Three teams. All using AI tools to get through the day faster. Two of them pasting client notes into a consumer chatbot to generate summaries. One pulling a contract draft into a free tool for a quick tidy-up. Nobody was trying to cause a problem. The issue was simpler. Nobody had ever told them which data could go where.
A data classification table fixes that. One page. Four tiers. A column for each tool category. The daily decision becomes visible before anyone opens a browser tab.
What is a four-tier data map?
A four-tier data map sorts everything your business holds into one of four categories. Public is anything already in the public domain. Internal is the operational material used inside the firm. Confidential is what clients or partners have shared in trust. Restricted sits at the tightest end, governed by UK GDPR Article 9. Each category carries a rule about which AI tools may process it.
Professional services businesses encounter all four tiers on the same working day. A piece of marketing copy is public. The strategy document that preceded it is internal. The client brief that shaped the campaign is confidential. A health disclosure that came up during an HR conversation is restricted. Four categories, not because the framework demands it, but because the regulatory and contractual obligations that attach to each one are genuinely different.
Public data carries no confidentiality risk and typically involves no personal data. Free AI tools, including consumer tiers that train on user inputs by default, are appropriate here. Drafting a social media post from a publicly available report, generating a case study template, brainstorming ideas from market data already published online. All safe.
Internal data carries no third-party confidentiality obligations, though it may include employee personal data and commercially sensitive material. A paid commercial AI tool with a Data Processing Agreement is appropriate for this tier, meaning the vendor has contractually committed not to train on your inputs and has specified where your data is stored. The major commercial providers now offer this at their paid tiers.
Confidential data carries one additional condition. The third party whose information it is must have approved the AI processing. An updated engagement letter or explicit consent, alongside a DPA with the vendor, is the minimum. For a law firm, this is where client matter files sit, and the Solicitors Regulation Authority’s guidance on AI in legal practice applies directly.
Restricted data under UK GDPR Article 9 sits at the strictest end of the scale, and external AI tools are off the table in almost every case.
Why does it matter to your business?
In 2023, Samsung engineers used free ChatGPT to complete work tasks and inadvertently shared semiconductor design specifications, source code, and internal meeting notes. The free tier trained on those inputs, absorbing proprietary information into its model. The engineers were working normally. The governance failure had a single root cause. Nobody had told them which data could go into which tool.
The UK Information Commissioner’s Office has made clear that adopting AI does not exempt businesses from UK GDPR obligations. When AI processes personal data, several requirements activate immediately. Lawful basis under Article 6 must be established, privacy notices must be updated to disclose AI processing, and a Data Protection Impact Assessment is likely required under Article 35 if the processing poses high risk to individuals. These obligations are triggered by the act of adopting the technology, not by the size of the firm.
For a team of fifteen or twenty people, the exposure is comparable to Samsung’s, scaled down in visibility but not in consequence. An employee who pastes a client brief into a free consumer chatbot may have breached a professional confidentiality obligation, violated UK GDPR, and disclosed commercially sensitive information, all in the time it took to generate a three-paragraph summary. A classification map makes that error structurally impossible before it happens.
Where will you actually meet the tier question?
The classification decision comes up every time someone reaches for an AI tool. A proposal draft where the market overview is public but the client brief is confidential. A board meeting summary that is internal. An employee sickness record that is restricted. The tier decides which tool is acceptable and whether a Data Processing Agreement needs to be in place.
The free tier versus paid tier question is where many teams get caught. Consumer versions of the major AI tools, the ones that cost nothing, default to training on user inputs. The paid commercial tiers of the same tools offer the data protection commitments that make them appropriate for internal data. The practical test is simple. Does the vendor have a current DPA, a commitment not to train on your inputs, and clarity on where the data is stored? If the answer to any of those is no, the tool belongs in the public-tier column only.
A one-page table removes the daily calculation. Down the left side, the four tiers, each with two or three examples drawn from your own business. Across the top, four tool columns, one each for free tools, paid commercial tools with a DPA, approved specialist platforms, and on-premise or locally-run models. Each cell gets a yes or a no. A team member who is uncertain knows which row their data sits in and which column their tool occupies. The answer is in the cell.
When is data off-limits for external tools?
Restricted data under UK GDPR Article 9 covers health records, biometric data, genetic information, racial or ethnic origin, religious belief, sexual orientation, trade union membership, and criminal conviction data. For this tier, external AI tools are off the table in almost every scenario. The law requires explicit legal basis, additional conditions for processing, and exceptional technical safeguards before any automated handling can proceed.
Healthcare practices face this most directly. Patient health records are restricted by definition. Any AI system processing patient data must have a Data Processing Agreement, a training opt-out, GDPR compliance, and, for US patients, a HIPAA Business Associate Agreement. Many cloud-based AI services cannot satisfy all of those requirements. The practical default is that patient data stays on-premise or within a healthcare-specific system built explicitly for this compliance profile.
The same logic applies to HR functions in any business. An employee’s sick leave history, disability disclosure, or diversity monitoring data is restricted under Article 9. An AI tool summarising HR records for workforce planning is processing special category data and requires a lawful basis that goes beyond legitimate interest. The restricted tier rule is straightforward. If in any doubt, the data stays out of external systems until you have legal advice confirming otherwise.
What connects the map to your broader AI governance?
The tier map works alongside three other governance steps. A Data Processing Agreement with your AI vendor locks in the protections that make the confidential tier workable. A Data Protection Impact Assessment is legally required under UK GDPR Article 35 before processing that poses high risk to individuals. A shadow AI survey surfaces which tools your team already uses, so the map reflects your actual current state.
The DPA is not complicated to obtain. Every major AI provider now offers a standard template at their commercial tier. Ask the vendor three things. Do you have a current DPA, do you commit not to train on our inputs, and where is our data stored? If they cannot give clear answers to all three, they belong in the public-data column only, regardless of what their marketing suggests.
Shadow AI, the tools your team has already adopted without formal approval, often sit in the confidential or restricted category without anyone having noticed. A brief, anonymous survey of current AI tool use, run before you build the map, typically surfaces three or four tools that need immediate attention. The survey also tells you which approved tools people are actually finding useful, which is the more productive half of the information.
Once the table exists, the governance burden drops considerably. The table replaces the policy document. Every team member asks two questions. Which tier is this data, and which column is this tool in. One page, four rows, four columns.
If you are building out your AI governance foundation, that is a good place to start a conversation. Book a conversation
