The AI Saved You Three Hours. Did It Save You Money?

In November 2025 the US Census Bureau changed one phrase in a question it puts to hundreds of thousands of businesses. The old version asked whether a company had used AI in producing goods or services. The new one asks whether it used AI in any business function. Nothing about the economy changed that month. The reported adoption rate nearly doubled, from roughly 10% to 17%.

That single fact is the most useful thing a small business owner can know about AI return on investment, because the same trick is being played on you, usually by accident, every time somebody tells you what AI is worth. The number depends almost entirely on what you decided to count.

Why did the official AI number double without anything changing?

Because the first question only counted AI inside production, and production is where almost nobody uses it. Once the question widened to any business function, all the drafting, summarising and searching that was already happening got counted.

The St. Louis Fed took that puzzle apart in a piece on why measuring AI adoption depends on how you ask. Worker surveys were reporting 35% to 40% adoption while the main firm survey reported 5% to 7%, a gap wide enough to suggest one of them was simply wrong. Neither was. European data, which asks about any business purpose, shows production accounting for only about 21% of AI applications, well behind marketing and sales at roughly 35% and business process organisation at about 31%. Across countries, an any purpose measure comes out around five times larger than a production focused one.

Apply that to your own shop for a moment. If somebody asks whether you use AI to make your products, the answer is probably no. If they ask whether you used it this week to write a product description, answer a customer, or summarise a supplier email, the answer is probably yes. Both answers are true and they will produce wildly different pictures of your business.

How many small shops actually use AI?

Fewer than the coverage implies, and retail is at the bottom of the table. The Census Bureau put overall business AI use between 17% and 20% in the six months to May 2026, with retail trade at around 14%.

The Census Bureau's May 2026 breakdown of AI use by firm size is worth reading slowly, because the shape of it matters more than the headline. Adoption is strongly tied to headcount.

Read the third row again. Among firms with fewer than twenty employees, adoption did not move at all over six months, while every larger category climbed. The smallest businesses are not quietly racing ahead with AI. If you run a shop of one or two people and you feel behind, the data says you are roughly where everyone like you is.

Do the firms using AI actually report a return?

Mostly they report nothing measurable. A survey of nearly 6,000 senior executives across four countries found roughly 70% of firms had adopted AI and 89% reported no productivity impact from it over the previous three years.

That study, summarised by the National Bureau of Economic Research in its May 2026 digest, is the coldest water available on this subject and it deserves to be read carefully rather than as a verdict. United States adoption came in at 78%, Australia lowest at 59%. Text generation was the dominant use, at 41% of firms. More than 90% of executives reported no effect on employment. Among the minority who did report a productivity effect, the estimated boost was 0.29%.

Set that against what the same executives expect. They forecast a 1.4% labour productivity increase over the next three years, a 0.7% employment reduction, and a net output gain of about 0.8%. So the people running these companies have measured almost nothing so far and expect a great deal shortly. Both halves of that sentence are informative, and the gap between them is where most AI budgets currently live.

What do the shops that did measure a return report?

Specific hours on specific tasks, never a percentage across the business. The examples that survive contact with a survey are all narrow, and that narrowness is the finding rather than a limitation of it.

The Federal Reserve Bank of St. Louis surveyed firms in its district and published the results on AI adoption, employment and productivity in July 2026. A law firm documented two to three hours of weekly efficiency gains per attorney from AI assisted research and document preparation. One professional services firm reported 35% revenue growth over a year without adding staff, another 15% growth in revenue per employee. Meanwhile 34% of firms used AI tools with only a small portion of their workforce, 25% were piloting without regular deployment, and 38% of the non adopters cited missing skills, data or infrastructure.

Notice what the credible figure looks like. Two to three hours per attorney per week is a claim you could check. A 35% revenue increase attributed to AI is a claim nobody can check, including the person who made it, because a year of a business contains a hundred other causes. When you set your own target, aim for the first kind.

What should you measure instead of return on investment?

One task, twice, with a stopwatch. Return on investment across a whole small business is unmeasurable in any honest sense, and pretending otherwise is how owners end up paying for four subscriptions they never evaluate.

The reason a single task is the right unit comes straight from the data. A Census Bureau working paper on how AI spreads through firms found that among companies using it, 65% applied it to three or fewer tasks, with sales and marketing the most common function at 52%. Firms are not transforming themselves. They are picking a few jobs, which means a few jobs is what you can sensibly evaluate.

Pick something you do at a known cadence with a clear beginning and end. Writing the week's product descriptions. Answering the repeat questions in your inbox. Reconciling the month's invoices. Then time it for a week the way you normally do it, and time it for a week with the tool, doing the same volume. Convert the difference into money at whatever an hour of your time is genuinely worth, which for most owners is higher than they admit. Then subtract three things: the subscription, the time you spend checking the output, and the work you redo when a bad answer slips through.

That last subtraction is the one people skip and it is often decisive. A tool that saves forty minutes of writing and costs twenty five minutes of checking has saved fifteen minutes, not forty. Our breakdown of what small business owners actually pay for AI goes through the subscription side of that arithmetic in more detail.

Why does the return show up so late?

Because the tool arrives before the change in how you work, and only the second one produces savings. Buying the subscription takes a minute and rearranging a habit takes months.

This is visible in the survey data as a kind of holding pattern. A quarter of firms in the St. Louis Fed survey were piloting AI without deploying it regularly, and many described it as an active topic of discussion while making no operational changes. That is not laziness. Changing a process you rely on during a working week is genuinely risky, and most small businesses cannot run a parallel version of themselves to test it.

The practical consequence is that the first month of any AI tool will look like a loss, because you are paying for it while still doing the job the old way half the time. Judge it in month two or three. If it still has not cleared its own cost by then, it is not going to, and the honest move is to cancel rather than to keep it as insurance against a future you have not planned. Deciding what to try first, in the right order, is its own question, which we worked through in a guide to what to automate first in a small business.

What does a negative result actually look like?

A task that takes the same time and now requires supervision. That is the common failure and it rarely announces itself, because the output looks finished.

There are three shapes it takes. The first is a task where the checking costs as much as the doing, which happens with anything where an error is expensive and hard to spot, such as pricing, stock figures or anything a customer will treat as a promise. The second is a task you did not actually do before. Producing eight social posts a week with AI is not a saving if you previously produced none, it is a new cost with an unproven return. The third is quality drift, where output that is acceptable in isolation gradually stops sounding like your business, which costs you nothing measurable until it costs you a customer.

None of these show up in a productivity statistic. All of them show up in a one task timing test, which is the argument for doing the small boring version of the measurement rather than the impressive one.

How much is AI actually being used inside a working week?

About an hour and a half, by the people most likely to tell a survey they have adopted it. Senior executives in the four country study averaged 1.5 hours of AI use per week, and United States workers 1.8 hours.

Those numbers are worth holding next to the adoption headline. A firm where the chief executive spends ninety minutes a week with a chatbot counts as an adopting firm. It is a perfectly honest answer to the question asked, and it explains how 70% adoption and 89% reporting no productivity effect sit in the same dataset without either being false. Adoption measures whether the tool is present. It says nothing about dose.

The Census Bureau working paper on how AI diffuses through firms, functions and worker tasks gives the same picture from the other end. Writing, document analysis and information search were the primary uses at worker level. Sales and marketing was the most common function at 52%, strategy and business development at 45%, IT at 41%. The paper put firm level use at 18%, rising to 32% when weighted by employment, which is another reminder that a big company using AI in one department drags every national average upward.

For a merchant this is oddly reassuring. The dominant real world use of AI in business is writing things and finding things, which is exactly what a shop of one or two people needs it for. You are not missing a sophisticated application that larger competitors have unlocked. They are drafting emails too.

Does using AI mean cutting someone?

The evidence says no, and unusually for this subject the surveys agree with each other. Employment reductions attributable to AI showed up in 2% of firms in the Census working paper, and more than 90% of executives in the NBER study reported no employment effect over three years.

The St. Louis Fed survey asked about the next twelve months rather than the past three years and found 49% expecting no staffing change at all, 18% expecting a shift in the skills they need rather than fewer people, and close to 20% expecting a modest decrease. Its own summary is that AI is helping businesses expand capacity with the workforce they already have.

That distinction matters for how you value a saving. If AI gives you back four hours a week and you have no intention of reducing headcount, the return is not a payroll cut, it is whatever you do with four hours. For a one person business that might be genuinely valuable or it might be nothing, depending entirely on whether there is revenue generating work waiting that you were too busy to reach. Be honest with yourself about that before you count the hours as money. Time returned to a business with nothing queued behind it is rest, which has real worth, but it is not a return on investment and should not be written down as one.

What to do with the number once you have it

Very little, and that is the point. The output of this exercise is a decision about one subscription, not a strategy.

If the task clears its cost, keep the tool and leave the rest of your business alone until that task is stable. If it does not, cancel it and try a different task rather than a different vendor, because the task is usually what determined the result. The firms in every one of these surveys that reported something real reported it about a specific job, and the ones reporting nothing had generally spread the tool thinly across everything.

Keep a note of what you measured and when. In six months the tools will be better and your answer may flip, and you will want to know what you compared against rather than reasoning from memory. That record is also the only defence against the thing this article started with, which is a number that moved because somebody changed the question.

If part of what you are evaluating is the cost of the platform your shop runs on rather than the tools bolted onto it, our pricing page sets out what building and running a store here actually costs, in the same per task terms this article argues for. And if the answer to your timing test is that AI saves you nothing on the job you tested, that is a real result. Write it down, keep the money, and test a different job next quarter.