Last month, a small services firm owner found out that her best writer had been pasting entire client briefs into a public AI tool every week for six months. There was no policy either way. The client was a law firm.
Nothing bad had happened, as far as anyone knew. But the question she couldn’t answer was whether the data had been stored, shared, or used to train the model. Nobody had thought to ask the vendor.
That gap between what staff are using and what the business has actually assessed is where avoidable AI risk lives.
What counts as avoidable AI risk for a small firm?
For a 5-50 person services firm, the main AI hazards are governance-level: data leaks from unvetted tools, decisions taken on unchecked output, bias in recruitment or performance management, and non-compliance with existing rules. UK GDPR is the primary legal framework for AI use. Sector regulators such as the FCA add further obligations in regulated industries. Treat AI like any other high-impact software that handles client or staff data.
The ICO has already investigated firms for unlawful use of algorithms in recruitment and credit decisions, focusing on fairness, transparency, and data protection. The FCA has confirmed that AI use in financial services must still meet existing conduct, governance, and operational resilience rules. There is no AI exemption in either framework, which means the question is never whether existing rules apply, only how they apply to your specific use case.
The practical implication is the same as for any other high-impact system: define your use cases, assign accountability for outputs, test before you rely on results at scale, and document the decisions you make along the way. You don’t need an AI strategy department. A founder plus one operational lead is enough to own this, working from public regulator guidance.
Why does unmanaged AI use create bigger exposure than you might expect?
A 2024 Microsoft study found 71% of UK employees are already using AI tools at work, frequently without employer approval. The NCSC warns that prompts entered into public AI tools may be retained and used to train models, putting sensitive client information at risk. Assume AI is already in your business. Your first safeguard is a clear policy and an approved tools list.
The instinct to ban AI tools tends to push the behaviour into the gaps where you cannot see it. A clear policy with an approved list keeps use visible and manageable, and it gives staff a legitimate route rather than an informal one.
The assumption that small firms are too small to attract regulatory attention does not hold. The ICO fined Clearview AI £7.5m in 2022 for scraping facial recognition data without consent. That case involved a large-scale data operation, but the underlying principle applies at every scale: unlawful data processing attracts scrutiny regardless of firm size. Opportunistic cyber attacks on UK SMEs follow the same pattern, as the NCSC’s annual guidance consistently confirms.
The HSBC small business guide draws a useful practical conclusion: start with AI for internal productivity tasks, keep human oversight on any client-facing content or advice, and choose tools with clear data-handling policies from the outset.
Where does AI risk actually show up in a services business?
AI risk tends to concentrate in three places for a services firm: client-facing content that hasn’t been independently checked, people decisions that lack genuine human review, and data flows where client or staff information reaches unvetted tools. The first creates errors in deliverables. The second triggers GDPR automated decision-making obligations. The third is a quiet data protection exposure.
On the content side, the CMA and NCSC have both highlighted hallucinations as a material risk. Generative models can produce fluent but wrong answers. For a financial promotion, a technical specification, or a regulatory submission, an unchecked error can carry legal consequences as well as reputational ones.
On the people side, UK GDPR gives individuals the right not to be subject to decisions with significant effects where those decisions are made solely by automated means. The Court of Appeal addressed this directly in the 2023 Uber drivers case, affirming the need for genuine human review rather than nominal sign-off. The ICO’s AI recruitment guidance makes the same point: automated profiling can entrench bias and must be accompanied by meaningful human involvement and transparency to candidates.
On the data side, any AI vendor that processes personal data on your behalf is a data processor under UK GDPR. That means a data processing agreement, a check on where data is stored, and confirmation that your inputs are not used to train the model.
When do you need formal safeguards versus informal oversight?
The line falls between reversible and significant. Drafting internal communications, summarising meeting notes, and generating ideas are low-risk tasks where a read-through before sending is enough. Any process that affects a person’s employment, financial position, or access to services, or that constitutes regulated advice, needs a named reviewer, a documented check, and a record that a human signed off. UK GDPR gives individuals the right to human intervention in decisions with significant legal effects.
The ICO’s guidance warns explicitly against rubber-stamping automated outputs. A human reviewing an AI recommendation without the capacity to understand, challenge, or override it does not meet the meaningful oversight standard. The FCA takes the same position for regulated firms: boards remain responsible for outcomes and must ensure adequate skills and genuine challenge around the AI models they use.
A useful rule of thumb before deploying AI in any process: ask whether an error in the output could significantly affect someone’s livelihood, safety, or legal rights. If yes, build in a human sign-off step before the output leaves the business, and keep a record that the step happened.
What are the practical safeguards that reduce risk from day one?
The safeguards that matter for a 5-50 person firm fit on one page and don’t require a specialist team. A short AI policy sets which tools are approved, what data types cannot be entered, and who reviews outputs before they inform decisions. Add a vendor check for data hosting and model training terms, basic technical hygiene on access controls, and a one-hour staff briefing.
UK and SME-focused guidance converges on a practical sequence. Map candidate use cases first and prioritise low-risk, internal tasks. Check whether personal or confidential data is involved in each one. Write a one or two page policy before rolling out to the whole team and reinforce it through induction. Choose vendors that can answer clearly on data hosting, model training, and retention. Enable multi-factor authentication on AI tools and admin accounts, and use role-based access so staff can only connect AI to the data they genuinely need. Run a short pilot with a small group before scaling. Review the whole set-up every six to twelve months, since both the tools and the regulatory landscape will change.
On the EU AI Act: if your firm recruits in the EU, provides digital services into EU markets, or relies on EU-based AI vendors, expect Act obligations to appear in due-diligence questionnaires from larger clients within the next year or two. The UK government has signalled a pro-innovation approach that relies on existing regulators rather than new AI-specific legislation, but EU-facing firms should watch for deployer requirements around transparency and record-keeping for high-risk systems.
The firm owner from the opening scenario did something straightforward when she found out about the ChatGPT use. She started a conversation with the team, and that conversation became a short policy, an approved tools list, and a vendor review of the tool already in use. Getting ahead of avoidable risk takes a conversation, a document, and a decision.
