What Data Minimisation Requires Under GDPR
GDPR Article 5(1)(c) states that personal data must be adequate, relevant, and limited to what is necessary in relation to the purposes for which they are processed. These three elements are distinct. Adequate means you must collect enough data to fulfil the stated purpose. Relevant means the data must have a genuine connection to that purpose. Limited to what is necessary means you should not collect data speculatively, even if it might be useful later. The ICO has been clear in its enforcement decisions that broad or precautionary data collection is not compatible with data minimisation. The principle operates alongside the purpose limitation principle, which requires that data collected for one purpose is not repurposed for something else without a fresh lawful basis. For AI products, the combination of data minimisation and purpose limitation creates a framework that challenges the common practice of logging everything for potential future model improvement. The question is not whether you might find the data useful someday - the question is whether collecting it now is necessary for the specific, documented purpose.
Data Minimisation in LLM Prompt Design
LLM prompts are a primary data minimisation risk area for AI products. Users frequently include personal information in prompts, sometimes inadvertently. Customer names, employee details, addresses, financial figures, health information, and third-party personal data all appear regularly in user-generated prompt content. The data minimisation principle applies to how your system handles that data once it arrives. Prompt pre-processing that detects and redacts common PII patterns, such as email address formats, national insurance number patterns, and UK phone number formats, before the prompt is logged or stored reduces your data retention surface without degrading the AI's ability to respond to the query. Where prompts must be logged, a short retention window is more defensible than indefinite storage. Your privacy notice must accurately describe how prompt content is handled, how long it is retained, and whether it is used to improve models. Vague descriptions of data handling in AI products are an ICO enforcement risk.
Training Data and Minimisation
If you are fine-tuning models or using proprietary data for retrieval-augmented generation, the minimisation principle applies to what you include in training datasets and knowledge bases. Before including any dataset in fine-tuning, assess whether personal data it contains is necessary for the model to learn the intended behaviour. In many cases, the answer is no: a customer service model trained on historical support tickets does not need to retain customer names and email addresses in the training examples to learn appropriate response patterns. Anonymising or pseudonymising personal data from training datasets before fine-tuning satisfies minimisation without sacrificing model quality, in most cases. Where you do need personal data in training or knowledge bases, document why, establish retention periods for the training data itself, and ensure that individuals whose data appears in training datasets are covered by your privacy notice and have an appropriate lawful basis for that processing.
Analytics and Behavioural Data Minimisation
Analytics is another area where data minimisation is frequently ignored in favour of collecting comprehensive behavioural data. UK GDPR and the Privacy and Electronic Communications Regulations require cookies and behavioural tracking to have a lawful basis, usually consent, and that the tracking is proportionate to the analytics purpose. For AI products, analytics data can be particularly rich: you can capture what users asked the AI, how they iterated on prompts, what outputs they accepted or rejected, and how long they spent reviewing responses. All of this is personally identifiable when combined with user identifiers. The minimisation principle requires you to ask what analytics data is actually needed to make product decisions, and collect only that. Aggregated, anonymised analytics at the cohort level often serves product decision-making as well as individual-level tracking and avoids the compliance overhead entirely. Where individual-level tracking is genuinely necessary for a specific product function, pseudonymisation with time-limited identifier retention is a minimisation-compatible approach.
Applying Minimisation to Third-Party AI APIs
When your product passes user data to a third-party LLM API, minimisation applies to what you include in that API call. If you are using a RAG system to provide context to the LLM, only the retrieved chunks genuinely relevant to the user's query should be included in the prompt context, not entire documents that might contain personal data about third parties. If your system enriches user prompts with profile data from your database before sending to an LLM, assess whether that enrichment is necessary for the AI function or whether it is a convenience that happens to pass personal data to a third-party processor unnecessarily. Every piece of personal data in an API call to a third-party LLM is data you are transferring to a processor under GDPR, requiring a valid Data Processing Agreement, a transfer mechanism if the API routes data outside the UK or EEA, and a documented justification that the transfer is necessary.
Practical Steps to Implement Data Minimisation
Building data minimisation into AI products requires treating it as a design constraint from the start of scoping rather than a compliance task applied after build. During requirements definition, for each data point you plan to collect, ask: what specific product function does this enable, is that function in scope for this sprint, and what is the minimum collection period needed? Apply these questions to analytics events, prompt logs, user profile fields, document uploads, and RAG knowledge base content. Implement automated deletion schedules for data with defined retention periods rather than relying on manual data hygiene. Review your data map at the start of each sprint to check whether new features introduce new personal data collections that need minimisation treatment. At SpeedMVPs, data minimisation is a scoping-phase conversation, not a post-launch compliance exercise. The products we deliver have defined data retention periods and PII exposure surfaces that are documented before the first line of code is written.