How to choose AI tools for a small business

Most small businesses do not have a tool problem. They have twelve tabs open, four trials running, two of them auto-renewing, and no way to say which one is earning its money. The market rewards this. Every vendor page is written to make its product look like the obvious first purchase, and none of them will tell you that the job you are trying to do is already covered by something you pay for.

So this guide is not a list of tools. It is a procedure you run in order, and it eliminates candidates at every step: name the job, count the seats, apply your constraints, read the price, read the data policy, then check how you would leave. Run it and you will usually finish with one tool and a reason. If you are an enterprise buyer with a procurement function and a security questionnaire, this is too small for you. It is written for a one-person business or a team under about twenty, spending its own money.

Start from the job, not the tool

Write down the job in one sentence, in the words you would use to a colleague. “Turn twelve pages of meeting notes into a client-ready summary each Friday.” “Reply to the same forty support questions without me typing them.” “Get a first draft of a landing page.” “Pull the numbers out of 300 supplier PDFs.”

Then classify it into one of six buckets, because the bucket, not the brand, decides the shortlist. Writing and thinking work goes to a general assistant such as Claude, ChatGPT or Gemini. Code goes to a coding tool such as Cursor. Moving data between apps on a schedule goes to an automation tool such as Zapier, Make or n8n. Analysis of your own files goes back to a general assistant, because the file is the input and no specialist tool is needed for one spreadsheet. Customer support goes to whatever already holds your inbox or helpdesk. Research with citations goes to something built for retrieval, such as Perplexity. Images are their own bucket and usually the least urgent.

Two rules save most of the money here. First, if the job happens fewer than about four times a month, do not buy a tool for it; do it inside the assistant you already pay for. Second, if the job is a chain of steps that runs whether or not you are awake, you need an automation tool, and a chat assistant will not do it no matter how good the prompt is. Getting this one distinction right is worth more than any feature comparison.

Team size decides the shape of the bill, not just its size

Below about three people, buy individual plans and stop. Claude Pro is $20 per month billed monthly, or $17 per month on annual billing, billed at $200 up front [1]. Cursor Pro is $20 a month [3]. These are cheap enough that the correct move is to buy one, use it for a month, and cancel if it does not stick.

From roughly three to twenty people, the question changes from “which plan” to “who administers this”. That is what the team tiers sell. Claude Team lists central billing and administration, usage analytics and project sharing among its features [1]; ChatGPT Business lists centralised billing across users, admin controls for managing users, roles and access, and usage visibility with spend controls [2]. Claude Team standard seats are $25 per seat per month billed monthly and $20 billed annually, with premium seats at $125 and $100 respectively [1]. ChatGPT Business is priced the same way, at $25 per user per month billed monthly and $20 billed annually, with premium seats at $125 monthly and $100 annually [2]. Those numbers are close enough that price should not be your tiebreaker between them; the tiebreaker is which one your team already opens without being asked.

Two team-size traps are worth naming. Seat sprawl is the first: you buy five seats because five people said yes, and three of them log in twice. Set a calendar reminder for day 60 to check actual use and drop the dead seats. The second is the premium-seat reflex. Premium seats buy usage, not capability. Anthropic describes its premium seat as 5x the usage of a standard seat, with the same feature set [1], and OpenAI says premium seats include all the same access as standard seats but with higher usage limits [2]. If nobody has hit a usage limit, you do not need them.

Constraints eliminate more candidates than features do

Constraints are the only part of this procedure that produces hard eliminations, so apply them before you compare features, not after.

If nobody on your team writes code, you need tools that are usable without it. Zapier sells itself on exactly that, listing a visual no-code editor among its plan features and answering in its own FAQ that it lets you move data between your apps “without being a developer” [4]. n8n takes the other road: its Community Edition is a standard self-hosted version you install from GitHub, following its own hosting documentation [5], which means someone on your side owns the server, the upgrades and the outages. It is the more flexible option and the more expensive one in attention. Choose it only if that person exists and has time.

If you are in the EU or handle EU personal data, your constraint is where the data rests and who processes it. Ask for the vendor’s data residency options in writing before you buy. OpenAI, for example, says eligible customers can store sensitive content at rest in the US, Europe, the UK and several other named countries, and can opt into in-region inference in the US or Europe [7]. Treat any vendor that cannot answer this in one page as unsuitable for that data, and keep using it for work that contains none.

If your business runs on Google Workspace or Microsoft 365, integration is a real constraint and not a nice-to-have. A tool that cannot see your documents makes you the integration, pasting content back and forth. Check the specific connection you need, not the logo on a partners page.

How to read a pricing page in four minutes

Find the billing unit first. There are only two common ones and they behave differently under growth. Seat pricing multiplies by headcount and is predictable: Claude Team at $20 per seat annually times 6 people is $120 a month, forever, whether they use it hard or not [1]. Usage pricing multiplies by volume and is not predictable until you have run it for a month. Zapier bills tasks, with a free tier of 100 tasks a month and Professional from $19.99 a month, and its Zap workflows, AI steps, code, MCP and SDK all draw on the same task allocation [4]. n8n bills full workflow executions rather than steps, with Starter at €20 a month for 2,500 executions and Pro at €50 for 10,000, both billed annually [5]. A workflow that fires once per new order costs you nothing to think about; one that polls every five minutes is 8,640 runs a month before a single customer does anything.

Then find the monthly versus annual gap and treat it as the price of your own uncertainty. On Claude, annual billing saves $3 per month on Pro and $5 per seat on Team [1]; ChatGPT Business is $5 per user cheaper annually [2]. Paying monthly for the first two months is worth that premium, because the failure mode you are insuring against is an annual commitment to a tool nobody opened in March.

Third, if you or a contractor will call an API rather than use the chat app, check whether there is a batch price. Work that does not need an answer in the next few seconds is much cheaper: Anthropic’s Message Batches API charges 50% of standard API prices, with most batches finishing in under an hour and results expiring if a batch does not complete within 24 hours [6]. Classifying 4,000 support emails overnight is a batch job. Answering the one in front of you is not.

Fourth, look for the floor and the ceiling. The floor is the cheapest configuration that actually does your job, which is often one paid seat and a free tier of something else. The ceiling is what the bill becomes if the thing works and you triple the volume. Compute both before you enter a card number.

How to read a data policy in three questions

You are asking three things, and vendors answer all three in public documentation. If the answers are not public, that is your answer.

First, whether your content is used to train models. The honest current position varies by product tier, and by plan within a vendor. For its consumer plans, Anthropic uses Claude chats for training where you have turned on the model improvement setting, where a conversation is flagged for safety review, or where you joined a programme such as its trusted tester scheme, and Incognito chats are not used even when model improvement is on [8]. OpenAI states that by default it does not use data from ChatGPT Business, Enterprise, Edu or its API platform, including inputs and outputs, for training or improving its models, with explicit opt-in available in the API dashboard [7]. Note the shape of both statements: business tiers are excluded by default, consumer tiers depend on a setting you can check today.

Second, how long it is kept and whether you control that. Retention is separate from training and is the one people skip. OpenAI says qualifying organisations can configure how long it retains business data, including opting for zero data retention on the API platform [7]. If you handle client material under an NDA, this is the paragraph your client will ask about.

Third, where it lives. Region matters if you are bound by EU rules or a client contract that names a jurisdiction, and it is a per-vendor, per-tier answer rather than a universal one [7]. Check it once, write the answer in a file, and stop relitigating it every time someone asks whether a tool is “GDPR compliant”, which is a property of how you use it, not a badge a vendor wears.

The exit check before you enter

Before you pay, work out how you would leave, because the cost of a wrong choice is not the subscription, it is the month you spend extracting yourself.

Ask where the work lives. A chat assistant holds conversations and you lose little by switching. A workspace tool such as Notion holds your documents and structure, so switching means an export and a rebuild. An automation platform holds logic, and the logic is the part you would have to build again in whatever comes next. The deeper the tool sits, the more you should demand of it before entering.

Then check three things you can verify in the trial. Can you export your content in a format you can read without the vendor, today, from a menu, without asking support. Do you own the credentials to the accounts it connects to, so that removing the tool does not break your email or your store. And is the output portable, meaning prompts, documents and templates you wrote are text you keep rather than objects locked in a UI. A tool that fails all three is not disqualified, but it should be earning noticeably more than the alternative that passes.

The site’s Decision Engine at /decide runs this same procedure as a set of questions and hands you a shortlist with the reasoning attached, if you would rather answer prompts than hold the steps in your head.

What still goes wrong

Prices change, and this page is a snapshot verified on the date in its header. Vendors restructure tiers, rename plans, move features between them and adjust what a “task”, “credit” or “execution” means. Every number here is cited to the vendor’s own page for exactly that reason: check the source before you commit money, and treat any figure you see quoted in a blog post, including this one, as a starting point rather than a quote.

The procedure also cannot tell you the thing you most want to know, which is whether the tool will actually do your job well. Nothing can, short of running your real work through it during the trial. So spend the trial on one real task you were going to do anyway, not on the demo the vendor suggests. If you finish the trial with a piece of work you actually shipped, you have your answer; if you finish it with an impression, you do not.

Finally, this procedure optimises for reversibility, which is the right bias when you are spending your own money and cannot absorb a bad year. It is the wrong bias if you are deliberately building something deep on one platform, where committing hard and early is the point. Know which of those two you are doing before you start, because the same caution that protects a small business from lock-in will also stop it from ever getting good at anything.