Practical AI Risks for Small Businesses

A lot of advice on AI right now sounds either very excited or very afraid. Neither is much help to a small business owner trying to decide whether to put a tool in front of their team next month.

The useful version of this conversation is more boring and more specific. There are a handful of real concerns that deserve to be taken seriously by anyone signing a contract or rolling something out. Most of them are not about robots or runaway intelligence. They are about data, accountability, suppliers, energy, and who is on the hook when something goes wrong.

The first question is who is accountable

The question that sits behind almost every other concern is the same one: when the AI gets it wrong, who carries the cost?

In the UK, the answer is usually the business that deployed it. Regulators have been clear that outsourcing a decision to a system does not outsource the responsibility. If an AI tool sends the wrong invoice, mishandles a customer's data, gives bad advice, or refuses someone unfairly, the supplier of the tool is rarely the one defending it.

This is a reason to design any AI workflow with a named human owner, a named reviewer for the risky steps, and a clear record of what was decided by whom. If a tool cannot tell you who reviewed the action that just went out of the building, fix that before adoption, not after.

Data risks hide in supplier terms

Small businesses tend to be protective about customer information in conversation, then quite relaxed when it comes to which AI tool sees it. The questions that matter are not exotic. Where is the data going. Who can see it. Whether it is used to train a model. How long it is kept. Where in the world it lives. Whether you can delete it cleanly.

Under UK data protection law, the business deploying the AI is the data controller, and that is the role with the legal duty. The ICO's new code of practice on AI and automated decision-making comes into force on 12 May 2026, and its direction is clear: be able to show what your AI does, what data it uses, and how someone can challenge a decision made about them.

For most small businesses the practical answer is mundane. Use business plans rather than consumer ones, because the data terms are usually better. Decide which classes of customer information are not allowed to leave your systems at all, and write that down where the team can see it. One careless paste into the wrong chat window can be the most expensive mistake a small team makes this year.

Made-up answers are not edge cases

The single most consistent thing about AI in 2026 is that it still gets things wrong with confidence. Independent research keeps putting error rates on real tasks somewhere between low single digits on basic summarising and frighteningly high numbers on legal or medical specifics. The Stanford AI Index has the same general shape, and so does OWASP's top-ten list of AI risks.

For small businesses, two things follow. In any task where being wrong has a cost, AI output needs to be reviewed by someone with the context to spot the mistake: anything sent to a customer, anything quoted as a price, anything that names a person, anything that summarises a contract, anything that gives professional advice. The model does not always know when it is guessing.

The other useful decision is to anchor the AI to real source material. An assistant that answers from your own product information, your own documents, and your own historical decisions will make fewer things up than one drawing on a general knowledge base. That is the difference between a tool a customer-facing team trusts and one they quietly stop using.

Provider lock-in can become expensive

A second slow-burning concern is that the easiest AI tools to adopt are usually the most tightly bound to one provider. The chat product, the agent builder, the connectors, the memory, the file system, the billing, all the same logo. It feels efficient until the day the provider changes its pricing, retires a feature, or quietly degrades a capability the workflow depends on.

The useful buying heuristic is to prefer designs that treat the AI provider as replaceable. Can the data be exported in a format another tool can use, or is it locked behind a proprietary surface. Is the workflow rebuildable on another provider, or does it depend on a specific feature only one offers. The switching cost is rarely the licence fee. It is retraining the team, rebuilding integrations, and re-establishing trust in a new tool, and that cost goes up the longer the workflow runs.

Better tools come along quickly. The team that is not locked to one provider can take advantage of them. The team that is locked usually waits, complains, and pays.

Prompt injection is now a practical risk

This is the concern small businesses are least likely to have heard about and most likely to be surprised by.

When an AI agent reads emails, browses the web, processes files from customers, or pulls in supplier documents, it is exposed to text written by someone outside the team. That text can include hidden instructions designed to make the agent behave differently. OWASP has had prompt injection at the top of its AI risk list for two years running, and by early 2026 researchers were documenting these attacks happening in the wild, including against live commercial platforms.

The practical implication is simple. The more an AI tool can do automatically, the more careful the design has to be about what it is allowed to read, what it is allowed to act on, and which actions need a human to confirm. Keep destructive actions, anything that sends, edits, deletes, refunds, posts, or pays, behind a review step. Be wary of anything that auto-acts on documents or web pages that come in from outside the business.

You do not need to become a security expert to take this seriously. You need to ask, of any new AI tool, three questions: what can it touch, what can it change, and what stops it doing something stupid.

Check a supplier's energy and water use

The environmental cost of AI is real. IEA projections put data-centre electricity demand well over 900 TWh by 2030, and the UK government's recent report on water use in data centres made the same point in plainer terms: this is not free.

For a small business this is not a reason to avoid AI, any more than avoiding email is a sensible response to electricity use. It is a reason to ask suppliers about reporting and efficiency, to be sceptical of vendors that hide the topic, and not to leave heavy AI workflows running on a schedule for tasks that are not actually being read. The cheapest, greenest AI work is the work nobody needed in the first place.

Bias is a concern when AI decides about a person

AI concerns escalate quickly when the tool is helping decide about a specific person: hiring, lending, pricing, fraud flags, refusing service, ranking customers. The ICO's recent work on automated decision-making in recruitment is a clear sign of where regulation is heading. Where a person is the subject of the decision, evidence, oversight, and a route to challenge are not optional.

For most small businesses this rules out handing off decisions about people to a model without a person reviewing them and a record of why. The risk of getting it wrong is too high to save time on the review.

AI can put quiet pressure on staff

The last concern is human, and the one that gets least attention in vendor decks. Introducing AI into a small team can quietly change the way work feels: people unsure whether they should be using the tool, whether their work will be measured against it, whether saying "I am not sure" is still allowed. The risk is a team where everyone pretends they are using AI well, nobody admits when it has produced a poor answer, and anxiety builds underneath the productivity claims.

The response is the same as the rest of this list: name a reviewer, write down what is allowed, make it safe to question the assistant, and say out loud that the team is not expected to be replaced by the tool.

Use these risks as buying questions

These concerns are not arguments against using AI. We use it ourselves, and the pattern is consistent: the businesses that get value from AI are the ones that took the concerns seriously enough to design around them. They wrote down what could leave their systems. They named who reviewed what. They kept their workflows portable enough to swap providers when better tools appeared. They put the risky actions behind a person.

The businesses that ended up regretting their AI investment did not run into a futuristic problem. They ran into a boring one: a workflow nobody owned, a data leak nobody anticipated, a vendor change nobody planned for, a customer interaction the model handled badly that nobody caught until later.

That is the version of AI risk worth attention.

If you want help working out which of these concerns apply to a tool you are considering, that is the kind of conversation an AI Audit is for. We will tell you which risks are real for your situation, which are noise, and which need to be designed in before the tool goes anywhere near a customer.