What a Small Business Really Spends on AI Each Month

Ask ten people what AI costs a small business and you will get ten answers spanning three orders of magnitude, because almost nobody is measuring it. The figures that circulate come from vendor surveys, consultancy decks and content marketing, and they describe an aspiration rather than a bank statement.

There is one dataset that avoids the problem entirely by not asking anyone anything. The JPMorganChase Institute looked at what small businesses actually paid, in their actual accounts, to actual AI vendors.

What does a small business actually pay for AI?

About $28 a month, at the median, among those paying anything at all. The Institute's research on AI use among small businesses tracked de-identified business banking transactions for 4.6 million firms from January 2019 to December 2025, which makes it a payments record rather than a survey response.

Three quarters of small businesses that spend on AI at all spend less than about $90 a month. The picture of a small firm running a stack of eight specialised tools is not wrong, it is just rare enough that it sits above the 90th percentile of a very large sample.

The Institute also measured how the entry price moved. Starting spend fell from roughly $50 a month in 2019 to $20 to $30 in 2025, a drop of about 60%, and adoption speed changed accordingly. Businesses in the 2019 cohort took 77 months to reach 10% adoption. The 2025 cohort got there in six.

Why does the market talk about thousands?

Because the market is describing firms with employees, and most shops do not have any. The gap shows up cleanly in the Institute's split: 26.1% of employer firms had adopted AI by December 2025 against 15.3% of firms with no employees, a difference of nearly eleven percentage points.

The United States Census Bureau's business survey finds the same shape from a different direction. Its reporting on AI use by firm size covering December 2025 to May 2026 puts overall use between 17% and 20%, with 37% among firms of at least 250 employees, 32% among firms of 100 to 249, and under 20% among firms with fewer than 20. Retail trade sat at 14% against a national rate of 19.8%.

The finding inside that data deserves more attention than it got. Between December and May, AI use rose among firms with at least 20 employees and did not change significantly among firms with fewer than 20. The much discussed adoption curve is happening mostly above the size where a business is one person and a laptop.

Should you buy seats or pay for what you use?

For one person, seats. For anything automated, usage. The two pricing models are not competing offers for the same thing and choosing between them by price alone is how small firms end up paying twice. Which shape suits you depends on the work, so it helps to decide first which jobs in a small business are worth handing over at all.

A seat is priced for a human sitting at a keyboard. Published list prices show the shape clearly: Anthropic's plan pricing, checked on 4 September 2026, lists Pro at $20 a month billed monthly or $17 with annual billing at $200 up front, and a team seat at $25 monthly or $20 annually with a two seat minimum. That is the common structure across the consumer grade tools, and for a shop owner who wants a capable assistant open in a tab, one seat is the whole answer.

Usage pricing is for the things that run without you: a script that categorises orders, a job that drafts product copy overnight, an agent that answers a first line support question. Here you pay per unit of work, and the cost of a badly designed job is unbounded in a way a seat never is. We went through where that bill comes from in the piece on the AI bill you have not seen yet, and the underlying hardware economics that drive the numbers in the breakdown of why the same GPU costs anywhere from under a dollar to fifteen an hour.

The rule that keeps a small shop out of trouble is to put a hard spending cap on every usage based key on the day you create it, before you write any code against it. Not a budget alert, a cap. Alerts tell you about the money after it has gone.

What is not on the invoice?

The part that actually decides whether the tool was worth buying.

Checking the output is the big one and it is never counted. A tool that drafts forty product descriptions in a minute has moved your work from writing to reviewing, and reviewing forty drafts properly takes longer than most owners expect and shorter than writing them. The saving is real. It is just smaller than the demo implies, and it disappears entirely if the review is skipped, because the cost then arrives later as a wrong claim on a listing.

Seats you stopped using. The commonest waste in a small firm's software spend is not overpriced tools, it is correctly priced tools nobody opens. Because AI subscriptions are individually cheap they escape the scrutiny a $300 line item gets, and a $20 tool nobody uses costs $240 a year exactly like a useful one.

Annual lock in. The discount for paying yearly is typically in the range of 15% to 20%, and it is genuinely good value for a tool you have used for six months. On a tool you have used for six days it is a bet on a category that reprices every quarter. Pay monthly until a tool has survived a full billing cycle of your actually using it.

Credit systems. Any product that meters in credits rather than currency has a conversion rate it controls and can change. That is not sinister, it is how the underlying costs get passed on, but it means your budget is denominated in something the vendor mints. Check what a credit buys before you buy a bundle of them, and check again after a repricing.

Leaving. Ask what happens to your data and your prompts when you stop paying, before you start. A tool holding two years of your customer conversations with no export is more expensive than its price, and the bill arrives on the day you want to switch. We wrote about that failure at length in the piece on the AI tool you depend on being switched off.

What should the first twenty dollars buy?

A general assistant, not a specialist. This is counterintuitive to anyone shopping by category, because the specialist tools describe your problem back to you in your own words and the general one does not describe anything. But a specialist tool solves one job and a general assistant solves whichever job you happen to have on a given Tuesday, and for a business where the jobs change weekly that flexibility is worth more than any single integration.

The exception is a job you do at volume on a schedule, where the specialist earns its place by removing setup every time. If you process a hundred supplier documents a month, a tool built for that beats pasting them into a chat window. If you process four, it does not.

One thing worth spending the second twenty dollars on, if you spend it at all, is whatever removes the job you most dislike. That is not an efficiency argument, it is a persistence one. Tools bought to fix the task you avoid get used, and tools bought to optimise the task you already do well get abandoned in month two, which is how a stack accumulates without anyone deciding to build one.

The break even calculation nobody writes down

A $20 subscription needs to save you a little under an hour a month to be worth keeping, if you value your time at anything like what your business earns per hour. That sounds like an easy bar and it is, which is why the calculation is usually skipped and why the answer is so often no anyway.

The reason it fails is that the hour has to be a real hour that comes back to you, not a notional one. If a tool saves twenty minutes on a task you do weekly, that is roughly 87 minutes a month and the subscription pays for itself. If it saves two hours on something you do twice a year, it does not, however impressive the two hours felt. The frequency matters more than the size of the saving, and small businesses consistently buy for the size.

Run the arithmetic on the three things you actually do most often. For a shop that is usually writing product copy, answering the same handful of customer questions, and reconciling something. Estimate the current minutes, honestly, by timing yourself once rather than guessing. Then ask what fraction of that a tool removes rather than shifts. A tool that turns forty minutes of writing into fifteen minutes of editing has saved twenty five minutes. A tool that turns it into forty minutes of editing has saved nothing and cost $20.

There is a second term in the equation that only appears later. Every tool you keep is a thing to maintain: an account, a password, a subscription to review, a vendor whose terms change, a place your data sits. Six tools at $20 is not six times the burden of one tool at $20, it is considerably more, because the coordination between them is where the time goes. This is the strongest practical argument for the pattern the payments data already shows, where 72% of small businesses paying for AI pay for exactly one thing.

Why is retail so far down the adoption table?

Because most of the tooling was built for people who work in documents. Look at the sector spread in the census figures: Information at 39.7%, Finance and Insurance at 33.9%, Retail Trade at 14%. The Institute's industry split runs the same way, with information at 39.3% and professional services at 30.3% against transportation and warehousing at 5.4% and construction at 8.9%.

The pattern is not that shopkeepers are slower to adopt technology. It is that a general purpose assistant is immediately useful to somebody whose output is text and only indirectly useful to somebody whose output is a parcel. The jobs in a shop that AI genuinely improves are mostly the ones sitting between the physical work: the listing copy, the supplier email, the reply to the review, the question about a delivery. Those are real and they are a smaller share of the day than they are for a consultant.

Two consequences follow for a merchant reading this. The first is that you should expect a lower return than the case studies imply, because the case studies come from sectors where the whole job is the kind of work these tools do. The second, more useful, is that the adoption gap is not a warning. Retail sitting at 14% while information sits near 40% means the competitive pressure to adopt is far lower in your sector than the marketing suggests, and you can afford to be slow and deliberate about it rather than buying a stack because a newsletter said the window was closing.

A budget that fits the evidence

If the median is $28 and the 75th percentile is $90, a sensible target for a one person shop is somewhere in that range, arrived at deliberately rather than by accumulation.

One general assistant seat, at around $20 a month, does most of what a shop owner needs: drafting, summarising, reading a document, working through a problem out loud. That is the single highest value line item and for many businesses it should be the only one for the first six months.

Then add tools only against a named job you are currently doing badly or slowly. Not a category, a job. Writing the weekly supplier order. Replying to reviews. Getting line items out of PDFs. If you cannot name the job in one sentence, the tool is a purchase in search of a use, and it will become one of the 72% of subscriptions that exist because somebody signed up during a trial.

Review the whole stack quarterly with one question per line: what did this do in the last three months that I can point at? Anything that fails is cancelled the same day. This takes twenty minutes and it is the only reliable cost control in a category where the individual prices are too small to trigger anybody's judgement.

Does spending more get you better results?

Not reliably, and the shape of the data suggests why. If nine percent of small businesses pay for three or more services while three quarters pay under $90, the firms spending heavily are not mostly buying better AI. They are buying more categories of software, each of which needs setting up, connecting and checking by the same one or two people. Past a certain point the constraint stops being the tools and becomes the attention available to supervise them.

The tools worth having are the ones that fit a job you already do on a schedule, which is the filter we applied in the piece sorting AI tools by job rather than by category. A shop with one assistant seat and a clear weekly routine gets more out of AI than a shop with six subscriptions and no routine, and it pays a quarter as much.

That is also the reasoning behind how we price our own product, which bills for the work done rather than for a seat that sits there: for a business at this size, a bill that tracks usage is easier to reconcile against value than one that arrives whether you opened the tab or not. Our own numbers are on the pricing page, and the honest comparison for any tool is the same one this article argues for throughout. What job does it do, how often, and who checks it.

The median small business paying for AI spends about $28 a month. The interesting question is not how to spend more. It is which single job the next twenty dollars should do.MaShop, reading the JPMorganChase Institute data, 4 September 2026