Key takeaways
- A DPIA is required before processing that is likely to be high risk, which includes large-scale processing of health data; most new clinical or patient-facing systems qualify.
- UK GDPR sets a minimum content: a description, a necessity and proportionality check, a risk assessment and the measures to address the risks.
- You must seek your data protection officer's advice, and suppliers' templates are a starting point, not a finished assessment.
- AI tools add questions about transcripts, training data, accuracy and whether the product is a regulated medical device.
When you need a DPIA
UK GDPR requires a data protection impact assessment before any processing that is likely to result in a high risk to people's rights and freedoms[1]. It singles out large-scale processing of special category data, such as health information, as a case where an assessment is required[1]. In a practice, that covers most new systems that touch clinical records or patient conversations.
- A new or migrated clinical system, or moving records to a cloud service.
- Online consultation, triage or messaging tools.
- Cloud telephony with call recording, transcription or automated answering.
- Ambient voice technology that listens to consultations and drafts notes.
- A new data-sharing arrangement across a primary care network or with a private partner.
- CCTV, door entry or staff monitoring systems.
- Patient-facing apps, portals or chat tools.
Do the DPIA before the contract is signed and before any patient data flows. An assessment written after go-live describes decisions already made rather than shaping them.
Why this matters now
AI is arriving in general practice quickly, often from the bottom up. In a 2025 Nuffield Trust and RCGP survey, 28% of GPs said they used AI tools in clinical practice[2]. Of those, 57% used them for clinical documentation and note-taking and 44% for administrative tasks[2]. Strikingly, 11% of all GPs used tools they had obtained independently, almost as many as the 13% using tools provided by their practice[2].
Source: Nuffield Trust, 2025[2]
Show the data
| Item | Value |
|---|---|
| Provided by their practice | 13% |
| Obtained independently | 11% |
| A combination of both | 4% |
National policy will accelerate this. The 10 Year Health Plan commits to rolling out ambient voice technology, digital triage and the single patient record to cut clinical admin[3]. Each of those needs a DPIA at the practice level, even when a national or regional body has done its own assessment.
Screening: do we need a full assessment?
A short screening form helps decide quickly. If you answer yes to any of these, complete a full DPIA:
- Will the system process health information about many patients?
- Does it use new technology, such as AI, voice recognition or automated decisions?
- Will data go to a new supplier or leave the UK?
- Does it combine data from different sources or match records?
- Will patients or staff be monitored, recorded or profiled?
- Could a failure, breach or error cause patients significant harm or distress?
If every answer is no, record the screening and the reasons. That note is your evidence that you considered the question.
The walkthrough
1. Describe the processing
What data, about whom, collected how, stored where, accessed by whom, kept for how long and for what purposes. Include a simple data flow diagram from patient to supplier and back.
2. Consult
UK GDPR requires you to seek the advice of your data protection officer[1]. Also ask the staff who will use the system and, where appropriate, your patient group.
3. Test necessity and proportionality
Is there a less intrusive way to achieve the same aim? Is every data item needed? Is the retention period justified?
4. Identify the risks
List what could go wrong for patients and staff: unauthorised access, inaccuracy, loss, unexpected secondary use, lack of transparency, excessive retention.
5. Assess likelihood and severity
Rate each risk before mitigation, using a simple scale such as low, medium and high.
6. Agree measures
For each risk, set out the safeguards: contractual, technical, organisational. UK GDPR expects the assessment to include the measures to address risks and demonstrate compliance[1].
7. Decide and sign off
Record the residual risk and who accepted it, normally a partner or the senior information risk owner. If high risk remains that you cannot reduce, consult the regulator before going ahead.
8. Build it into the project
Turn the measures into tasks with owners and dates, and set review triggers.
A sample risk register
| Risk | Before mitigation | Measures | After mitigation |
|---|---|---|---|
| Transcripts or recordings kept indefinitely | High | Retention set in the system; automatic deletion; policy | Low |
| Supplier uses patient data to train its models | High | Contract clause forbidding it; written confirmation | Low |
| Draft note contains errors that enter the record | Medium | Clinician reviews every draft; audit sample monthly | Low |
| Patients unaware the tool is in use | Medium | Notice in waiting room and privacy notice; verbal explanation | Low |
| Staff use unapproved free tools instead | Medium | Approved tool provided; acceptable use policy; training | Low |
| Data stored outside the UK without safeguards | Medium | Confirm hosting; adequacy or transfer safeguards documented | Low |
Keep the register honest. Reviewers and inspectors are more reassured by a risk rated high with clear measures than by a list where everything is low from the start.
Questions specific to AI tools
- Does the tool record audio, and if so, is the audio deleted after transcription?
- Are transcripts, summaries and prompts stored, where and for how long?
- Is any patient data used to train or improve the supplier's models?
- How accurate is it, how was that tested and what happens when it is wrong?
- Can patients decline its use in their consultation, and how is that recorded?
- Does the tool make or suggest clinical decisions?
The last question matters for regulation as well as data protection. The MHRA says many software and AI products used in health and social care are regulated as medical devices[4]. Ask the supplier for its regulatory status in writing and record it in the DPIA. NHS deploying organisations are also expected to complete clinical safety work, so align the DPIA with your clinical safety case rather than running two separate exercises.
What to ask suppliers for
- Their own DPIA or a DPIA template for customers, plus the data flow diagram.
- The data processing agreement and any sub-processor list.
- Hosting location and transfer safeguards.
- Security certifications and their latest Data Security and Protection Toolkit status, where relevant.
- Clinical safety documentation and any medical device registration.
- Retention settings and how deletion is verified.
A supplier's DPIA describes their system. Yours describes how your practice will use it, with your patients, staff and workflows. Use theirs as evidence, not as a substitute.
What happens if you skip it
Article 35 is one of the controller obligations in Articles 25 to 39, which carry fines of up to £8.7 million or 2% of worldwide turnover[5]. More practically, a missing DPIA is one of the first things asked for after a breach or complaint involving a new system, and its absence makes every other decision harder to defend.
Keeping the assessment alive
A DPIA is a living document. Systems change after go-live: suppliers add features, settings drift and staff find new uses. Set review triggers when you sign the assessment off, and put the first review in the diary.
- The supplier announces a new feature, especially anything involving AI, analytics or new data sharing.
- The system is extended to new staff groups, sites or patient groups.
- A breach, near miss or complaint involves the system.
- The supplier changes hosting location or sub-processors.
- The law or national guidance changes in a way that affects the processing.
- A year has passed since the last review.
Each review need only answer two questions: has anything changed, and are the measures working? A short dated note is enough when nothing has changed; a fuller update is needed when something has.
Example: an AI scribe pilot
Example (fictional): a GP partnership agreed to pilot an ambient voice tool with three clinicians. The screening answered yes to new technology and health data at scale, so the manager ran a full DPIA with the DPO. The assessment found that the default setting kept audio for 30 days and that the contract was silent on model training. The practice changed the setting to delete audio once the transcript was produced and obtained a written contractual commitment that patient data would not be used for training.
Patients were told about the pilot by a notice in each consulting room and could decline. After three months the DPIA was reviewed with real data on accuracy and opt-outs before the tool was extended to other clinicians.
DPIAs connect closely with breach handling and confidentiality. See our guides to reporting a data breach, the most common breaches in practices and the Caldicott Guardian role.
Frequently asked questions
When is a DPIA mandatory for a practice?
Before any processing likely to result in high risk, including large-scale processing of health data[1]. Most new clinical, telephony or AI systems meet that test.
What must a DPIA contain?
At least a description of the processing and purposes, an assessment of necessity and proportionality, an assessment of risks and the measures to address them[1].
Can we use the supplier's DPIA?
Use it as evidence, but complete your own assessment covering how your practice will use the system with your patients and staff.
Do AI scribes need a DPIA?
Almost always. They process health information at scale using new technology, and they raise questions about audio retention, model training and accuracy.
References
legislation.gov.uk (The National Archives) (2026). UK GDPR Article 35: Data protection impact assessment (opens in a new tab) ↑
Nuffield Trust (2025). How are GPs using AI? Insights from the front line (opens in a new tab) ↑
GOV.UK (Department of Health and Social Care) (2025). Fit for the future: 10 Year Health Plan for England (opens in a new tab) ↑
GOV.UK (Medicines and Healthcare products Regulatory Agency) (2025). Software and artificial intelligence (AI) as a medical device (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2026). UK GDPR Article 83: General conditions for imposing administrative fines (opens in a new tab) ↑