A Data Protection Impact Assessment (DPIA) is a structured way to identify and reduce privacy risks before an organisation deploys an AI system. Harper James is a UK full-service commercial law firm advising businesses on data protection and the laws of England and Wales.
Under the UK GDPR, a DPIA is required before processing that is likely to result in a high risk to individuals. AI systems often meet this threshold when they involve profiling, automated decision-making, large-scale monitoring, special category data, biometric data, children’s information, or the use of personal data to train or refine a model. The assessment should cover the full processing activity, not just the software product.
A useful starting point is the ICO’s guidance and risk assessment resources. The organisation should describe the system’s purpose, data sources, lawful basis, retention period, suppliers, users and expected outputs. It should also record whether the system makes decisions about individuals, whether people can challenge those decisions, and whether the data is accurate and proportionate for the stated purpose.
The risk analysis should examine risks such as inaccurate outputs, discrimination, excessive collection, unauthorised disclosure, re-identification, insecure prompts, unsuitable training data and the loss of meaningful human oversight. Controls should then be specific and testable. Examples include data minimisation, access restrictions, pseudonymisation, retention limits, testing for bias and accuracy, human review, appeal routes, supplier due diligence and contractual limits on the provider’s use of submitted data.
The DPIA should also allocate responsibility. A data protection officer, where appointed, should provide advice and monitor the process. Procurement, information security, product and legal teams each need clear actions. If high residual risks remain after mitigation, the organisation must consult the Information Commissioner’s Office before starting the processing. The DPIA should be approved before launch and revisited when the model, purpose, data sources, supplier or decision-making process changes.
Generative AI has made DPIAs more operational. Organisations now need to distinguish between public tools, enterprise services and internally hosted models, while checking where prompts and outputs are stored, whether data is used for further training, and how errors are reported. A DPIA is not a one-off document for procurement. It is a record of the decisions, safeguards and monitoring arrangements that support responsible use of personal data throughout the system’s life.