows how bad your data is until they look
Training and habit change — a fortnight of reduced output while people learn, which is a real cost even though it never appears on an invoice
The checking overhead — if a human reviews every output, that review time is permanent and must be subtracted from the saving
Ongoing ownership — someone maintains prompts, permissions and exceptions, forever

That last one catches people out. A tool that saves ten hours and costs two hours a week to look after saves eight, not ten.

The payback test

For an SME, a sensible threshold is twelve months. If a project cannot pay back inside a year on hours alone — before you count any soft benefits — it needs an unusually good strategic reason.

Twelve months is not arbitrary. It is roughly how long before the tooling shifts enough that you would reconsider the decision anyway. Anything claiming a three-year payback in a field moving this fast is making a forecast nobody can support.

Where the numbers understate the case

Two benefits are real but rarely quantified well, and it is worth naming them rather than inflating the hours to compensate:

Speed of response. If quoting drops from four days to one, you win work you previously lost on responsiveness alone. Your win rate is the place to look for evidence, not your timesheet.

Reduced key-person risk. When the knowledge of how to price a job lives in one person's head, that is a risk on your balance sheet whether or not it is written down. Anything that captures it has a value that exceeds the hours saved.

A decent business case fits on a page

Baseline hours and their loaded cost. What the saved time becomes. Total first-year cost including integration, training and checking. Payback period. One sentence on what happens if it does not work.

If it takes more than a page, the project is probably too broad to succeed.