Your finance director has a point. The team is already using ChatGPT’s free tier, they’re getting things done faster, and the budget conversation is uncomfortable. Why pay £20 per person per month for something that appears to be working?
The answer sits in the terms of service. Free tiers of the major AI platforms default to using your team’s inputs to train and improve their models. Whatever an employee types in, client briefs, internal pricing, supplier contracts, board materials, can become part of the model that improves responses for everyone else on that platform. Including your competitors. The licence fee is not the real cost. The data is.
What does the free tier actually do with your team’s inputs?
Free tiers of the major AI tools default to training on user inputs. ChatGPT’s free tier and Google’s Gemini free tier use what employees type to improve their models. An opt-out exists, but it requires deliberate action each session. The default state, unless someone actively changes it, is that your team’s inputs contribute to a training pool shared across all users of that tool.
In 2023, Samsung engineers used the free ChatGPT tier to complete work tasks and inadvertently pasted semiconductor design specifications, source code, and internal meeting notes into the interface. That information entered OpenAI’s training data. Samsung’s response was to restrict ChatGPT to commercial tiers with data privacy protections. The governance failure was the absence of any policy specifying which data was allowed into which tool.
Your team is doing exactly what Samsung’s engineers did, using the tool available to get work done faster. The question is whether the data they are typing into it should be anywhere near a public training pipeline.
Why does this cost more than the licence fee?
The visible cost of free AI tiers is zero. The invisible cost is the data you hand over. A commercial tier with data privacy protections runs roughly £15 to £30 per active user per month, so a team of four regular AI users costs £60 to £120 a month. A single regulatory investigation or professional indemnity claim triggered by a data exposure would dwarf that figure.
For businesses in regulated sectors, the calculation is more specific. The Solicitors Regulation Authority has clarified that feeding client matter information into a public AI tool without explicit consent and a Data Processing Agreement constitutes a breach of professional duty and a likely UK GDPR violation. The Financial Conduct Authority holds similar expectations for firms using AI in client-facing processes. The General Medical Council does the same in healthcare.
For many owner-managed businesses in professional services, using a free tier for confidential client data amounts to a professional standards breach, dressed as a cost-saving decision.
Where will you actually meet this decision?
The choice shows up in specific, everyday moments. A team member drafting a proposal pastes in a client’s requirements. Someone in finance formats a board report with actual revenue figures. The operations lead summarises a supplier negotiation. Each of these tasks is exactly what employees reach for AI to help with, and each involves information that has no place in a public training pool.
A practical data classification separates content into three tiers. Public data is anything the business or its clients already intend to make available: published marketing material, public research, website copy. This can go into any tool, including free tiers, because there is nothing confidential at risk.
Internal data is what the business uses day-to-day but does not share externally: meeting notes, financial forecasts, strategy documents, internal processes. This should only go into a paid commercial tier where the vendor has a signed Data Processing Agreement and a contractual no-training commitment.
Confidential data is information provided by clients in confidence: matter notes, financial records, unpublished plans. This should only be handled by a tool with explicit client consent and a DPA in place, or kept on an on-premise system that stays within your infrastructure.
The classification is simple enough to summarise in a one-page table. The harder part is making it a team habit rather than something people assume someone else has already sorted.
When does paying for a commercial tier become the right call?
Pay for a commercial tier the moment your team starts putting internal or confidential data into an AI tool, and that moment is probably already here. For three or four daily AI users, the cost is roughly £45 to £120 a month. Document the decision either way: if you choose to stay free, write down which data your team is and is not allowed to input.
The documented decision matters because it draws a line between a governance choice and an oversight. If a data incident occurs and you have never committed to paper which AI tier your team uses and what data is permitted in it, the absence of that record becomes evidence that governance was not in place. If you have written it down, even in an email or a shared document, you have demonstrated that the decision was a considered one.
For businesses where the team uses AI only for brainstorming with publicly available information, free tier use is defensible. A marketing agency generating copy ideas from published sources is in a different position from a law firm where paralegals are summarising client files. The line is about data, not about whether you can afford the licence.
What else do you need to check before committing to a tier?
Choosing a paid tier is the starting point. Before you commit, get clear answers on three things from any vendor. Do they offer a Data Processing Agreement you can sign? Have they committed not to train on your inputs? Does data stay within the UK or EU? The major providers, OpenAI, Google, Anthropic, and Microsoft, offer these protections on commercial tiers, though which tier includes them depends on the provider.
Getting a contractual no-training commitment from OpenAI requires the API commercial tier or an enterprise agreement, not the Plus subscription. Gemini Business, part of Google Workspace, includes a training opt-out as standard. Anthropic’s Claude for Work includes a business agreement with data handling commitments. The tier naming changes; the question you ask stays constant: does this tier include a signed DPA and a no-training commitment?
If your business processes personal data systematically through AI tools, the ICO’s guidance indicates that a Data Protection Impact Assessment is often required before deployment under UK GDPR. That is a 1 to 2 hour document, not a month-long project, but it needs to happen before the tool goes live rather than after the first incident.
The rule is simple enough to state. Public data can go anywhere. Internal and confidential data goes only to a paid tier with a DPA and training disabled. Write the choice down, whichever way it lands, so it is deliberate rather than default.
