Buying an AI Tool
The best AI tool for your business is not the top-rated one
Quick answer: there is no best AI tool for small business. There is the tool that does your one job, that your people still open in week three, that you can trust with your data, and that fits under a ceiling you set before you looked. The top-rated tool on any list won on someone else's job. Most firms should start with the AI assistant already bundled into Microsoft 365 or Google Workspace, prove it cannot do the job, and only then shop. The free five-article method for doing that is at practicalmethods.ca/guides/buying-an-ai-tool.
What "best" means on a ranking
Every "best AI tools for small business" list is built the same way. Someone runs a set of general tasks through a dozen products, weighs features, speed and price, and produces an order. The order is honest. It is also useless to you, because it answers a question you did not ask.
A ranking cannot know three things. It cannot know your one job, the specific weekly task you want done, which is the only thing a tool should be scored on. It cannot know your people, and whether the tool will live in the tab they already have open or in a new one nobody visits by week three. And it cannot know your data: whether client files are going into it, where they are stored, and whether the vendor trains on them by default.
Those three things decide most purchases in a firm of five to fifty people. The ranking scores none of them. So the tool at the top of the list is the best tool for a generic job, run by an enthusiast, on public data. That is not your firm.
The tool you already pay for is usually missing from the list
Here is the part rankings skip because it does not make a good list. If you pay for Microsoft 365 or Google Workspace, you already own an AI tool. It sits inside the documents and the inbox where the job already happens. It costs nothing extra or a modest add-on. And on most lists it is either absent or ranked in the middle, because it loses on capability to standalone products.
It wins on something else. It is already there. A tool that is already in the document scores badly on a reviewer's benchmark and well on the only benchmark that matters in a small firm, which is whether six people used it in week three without being reminded. That is why the first step of any honest AI tool decision is not a comparison. It is proving the tool you already pay for cannot do the job.
The method calls this gate 0, the bundled test. Three checks. Do you pay for a suite with an assistant in it? Has someone run the actual job through it, with real inputs, at least three times? Can you name, in one sentence, the thing it could not do? If the second check is a no, you are not shopping yet. If you can write the sentence, that sentence is what you are shopping for, and it will cut most of the ranking on its own.
Why "best" fails in four predictable ways
Firms that buy off a ranking tend to fail in one of four ways, and each one maps to a criterion the ranking did not score.
It did not do the job
The demo was the vendor's inputs on the vendor's terms. Your inputs were longer, messier, and in a format the tool reformats first. The fix is a trial, run by the person who does the job, on the ugliest input you have, before anything is scored.
Nobody used it
The tool worked. It lived in a new tab. By week three it was one person's habit and by the quarter it was a renewing subscription with one active seat. The fix is to give three people the trial, say nothing, and count who is still using it in week three.
You could not stand behind the output
Two errors in ten that a client would have caught, or a cited source that did not exist when checked, or a data term nobody read. The fix is ten outputs checked by hand and the data processing terms read in full, not the landing page.
The invoice was not the pricing page
Per seat was cheap until the feature you needed lived one tier up. Usage was cheap until the tool worked. The fix is one annual all-in number per tool, in one currency, including the bundled option at its real marginal cost, which is often zero.
None of these are tool failures. They are decision failures, and a ranking cannot prevent any of them because it does not know the job, the people or the data.
What to do instead: shape, gates, criteria
The method is short and it is free to read. It runs in this order, and the order is the point.
Eight lines, before you open a pricing page: the one job, who does it weekly, seats in twelve months, the annual ceiling all-in, the data it will touch, the systems it must sit inside, what "worked" looks like at ninety days, and who signs. Half the market disappears on its own once these exist.
Bundled, a general assistant seat, a point tool for one job, or a custom build. Rankings mix all four as if they were one market. They are not.
Pass or fail: ceiling, path, data, security, fit, exit. A tool that fails any one is out and never gets scored. This is what stops the impressive one winning on points after failing something that should have ended the conversation.
Weighted: the job (5), the adoption (4), the trust (4), the deal (3). Forty-six questions, each rated one to five from trial evidence. The weights are yours to change. The order is not.
Each step has its own article, with the questions, in the free series at practicalmethods.ca/guides/buying-an-ai-tool. Start with the narrowing article. It is the one that replaces the ranking.
A worked example, briefly
A 22-person engineering consultancy on Google Workspace wanted a tool to draft the narrative sections of fee proposals. Six seats, a ceiling of six thousand dollars a year. Three tools passed the gates: the bundled assistant at no extra cost, a general assistant seat at about two thousand a year, and a proposal-builder point tool at about five thousand.
The general seat wrote the best draft by a distance and scored highest on the job. The bundled assistant won adoption, because five of six people were still using it in week three, and it won the deal, because zero is a real number. The totals were four points apart. The scorecard had done its job: it found that the decision came down to who already ran the admin console, where the draft landed, and a fee table that needed a template rather than a model. They kept the bundled tool, built one template, spent nothing, and put the general seat in the calendar as the ninety-day fallback.
No ranking would have produced that answer. Most would have ranked the tool they chose in the middle.
Frequently asked questions
What is the best AI tool for a small business?
There is not one. The right tool depends on the one job you want done, whether your people will use it in week three, whether you can trust it with your data, and whether it fits under a ceiling you set first. The free method at practicalmethods.ca/guides/buying-an-ai-tool walks through all four.
Should I just buy the top-rated AI tool?
No. The top-rated tool won on a reviewer's tasks, run by an enthusiast, on public data. It has not been tested on your job, your people or your data. Run the tool you already pay for on the job first, then shortlist three to five and put them through the gates.
Is the AI in Microsoft 365 or Google Workspace good enough?
Often, for jobs that live in documents, email and meetings. It usually loses to standalone tools on capability and wins on adoption because it is already where the work is. Test it on the real job three times before deciding it is not enough.
How do I know if I need an AI tool at all?
Run gate 0, the bundled test. If you can name in one sentence the thing your existing assistant cannot do on the job, you need a tool for that sentence. If you cannot, you do not need a tool yet.
How many AI tools should I compare?
Three to five, after the shape is written. Fewer and you have not looked. More and nobody will run the trials properly. The gates usually cut five to two or three.
What should I test in an AI tool trial?
The real job, with real inputs, run by the person who does it weekly. Time the job by hand first. Give the tool your worst input. Run the same input three times. Note where the output lands. Check ten outputs by hand for errors.
Why do employees stop using AI tools?
Usually because the tool lives in a new tab, needs a prompt expert to get value, or produces output that has to be copied somewhere else. Score adoption before buying by giving three people the trial and counting who uses it in week three unprompted.
Is it safe to put client data into an AI tool?
It depends on the terms, not the tool. Confirm training on your inputs is off by default, retention is stated, storage region is stated, and deletion works. If any of those is unclear, use anonymised data or do not use the tool for that job.
Criteria: Buying an AI Tool
Five articles. Narrowing, the job, the adoption, the trust, the deal.
Read the free methodThe list is not the decision. It is how you find the one or two things that are.
Not a review, not rankings, not software advice. A method.
