AI for small business: where it helps, what it costs and where to stop

How can I use AI in my small business?

Use AI for the reading and writing that eats the owner's and the team's evenings: first drafts of emails and proposals, summaries of long documents, tidying notes into a checklist, rewriting a policy in plain English, and turning a messy spreadsheet into a short explanation. These jobs share three features: they repeat, someone in the business can tell a good result from a bad one, and an error is caught before it does damage.

Small firms often lack the specialist roles that larger companies use to fill these gaps. A seven-person firm usually has no marketing writer, no operations analyst and no one whose job is to read supplier terms. A general assistant cannot replace those people, but it can produce a first draft that the owner then corrects, which is often faster than starting from a blank page. The value comes from the correction, not from the draft.

Should a small business use ChatGPT, Claude or something else?

Pick the assistant that fits how you already work, on a business plan rather than personal free accounts. If the firm runs on Microsoft 365, Copilot is the natural first evaluation; on Google Workspace, Gemini; otherwise compare ChatGPT and Claude business plans on the tasks you actually do. Try each one on the same two real tasks with fictional or approved data, and choose the one whose drafts need the fewest corrections from you. Our separate guides to ChatGPT for work and Claude for business cover each tool in more depth.

The plan matters more than the brand. A business plan typically gives you a shared admin, a way to remove a leaver's access and published terms about how your data is handled. For example, Anthropic states that by default it does not train its models on inputs or outputs from commercial products such as Claude for Work and its API, unless the customer gives feedback or otherwise opts in. Read the equivalent page for the tool you choose, and confirm it covers the plan you are paying for, not just the enterprise tier.

How much does AI cost for a small business?

The direct cost is usually a per-seat monthly subscription to a business plan, plus any specialist tools you add later. Prices change often and differ by region and billing period, so check each vendor's current plan page rather than trusting a figure in an article, including this one. Some tools are free to try, but free consumer accounts usually come with weaker admin control and different data terms, which is a poor fit for company information.

The larger cost is time: setting rules, learning to give good context and reviewing output. Budget a few hours of the owner's time in the first month to agree the boundaries and look at work samples, and some practice time for each person. A subscription nobody knows how to use well is the most expensive option. If you want structured practice for the team, Tosheo.ai's planned prices are on the pricing page; the product is pre-build.

What does a first month look like in a seven-person firm?

Kestrel Bookkeeping is a fictional seven-person firm serving local tradespeople. The owner, Priya, approves one business-plan assistant and two tasks for the first month: drafting reminder emails to clients who have not sent their receipts, and turning HMRC guidance pages into a plain-English checklist for staff. Client names, figures and bank details are not to be entered; staff use placeholders such as [CLIENT] and [AMOUNT] and fill them in afterwards.

In week one, a bookkeeper asks: 'Write a reminder email to a client about missing receipts.' The draft is polite but threatens a 'late filing penalty of £100' that nobody supplied, and sets a deadline that does not match the firm's calendar. Priya's review note is short: never let the assistant invent amounts, dates or penalties. The revised prompt gives the firm's actual deadline, the tone ('friendly, brief, no threats'), and the rule 'if a date or amount is needed, write [CHECK] instead'. The second draft is usable after a one-line edit.

In week three, the checklist task shows the other side. The assistant's summary of a guidance page is clear but leaves out an exception that matters for some clients. Priya decides the checklist stays, but every item must carry a link to the paragraph it came from, and a qualified colleague signs it off. At the end of the month, the firm keeps the reminder emails, keeps the checklist with sign-off, and says no to a third idea, drafting client tax explanations, because nobody wants to own the review risk yet. That is a good outcome.

What should a small business keep out of AI tools?

Keep out anything you would not want to explain to a client, a regulator or a journalist if it leaked: personal data about customers or staff, bank and card details, passwords, unreleased financial figures, and client documents you do not have permission to share. The NCSC advises against including sensitive information in queries to public large language models, and Acas reminds employers that existing data privacy policies apply to AI use.

Placeholders solve most of the problem. Write [CLIENT], [ADDRESS] or [TOTAL] in the prompt and fill in the real details afterwards in your own system. If a task genuinely needs personal data, stop and check your data protection position first; the ICO's guidance on AI and data protection is the UK starting point. This is general information, not legal advice.

What goes wrong most often?

The most common failures are confident invented details, promises the business cannot keep, and the owner checking less carefully as the weeks go by. NIST's Generative AI Profile describes the first as confabulation, confidently stated but erroneous or false content, and warns about automation bias, the tendency to defer to automated output. Both are worse when you are busy, which is exactly when a small business owner reaches for the tool.

Build the check into the task rather than relying on willpower. Ask the assistant to mark anything it was not given with [CHECK], keep a short list of facts the business must never get wrong (prices, deadlines, warranties, legal wording), and read those lines first on every draft.

When should a small business not bother with AI?

Skip it, for now, when the work is mostly physical or face-to-face, when there are only a handful of written tasks a month, or when nobody has time to review what the tool produces. A two-person landscaping firm that writes three quotes a month may get more from a good quote template than from a subscription.

Also hold back on anything that makes a decision about a person, such as screening job applicants, or that gives regulated advice to customers. Those uses need far more care than a general assistant and a quick read-through, and in some cases legal advice.

How do small teams learn to use AI well?

Small teams learn fastest by practising their own tasks with someone checking the result, not by watching long general courses. Agree two or three tasks per person, give a sample prompt and a review checklist, and look at one real output together each week for the first month.

If you want that structure without building it yourself, Tosheo.ai is being built to provide role-specific practice for teams of 5 to 200, with hands-on tasks graded against a visible rubric and a manager view of progress. It is pre-build, and our page on AI training for small teams explains the planned offer. For many firms of under ten people, the routine described on this page is enough to start.

Sources and scope

Anthropic Privacy Center, Is my data used for model training? (commercial products), checked 16 October 2026 ↗

NCSC, ChatGPT and large language models: what's the risk?, 14 March 2023 ↗

Acas, One third of employers think AI will increase productivity, 15 May 2025 ↗

ICO, Guidance on AI and data protection, dated 15 March 2023 ↗

NIST, AI 600-1 Generative AI Profile, July 2024 ↗