How to Measure AI ROI for a Small Business

To measure AI ROI for a small business, pick two or three tasks, record how long they take and how often they go wrong before you introduce any tool, run the tool for four weeks, then measure the same two numbers again. Multiply the time difference by a loaded hourly cost, subtract the subscription and the hours you spent setting it up, and be honest about whether the freed time turned into anything. That last step is where most ROI claims fall apart. This measures ROI after you have already committed to a tool; if you are still deciding whether a project is worth funding, how to estimate AI project ROI before you start covers the estimate you make first.

The hard part is not arithmetic. It is that almost nobody records the baseline, and a baseline reconstructed from memory after the fact is always flattering.

Why most AI ROI numbers are worthless

Three failures account for nearly all of them.

No baseline. "It used to take about two hours" is a number produced by someone who wants the tool to have worked. Real measured task times are routinely 40 percent off remembered ones, in both directions.

Hours counted as money automatically. If a salaried person saves six hours a month and those six hours become six more hours of the same job at a slower pace, you have bought comfort, not margin. Comfort is worth something, but it is not the number you should put in a spreadsheet.

Quality ignored. A task done in a third of the time with double the error rate is not a saving, it is a transfer of cost from the person doing it to the person fixing it, usually you.

The three numbers that measure AI ROI

Resist the urge to build a dashboard. Three numbers per task, tracked by hand, beat twenty tracked automatically and ignored.

A task done four times a month cannot repay a serious integration effort no matter how much faster it gets. Volume decides which tasks are worth measuring in the first place, which is the same filter that should decide which tasks to automate first.

Getting a baseline in one week

You do not need new software. You need one shared sheet with five columns: date, task, who, minutes, needed a second pass yes or no.

1. Pick two or three candidate tasks that happen at least weekly and have a clear start and end. Vague tasks like "admin" cannot be measured.

2. Tell the team what you are doing and why, plainly. People who think they are being timed for performance reasons produce bad data.

3. Log for two weeks with no tooling changes at all. One week works if volume is high.

4. Take the median, not the mean. One catastrophic instance will drag an average somewhere useless.

5. Write the baseline down somewhere you cannot quietly edit later.

Two weeks feels slow when you are keen to start. It is the difference between a number you can show a bank or a partner and a number you made up.

Turning hours into money, honestly

Use a loaded hourly cost, not a salary divided by 2,080. Add employer contributions, software, equipment and the fact that nobody bills eight hours in an eight hour day. For most small businesses the loaded cost lands somewhere between 1.3 and 1.6 times the base hourly rate. Pick a multiplier, write down why, and use the same one every time.

Then apply the conversion test. Saved hours become money in exactly three ways:

• Billable substitution. The freed hours are now spent on work you invoice. This is the only case where the money is unambiguous.

• Avoided hire. You were about to add a person or a contractor and now you are not. Real, but only if the hire was genuinely imminent rather than hypothetical.

• Volume increase. Same team, more output, and the extra output actually sells.

If none of those three apply, record the saving as time returned to the team and say so. That is a legitimate outcome. Presenting it as profit is how businesses end up with fourteen subscriptions and no measurable improvement.

A worked example

A four-person design studio. Two tasks measured over two weeks before any change.

Baseline: proposals cost 4.75 hours a week, invoice chasing 1.8 hours. Total 6.55 hours.

After four weeks with a drafting workflow for proposals and a template-plus-reminder setup for invoices, remeasured:

New total: 2.0 hours plus 0.8 hours, so 2.8 hours. Saving of 3.75 hours a week, about 16 hours a month, with rework slightly improved rather than worse.

At a loaded rate of 65 per hour that is roughly 1,040 a month of time. Costs: 60 a month in subscriptions, plus 11 hours of setup and prompt iteration amortised over a year, call it 60 a month. Net around 920 a month of time value.

Now the conversion test. The studio owner used the freed proposal time to take on one extra small project a month at 1,800, so a portion converts cleanly to revenue. The invoice chasing time went back to the office manager's general workload and converted to nothing measurable. The defensible headline is therefore not "920 a month saved", it is "one additional project a month plus about five hours a week returned to the team". That statement survives scrutiny. The other one does not.

That 11-hour setup figure only holds if the project stays scoped to what was actually measured; knowing how to handle scope creep on an AI project is what keeps a tidy ROI calculation from quietly turning into a loss once "just one more workflow" creeps in.

The subscription audit nobody runs

Separate from measuring a single tool, once a quarter list every AI subscription with its monthly cost, the person who owns it and the last date anyone used it. Small businesses accumulate these fast, often at the individual level where they never reach a budget review.

Anything unused for 60 days gets cancelled. Anything nobody can name an owner for gets cancelled. This single exercise usually recovers more money than any individual tool's ROI, and it pairs naturally with a view on how much a small business should be spending on AI tools in the first place.

When to stop

Kill a tool if after eight weeks the time saving is under 15 percent, or the rework rate went up at all, or usage is concentrated in one enthusiast and nobody else touched it. That third one is the most common and the least discussed: adoption that never spread is a pilot that failed, and it usually means the workflow around the tool was never built, which is a different problem from the tool being bad. Getting a team to actually use AI is its own piece of work.

For context on where measurable gains tend to concentrate by task type, Anthropic publishes aggregate usage data in its Economic Index, which is more grounded than most vendor productivity claims because it reports observed task categories rather than survey self-reports. Useful for choosing what to measure, not a substitute for measuring your own.

Common questions

How long before I can judge whether AI is worth it?

Four to six weeks after the tool is genuinely in use, not from the day you bought it. The first fortnight measures learning, not the tool.

What if the task is not repetitive enough to time?

Then it is probably not the place to start. One-off work has real value from AI assistance but no measurable ROI, so pick a repeatable task for the measurement exercise and treat the rest as a bonus.

Should I count the time I spend learning the tool?

Yes, as a one-off cost amortised over a year. Leaving it out is the most common way ROI gets overstated, and for small teams it is often the largest single line.

Is it worth measuring quality as well as speed?

It is the whole point of the rework column. A tool that halves the time and doubles the corrections has not saved anything, it has moved the work to whoever reviews it. If you are earlier than this and still deciding whether to start at all, the signs a business is ready for AI automation come first, and the wider picture for small businesses sets the context.