A consultant at a small professional services firm spent an afternoon using a free AI chatbot to improve client reports. The outputs looked good. The problem came later, when she realised she had named the clients and pasted in their financial figures, and the tool had no contractual commitment about what the provider does with input data. No policy existed to check. Nobody had sanctioned the account.
That situation is more common than the alternatives. The ICO has stated clearly that AI carries no exemption from UK GDPR: the same principles that apply to any other form of data processing apply the moment a staff member pastes client information into a chatbot. Your firm is the data controller regardless of whose server the tool runs on.
The practical question is what to do about it.
What does “safer AI use” actually mean?
“Safer AI use” means running AI tools in ways that keep your firm in control of where data goes, who can use it, and what happens to it afterwards. The goal is the productivity gain without unnecessary exposure. The ICO, NCSC, and UK data protection specialists all point to the same baseline: an approved tools list, a short written policy, and business-grade accounts rather than free personal ones.
That baseline is reachable for a 5-50 person services firm without a dedicated compliance team. The work is getting the sequence right.
Why does data exposure matter even for a small services firm?
The ICO applies UK GDPR regardless of company size. A 12-person accountancy practice carries the same legal obligations when processing personal data as a large institution. The penalty range reaches up to £17.5 million or 4% of global annual turnover for serious infringements, including inadequate security and unlawful processing. Those figures sit well above what a typical services firm could absorb without significant disruption.
The practical risk for a small services firm goes beyond a fine. Staff using free AI accounts without a policy can expose client information in ways that damage trust before any regulator gets involved. Mishcon de Reya, advising UK law firms, has made the point that uploading client documents to public generative AI tools can breach confidentiality obligations and UK GDPR simultaneously.
There is a quality risk alongside the compliance risk. A US law firm was sanctioned in 2023 after its lawyers submitted court filings containing fabricated case citations generated by ChatGPT. The filings looked right until the opposing side tried to find the cases. Verifying AI output before it reaches anyone outside the firm is the minimum standard for responsible use.
Where are the main exposure points in everyday AI use?
The two main exposure points in a typical services firm are free consumer AI accounts used without business controls, and staff pasting raw client or staff data into prompts without removing identifiable details first. Both are addressable. The harder one is the accounts problem: staff will reach for the tools they already use at home, and those are usually free-tier tools with less protective data terms.
Business-grade accounts change the picture. OpenAI’s ChatGPT Team and Enterprise plans include a contractual commitment that inputs are not used to train models by default. Microsoft 365 Copilot processes data within the customer’s Microsoft tenant, subject to enterprise terms. These commitments matter because, as the data controller, your firm needs a clear answer to where client information goes after a prompt is submitted.
Data minimisation also reduces exposure at the point of use. UK GDPR requires that personal data processed is limited to what is necessary. The practical application is straightforward: before pasting anything into an AI tool, consider whether the client can be anonymised, the figures rounded, or the name removed. A prompt asking for help summarising a client situation rarely needs the client’s name to produce a useful answer.
The Samsung case from April 2023 shows the speed at which this can go wrong. Engineers pasted proprietary source code and internal meeting notes into free ChatGPT. The company subsequently had to restrict employee use of generative AI tools across the board. The data had already left. A written policy and a business-grade account, introduced before that first use, would have changed the outcome.
When does the level of risk change?
The risk level shifts based on what the AI is doing, not how sophisticated it looks. Using AI to draft emails, summarise documents, or generate marketing copy sits at the lower end of the compliance spectrum. Using it to screen job applicants, assess customer creditworthiness, or price individual clients sits at the higher end, where UK GDPR’s automated decision-making rules apply and documented controls become mandatory.
The ICO uses Data Protection Impact Assessments to govern higher-risk uses. A DPIA is required, rather than recommended, when AI processing is likely to result in high risk to individuals. The ICO publishes an AI risk toolkit specifically for this purpose. For small services firms starting out with assistive AI, the DPIA question won’t arise immediately, but knowing the threshold helps you recognise when you’re approaching it.
The EU AI Act is also relevant for firms with EU-based clients. It classifies AI used in areas including employment, credit scoring, and access to essential services as high-risk, with stricter obligations around human oversight and documentation. UK financial services firms are watching the FCA’s approach to AI model governance as a directional signal for where broader expectations are heading.
What should you put in place first?
The sequence that UK data protection specialists recommend for a small services firm starts with a short written AI policy covering which tools are approved, what staff must never paste into any AI system (client names, financial data, HR records), and who to contact when someone wants to try a new tool. UK-specific templates are available; you do not need to write it from scratch.
The second step is standardising on a small number of business-grade accounts. Choose your tools deliberately rather than letting the business accumulate a long list. Ensure the “do not train on my data” option is active where available, and turn on logging and access controls. Microsoft 365 Copilot and ChatGPT Team are the most common starting points for UK services firms, and both offer contractual data protections that free accounts do not.
Third, brief your staff. This does not need to be a formal course. A short session covering what they can and cannot paste into AI tools, how to check AI-generated content before it reaches a client, and where to raise questions addresses the large majority of practical risk.
Fourth, run a Data Protection Impact Assessment for any AI that touches personal data at scale or influences decisions about specific individuals. The ICO’s AI risk toolkit guides you through the documentation. Fifth, revisit the whole setup every six months: what tools have staff adopted since your last review, and does the policy still reflect what is actually happening?
The question many founders reach for is “how do I use AI without sharing data?” The honest answer is that you manage what is shared rather than eliminating it entirely. A client’s name in a prompt is not automatically a problem if the account has the right terms, the firm has a written policy, and staff know the rules. The work is building the controls that make that judgement possible.
