You have been using ChatGPT to think through a pricing decision. You paste in the scenario, talk it through with the model, get a sharper sense of your options. It feels useful. Then someone mentions data protection, or a client asks whether you use AI on their account, and you find yourself wondering whether you should be doing this differently.
That question is worth taking seriously. About a third of owner-managed businesses in the UK now use AI tools, according to a 2024 YouGov survey of over a thousand business decision-makers. A good share are using it for exactly this kind of work: ideas, options, decisions. Very few have thought clearly about which tool suits which kind of work, or what the governance boundary between them actually looks like.
What choice are you actually facing?
Whether you reach for ChatGPT, Claude, Gemini or Microsoft Copilot, the underlying choice is the same: a public cloud service where your inputs are processed on a third party’s infrastructure, or a controlled arrangement where you have clearer oversight over how your data is handled. Getting this right turns on what you are asking the AI to help with, not which provider you prefer.
Public cloud tools are immediately available, low-cost, and capable of a wide range of everyday thinking work. Their advantage is speed and access to broad knowledge. Their constraint is that your prompts pass through the provider’s infrastructure, and depending on which tier you are on and which settings you have enabled, that data may be used to improve their models.
Controlled or private arrangements, including enterprise tiers with training disabled, Azure OpenAI hosted in a UK or EU region, or a self-hosted open-source model such as Llama 3, give you more governance over data residency and logging. They are also more expensive and require IT resource to configure and maintain.
The UK Government’s 2023 AI White Paper takes a deliberately light-touch approach, directing existing regulators such as the ICO, FCA and CMA to oversee AI use within their sectors rather than creating new AI-specific rules. For owner-managers, that means UK GDPR data protection rules are the primary constraint to understand first.
When is a public cloud tool the right option?
If you are using AI to explore strategy, test a pricing idea, draft communications or work through a decision that involves no client-identifiable data, a public cloud tool on a paid plan is proportionate. The paid tier matters: ChatGPT Plus, Claude Pro and Gemini for Workspace all offer settings that prevent your prompts from being used to train the provider’s model.
The YouGov data is instructive here. Among UK businesses already using AI, 54% apply it to task automation and 45% to marketing and advertising. A significant share of that work involves no sensitive personal data. For tasks in this range, the governance overhead of a private deployment is disproportionate to the risk.
The practical guardrails are straightforward. Use a paid account. Enable the opt-out from training data use: all major providers offer this on business accounts. Avoid pasting in names, addresses, financial account details, health information or commercially sensitive material specific to an identifiable person or client. Treat the AI like a well-informed generalist who works from a shared office: useful for thinking, not the place for a confidential file review.
The risk worth calling out is over-reliance. A 2023 Stanford and MIT study found that professionals using GPT-4 completed knowledge work tasks 55.8% faster on average, but some reduced their own critical checking as they grew more comfortable with the tool. The model works best when you treat its output as a first draft of an idea, not a settled conclusion.
When do you need a more controlled setup?
When your work involves client-identifiable data, financial accounts, health information or processes inside a regulated sector, you need a controlled or enterprise-grade arrangement with clear data handling contracts. The ICO’s guidance on generative AI requires organisations to have a lawful basis, appropriate contracts and transparency before sending personal data to any external AI platform.
In practice, this means any scenario where you want AI to analyse client files, run figures from your accounts, review employment records or support decisions that affect identifiable individuals. For higher-risk processing, a Data Protection Impact Assessment is required under UK GDPR before you start.
The alternatives are practical. Enterprise tiers from OpenAI, Microsoft and Google offer contractual commitments on data handling, including data processing agreements, the option to disable training on your inputs, and in some cases geographic data residency within the UK or EU. These arrangements carry higher minimum costs, typically running to several thousand pounds per year, and require IT resource to configure.
For businesses in financial services, the FCA’s guidance is clear: firms remain fully accountable for decisions made using AI tools. Delegating a regulated judgement to an AI model does not transfer the regulatory responsibility. That applies whether you are using a consumer tool or an enterprise deployment.
What does it cost to get this wrong?
Getting this choice wrong carries two distinct costs. The regulatory one: pasting client-identifiable data into a public AI tool without a lawful basis and appropriate contracts is a potential UK GDPR breach, regardless of whether anything goes publicly wrong. The operational one: treating AI output as a decision rather than an input produces faster conclusions, and periodically the wrong ones.
On the regulatory side, the ICO’s enforcement record is a useful reference point. When Clearview AI scraped images of UK residents without lawful basis, the ICO issued a £7.5 million fine and required the firm to delete the data. That case involved deliberate scraping at scale, not a business owner’s chat session. The underlying principle holds: personal data processed without lawful basis, proper contracts and adequate transparency creates real exposure.
On the operational side, the Stanford and MIT research found that while GPT-4 improved average performance on knowledge work tasks, users who relied heavily on the model without checking its outputs produced confident but incorrect results in a meaningful share of cases. Ethan Mollick, writing for HBR, has noted the same pattern: performance gains are highest when users treat AI as a tool to question, not one to defer to.
What should you ask before committing to any option?
Before you settle on a tool or a setup, four questions do the substantive work. What decisions are you actually asking AI to help with, and do they involve personal or commercially sensitive data? Where does the provider process and store your inputs? Can you get a data processing agreement? And what human-in-the-loop rules will you follow before acting on anything the AI recommends?
On data handling: ask any candidate provider where data is stored and processed, whether prompts are used to train their models and whether you can opt out, and whether they will sign a data processing agreement. If a provider cannot answer these questions clearly, treat that as a signal about their suitability for business use.
On integration: if your team runs on Microsoft 365, Copilot sits closer to the tools they already use. Google Workspace-centred businesses will find Gemini a more natural fit. If you mainly want a browser-based sparring partner for your own thinking, a standalone subscription is simpler to set up and easier to govern.
On decision governance: decide in advance which categories of decision AI may input to, and which must remain human-owned. Irreversible decisions such as senior hires, structural pricing changes, acquisitions and changes to client terms should stay with you. AI can be one voice in the room. The deciding one should not be.
The NCSC’s guidance on safe use of public generative AI is worth reading once if you have not done so. It sets out, in plain language, what staff should and should not share with public AI tools. One page of internal guardrails based on it covers the governance overhead this kind of use actually requires. After that, the thinking-partner question is a capability question: which tools are genuinely useful for which kind of work, and how do you use them well.
