Most businesses have not yet decided whether their staff should be using AI tools with company data. The problem is that their staff have already decided for them.
ChatGPT, Copilot, Gemini and a growing list of similar tools have become a routine part of how people get work done. Drafting emails, summarising documents, preparing for client meetings, writing proposals, analysing data. The tools are fast, capable and, from the user’s perspective, entirely unremarkable. They are just another way of getting something done.
What most users do not consider, and what most businesses have not addressed, is what happens to the information they put in.
What the risk actually is
The concern is not primarily that AI providers are malicious or careless with data. The concern is more practical than that.
When a member of staff pastes a client contract into ChatGPT to ask it to summarise the key terms, that text is submitted to an external system outside your control. Depending on the tool, the version being used and the account settings in place, that text may be used to train future models, may be retained by the provider, may be accessible to the provider’s staff for safety and quality review, or may be stored in ways that are subject to data protection laws in jurisdictions other than the UK.
The same applies to financial data pasted in to build a model, HR information submitted to draft a sensitive communication, technical specifications shared to generate documentation, and pricing or commercial information used to prepare a pitch.
None of this requires a breach, a hack or any malicious intent. The data simply leaves your control the moment it is submitted, and in most SMEs nobody has assessed whether that is acceptable.
Why this is harder to manage than it looks
The difficulty with AI data risk is that the tools are genuinely useful. Banning them outright is both practically ineffective and commercially counterproductive. Staff will use them anyway, through personal devices or personal accounts, and you will simply lose visibility rather than reducing the risk.
The second difficulty is that the risk is not uniform. Submitting a generic marketing brief to an AI tool carries a different risk profile from submitting a due diligence report, a personnel file or a client’s financial statements. Most governance approaches treat all AI use as equivalent, which means they either overreact to low-risk activity or underreact to high-risk activity.
The third difficulty is pace. New tools emerge constantly, existing tools add AI features without announcement, and the default settings of tools your business already uses may have changed without anyone noticing. The Microsoft 365 suite in particular has introduced AI capabilities across Word, Outlook, Teams and Excel that many businesses have not consciously adopted but are already using.
What good governance looks like
The starting point is visibility. Before you can govern AI use, you need to understand what is already happening. That means asking directly, reviewing software in use across the business, checking the settings of tools already deployed, and accepting that the answer will be more extensive than most leaders expect.
The second step is classification. Not all data carries the same sensitivity, and not all AI tools carry the same risk. A governance framework that distinguishes between genuinely sensitive data, such as client personal information, commercially confidential material, financial data and HR records, and lower-sensitivity material gives staff a workable basis for decision-making rather than a blanket prohibition they will quietly ignore.
The third step is tooling. For businesses that want to allow staff to use AI tools productively while controlling what data is submitted, enterprise versions of the major tools offer substantially better data protection terms than free or consumer accounts. Microsoft 365 Copilot, ChatGPT Enterprise and Google Workspace with Gemini all operate under contractual terms that prohibit training on your data and offer stronger retention controls. The cost is meaningful but for most SMEs it is recoverable in the productivity benefit alone, and the governance position it creates is significantly more defensible.
The fourth step is policy. A written AI use policy does not need to be long or complex, but it does need to exist. It should cover which tools are approved for which types of use, what categories of data must not be submitted to external AI tools under any circumstances, what the approval process is for using a tool that is not on the approved list, and what the reporting expectation is if something goes wrong. Most SMEs do not have this in place. Having it means you have a defensible position under UK GDPR, a clear expectation for staff, and a governance basis for your insurers.
The regulatory dimension
UK GDPR places clear obligations on organisations that share personal data with third parties. An AI tool that processes personal data on your behalf is a data processor under that framework, which means you need a Data Processing Agreement in place, you need to have assessed that the tool meets appropriate security standards, and you need to be able to demonstrate that you have done both.
Most businesses using consumer AI tools with personal data have none of this in place. That is not a theoretical risk. The ICO has been increasingly active in its scrutiny of AI-related data processing, and the position “we did not realise the tool was processing personal data” is not a defence.
The same consideration applies to professional and sector-specific obligations. Legal firms, financial services businesses and healthcare adjacent organisations carry additional confidentiality requirements that extend well beyond GDPR. A solicitor asking an AI tool to draft a letter using client case details, or an accountant summarising a client’s financial position, may be breaching professional conduct obligations regardless of whether personal data is technically involved.
A practical starting point
For most SMEs the right starting point is a short audit of current AI tool use, followed by a classification of what data the business holds and which categories of that data warrant specific controls.
That work does not need to be lengthy or expensive, but it does need to be done by someone who understands both the commercial context and the data governance obligations. Leaving it to IT support without board involvement tends to produce either an unworkable blanket ban or a policy that exists on paper but has no practical effect.
The businesses that handle this well treat it as a governance question first and a technology question second. The question is not which tools to allow. The question is what data you are responsible for protecting, what your obligations are to the people and organisations whose data you hold, and what controls you need to have confidence that those obligations are being met.
AI tools are not going away, and the productivity case for using them is real. The governance case for using them properly is equally real, and for most SMEs it remains unaddressed.
If your business uses AI tools and you are not confident that your data governance keeps pace with that use, Northstar can help. We work with SME leadership teams to assess AI-related risk and put proportionate controls in place.