Choosing an AI tool for a small business: questions to ask about data, cost and fit

Every software product now seems to come with an AI assistant, and a small business owner can sign up for five of them before lunch. The risk is less about picking a bad tool and more about picking one without asking where your data goes, what it will really cost once the team uses it, and whether it does the job any better than what you have.

This guide sets out the questions to ask before you commit, grouped into data, security, cost and fit, followed by a simple two-week trial and a worked comparison.

Start with the job, not the tool

The UK government’s AI Playbook lists “choose the right tool for the job” among its ten principles, and the idea applies just as well to a five-person firm. Before looking at products, write down:

• the task, in one sentence (for example, “draft first replies to customer emails about orders”);
• how often it happens and how long it takes now;
• what a good result looks like and who will check it;
• what data the task involves, especially customer or staff personal data.

Sometimes the answer is not AI at all. A saved reply template, a form or a simple automation rule may solve the problem more cheaply and predictably.

Questions about your data

This is the area where the gap between free consumer tools and business plans is widest.

1. Is my data used to train the provider’s models? OpenAI’s enterprise privacy page says that by default it does not use business data from its business plans and API for training. Anthropic’s privacy centre says the same for its commercial products. In both cases, sending feedback (such as a thumbs up or down) can count as opting in. Consumer plans have their own settings, so check them rather than assume.
2. How long is data kept, and can we control it? OpenAI, for example, says API inputs and outputs may be kept for up to 30 days for service delivery and abuse detection, and that admins on its business plans can set retention.
3. Will you sign a data processing agreement? If the tool will handle personal data, you need one. Both providers above offer them for business products.
4. Who is the controller? The ICO’s AI guidance says roles depend on who decides the purposes and means of processing, “regardless of how they are described in any contract.” A provider could be a processor for your use and a controller for training its own models. Either way, overall accountability for your use stays with you.
5. Do we need a DPIA? The same ICO guidance says that “in the vast majority of cases” using AI will involve processing likely to result in a high risk, which is the trigger for a data protection impact assessment. If you decide one is not needed, record why.

Questions about security

The National Cyber Security Centre’s guide, AI and cyber security: what you need to know, describes prompt injection as “one of the most widely reported weaknesses in LLMs”. In plain terms, text hidden in an email, document or web page can trick an AI tool into doing something you did not intend. That matters most for tools that can read your inbox, browse the web or take actions.

The NCSC suggests leaders ask questions such as: do we understand the data, model and software supply chain behind this tool, and what is the worst case, operationally or reputationally, if it fails? Add your own practical checks:

• Does it support single sign-on or at least two-factor authentication?
• Can we limit which files, mailboxes or systems it can reach?
• Does the provider publish independent security audits? OpenAI, for instance, states that its business products have completed SOC 2 Type 2 audits.
• Who in our business will respond if something goes wrong?

Questions about cost

AI pricing usually comes in two shapes: a monthly fee per user, or pay-as-you-go charges based on usage, often measured in tokens. Per-seat plans are easy to budget. Usage pricing can be cheaper for light use and expensive for heavy use. Ask:

• What is included in the plan and what triggers extra charges or limits?
• Do we need a higher tier for admin controls, data settings or integrations?
• What will setup, training and checking the output cost in staff time?
• What is the minimum contract length and notice period?

The hidden cost is usually review time. If a tool saves ten minutes drafting a reply but adds eight minutes of checking, the real saving is two minutes.

Questions about fit

• Does it work with what we already use? An assistant built into your helpdesk or accounts software may beat a better general chatbot that needs copying and pasting.
• Does it do our task well? Test it on your own examples, not the vendor’s demo.
• Can we get our data out? Check export options for conversations, templates and any knowledge base you build.
• What do the terms say about outputs? Check commercial use rights and any limits, especially for marketing content.
• Who will own it internally? Every tool needs one person responsible for settings, access and policy.

Run a two-week trial

1. Pick one task and collect 20 real examples, removing personal data unless the business plan and agreements are already in place.
2. Write three to five scoring criteria, such as accuracy, tone, time saved and edits needed.
3. Run the same examples through each shortlisted tool and score the results.
4. Time the full job, including checking, against your current method.
5. Ask the people who will use it daily which they would keep.

A worked comparison

An example (illustration only): a six-person online gift shop wants help drafting replies to customer emails. It compares a general AI chat plan with an AI add-on inside its existing helpdesk. All figures are invented for illustration.

The add-on costs a little more but avoids staff pasting customer details into a separate tool, scores better on real emails and saves more time. The shop picks it, completes a short DPIA and names one person to manage settings.

Common mistakes

• Staff using personal free accounts for client or customer data because no approved tool exists.
• Choosing on a polished demo instead of your own examples.
• Forgetting to count checking time when estimating savings.
• Connecting a tool to every mailbox and drive on day one.
• Skipping the DPIA and data processing agreement because the tool “only drafts text”.
• Having no exit plan for data and templates if you switch provider.