AI ROI for Small Business: Do the Math First

How do you calculate the ROI of an AI project?

You calculate AI ROI by estimating the annual value — time saved, revenue gained, and cost or risk avoided — subtracting the full cost to build, run, and supervise it, then comparing the result against a baseline number you measured before you started. No baseline, no honest ROI.

The math is simple. The discipline is not. Most small businesses skip the part that matters — the baseline — and end up with a number they can’t defend. That’s how you land in MIT’s pile: Project NANDA found 95% of enterprise generative-AI pilots delivered no measurable return and concluded the cause was mismanaged adoption, not weak models (Fortune). The projects that pay off do the arithmetic before they spend.

Here is the whole formula in one box. Value on one side, total cost on the other, and a baseline underneath both so the comparison is real and not a story you tell yourself.

Notice what makes it honest: the baseline. ROI is always change against a starting point. If you don’t know what the workflow costs today — in hours, errors, or lost revenue — you have nothing to subtract from, and any percentage you quote afterward is invented. Measure first. Always.

How do you estimate the value of an AI project?

Estimate value across three buckets: time saved (hours removed × loaded hourly cost), revenue gained (faster response, more capacity, higher conversion), and cost, errors, or risk avoided. Use conservative numbers, count only value you’ll actually capture, and tie each bucket to something you can later measure.

Take the three buckets one at a time. Time saved is the easiest to over-claim, so be strict: count hours the AI genuinely removes, multiply by a loaded hourly cost (salary plus overhead, not the bare wage), and only credit time that gets reallocated to something valuable. Revenue gained comes from answering leads faster, handling more volume without hiring, or freeing your best people for higher-value work. Cost and risk avoided covers fewer errors, less rework, lower compliance exposure, and contractor spend you no longer need.

Make it concrete with an example — clearly hypothetical, not a cited figure. The card below shows how a single repetitive task turns into an annual number you can put next to a cost.

That €6,500-ish figure is a ceiling, not a promise. It only lands if those five hours actually get spent on revenue work instead of evaporating into longer coffee breaks. Which is exactly where most ROI estimates quietly lie — more on that below.

Time saved is only ROI if someone reallocates it. Otherwise it’s a slower lunch.

What costs do small businesses forget to count?

They count the build and forget the rest: ongoing usage and API fees, hosting, maintenance and updates, and — biggest of all — the human time to run, check, and correct the AI. Add the “failed-pilot tax” too: the projects you’ll abandon before one sticks. Real cost is build plus run plus supervise.

A custom build has a one-time price tag, and that number is the one everyone fixates on. But the meter keeps running after launch. The costs that wreck a rosy ROI estimate are almost always the recurring ones — and the human ones.

The supervision cost is the one small businesses underestimate most. An AI agent that drafts replies still needs someone to read and approve them; one that enters data still needs spot-checks. Budget that time honestly — it’s real payroll, and it doesn’t disappear just because the work moved to a model. We size it up front in every custom build we scope.

What’s a good ROI — and when is the number lying to you?

A good ROI clears your full cost inside a payback period you’d accept for any investment — often months for a focused workflow. The number lies when it counts time nobody reallocates, ignores recurring and supervision costs, or has no measured baseline. That’s vanity ROI: a big percentage attached to nothing real.

The cleanest way to judge a project is payback period: how many months of net value it takes to repay the total cost. A focused single-workflow build that pays back in a few months is a strong bet; a sprawling program that “will pay off eventually” rarely does. Short payback, tight scope, measured baseline — that’s the profile of a return you can trust.

Vanity ROI is the opposite. It’s the impressive multiple you’ll see in a vendor deck that quietly assumes every saved hour becomes billable, every cost is one-time, and the baseline was whatever makes the after-number look best. The most common form is saving time that nobody redeploys — a version of Solow’s old paradox, “you can see the computer age everywhere but in the productivity statistics.” We unpack that trap in The AI Productivity Paradox.

How do you de-risk the investment?

De-risk it in three moves: measure a real baseline before you touch any tool, pilot one narrow workflow with a clear success number instead of buying a platform, and fix the process first — because the cheapest ROI win is often a better workflow with no AI at all. Prove value small, then scale only what earns it.

The single highest-leverage step is also the dullest: measure the baseline. Spend a week counting how long the target task really takes, how often it goes wrong, and what that costs. Now your “after” number has something to beat, and your ROI is arithmetic instead of a guess.

Then pilot one workflow, time-boxed, with a single success metric — not a company-wide rollout. And before you spend on any model at all, ask the cheapest question in the building: could a better process fix this with no AI? McKinsey found AI high performers were roughly 3× more likely to redesign their processes than other firms (McKinsey). The return lives in the redesign. That’s our whole method — map, redesign, match, embed — and it’s why we sometimes tell clients they don’t need AI yet. If you’re buying from outside, the same rigor applies to the vendor: vet the consultant before you sign.

Want to pressure-test the math on a real workflow without committing a budget? You can chat through one workflow with Maggie and get a grounded read on whether the numbers work — before anything gets built.