Key takeaways
- Staff are already using AI tools, some obtained independently, so a policy is about safe use rather than permission.
- Keep a short approved list and a clear red line: no patient data in tools the practice has not approved.
- Every AI output that affects a patient is reviewed by a named person who stays accountable for it.
- New tools go through one approval route: DPIA, contract, clinical safety check where relevant, training.
- Review the policy at least yearly and whenever national guidance or your tools change.
Why a policy is needed now
AI tools have arrived in practices faster than governance. In a 2025 Nuffield Trust and RCGP survey of more than 2,000 GPs, 28% said they used AI tools in clinical practice; 13% of all GPs used tools provided by their practice and 11% used tools they had obtained independently[1]. Outside work, 40% of UK adults use AI chatbots for personal purposes at least monthly[2], so many staff will be familiar with them and tempted to use them for work.
Source: Nuffield Trust, 2025[1]
Show the data
| Item | Value |
|---|---|
| Provided by their practice | 13% |
| Obtained independently | 11% |
More is coming. The government's 10 Year Health Plan commits to rolling out ambient voice technology, digital triage and the single patient record within two years[3]. A policy written now gives the practice a way to say yes to useful tools safely, and a clear line on what is never acceptable.
What the policy should cover
Purpose and scope
Which tools count as AI for this policy (anything that generates text, audio or images, summarises, transcribes, predicts or classifies), and who it applies to: all staff, locums, students and contractors.
Approved tools
A short list of tools the practice has approved, what each is approved for, and who owns it.
Prohibited uses
Uses that are never allowed, whatever the tool.
Data rules
What data may go into which tools, and what never may.
Human review and accountability
Who checks AI outputs, and who is accountable for them.
Transparency
How patients are told when AI is used in their care or contact with the practice.
Approval route
How a new tool or use is proposed, assessed and approved.
Training
What staff must complete before using an approved tool.
Incidents
How to report an AI error or data concern, and what happens next.
Review
When the policy is reviewed and who signs it off.
Traffic lights for common uses
| Use | Status | Conditions |
|---|---|---|
| Drafting a staff rota email or a poster with no patient data | Green | Approved tool; check the output before use |
| Summarising a public guideline for a team meeting | Green | Check against the source; cite it |
| Ambient note-taking in consultations | Amber | Approved product only, after DPIA and clinical safety assessment; clinician reviews every note |
| Drafting a patient letter from clinical notes | Amber | Approved tool within the clinical system or assured platform; clinician signs |
| AI call handling for practice information and routing | Amber | Approved supplier, processor contract, agreed urgent-help wording, monitored |
| Pasting patient details into a public chatbot | Red | Never |
| Using AI to decide a diagnosis or treatment without clinician review | Red | Never |
| Recording consultations with a personal phone app | Red | Never |
Data rules in plain words
Information about health is special category data under UK GDPR Article 9[4], and serious infringements can attract fines of up to £17.5 million or 4% of worldwide turnover[5]. The policy should turn that into rules staff can remember.
- No patient-identifiable information in any tool that is not on the approved list. Removing the name is not enough; dates of birth, addresses, rare conditions and free text can still identify someone.
- Approved tools are used only for the purposes they were approved for.
- No staff personal data, such as appraisal notes or grievance details, in AI tools without approval.
- No practice log-ins, passwords or system screenshots in AI tools.
- Outputs containing patient data are stored in the record or the approved system, not in personal accounts.
Human review and accountability
AI can draft, summarise and suggest; a person decides. The MHRA says many software and AI products used in health and care are regulated as medical devices[6], but whatever the regulatory status, the professional who uses an output is accountable for it. The policy should say so plainly.
Sample clause: "Any AI output used in a patient's care or record, or sent to a patient, must be reviewed and approved by a named member of staff before use. That person is accountable for the content as if they had written it."
Clinicians' concerns are a good guide to where review matters most. Among GPs not using AI, 89% worried about professional liability and 82% about patient privacy and data security[1]. Spell out how the practice's indemnity and complaints processes apply to AI-assisted work.
More clauses you can adapt
- Approved list: "Only tools on the practice's approved AI list may be used for practice work. The list states what each tool is approved for, its owner and its review date."
- Personal accounts: "Staff must not use personal AI accounts or apps for practice work, including on personal devices."
- Recording: "Consultations, calls and meetings may only be recorded or transcribed by AI using an approved tool, with the patient or participants told in advance."
- Accuracy: "Users must check AI outputs against the source, including names, dates, doses and numbers, before use."
- Supplier changes: "The owner of each approved tool must tell the data protection lead before any change to how the tool processes data, including new features."
- Declining: "A patient may decline the use of an AI tool in their care without affecting their access to services."
Keep the wording short and concrete. Staff remember a few sharp rules better than a long document, and the traffic lights do most of the explaining.
Who does what
Name roles rather than people, so the policy survives staff changes. A typical small practice needs four: a policy owner, usually the practice manager, who keeps the approved list; a data protection lead or DPO who screens and reviews DPIAs; a clinical lead who signs off clinical safety for tools near clinical content; and a tool owner for each approved product, who monitors use, reports errors and leads its yearly review. In very small teams one person may hold two roles, but the clinical sign-off should always sit with a clinician.
The approval route for new tools
- Proposal: the member of staff describes the tool, the problem it solves and the data it would use.
- Screening: the practice manager and data protection lead decide whether a DPIA is needed; for patient data it almost always is.
- Assessment: contract and processor terms, security, data location, whether data trains the supplier's models, and the regulatory position.
- Clinical safety: for anything near clinical content, a clinical safety assessment and sign-off by a clinician.
- Decision: approve, approve with conditions, or decline, recorded with reasons.
- Go-live: training completed, tool added to the approved list with an owner and review date.
Make the route quick for low-risk uses. A policy that takes months to approve anything will drive staff back to unapproved tools.
Telling patients
Patients should know when AI is involved in their contact with the practice or their care, such as an automated assistant answering calls or a note-taking tool in a consultation. Healthcare is the area UK adults most often say needs careful management of AI, chosen by 29%[2]. The policy should require the privacy notice to list AI tools that process patient data, and require clinicians to tell patients before using ambient note-taking, with the option to decline.
Training and incidents
- A short session for all staff on the traffic lights and data rules, at induction and yearly.
- Tool-specific training before anyone uses an approved tool.
- A simple way to report an AI error, such as a wrong summary or a misheard call, through the normal incident process.
- A rule that significant AI errors are reviewed at the clinical governance meeting and shared as learning.
Example: adopting the policy
Example (fictional): a five-partner practice discovers during an appraisal that a salaried GP has been using a free transcription app on a personal phone.
In this example the practice does not discipline anyone. It drafts the policy within a month, publishes the traffic lights, and asks staff to declare any tools they use. Four tools are declared. One is approved with conditions, one is replaced with a product already assessed by a neighbouring practice, and two are stopped. The partners agree to assess an ambient note-taking product properly, with a DPIA and clinical safety sign-off, because the demand from clinicians is clearly there.
Review
Review the policy at least once a year, and sooner if national guidance changes, a new tool is approved, or an incident shows a gap. Keep the approved list as a separate, frequently updated page so the policy itself does not need rewriting every time a tool changes.
Frequently asked questions
Does a practice need an AI policy?
Yes. Staff are already using AI tools, some obtained independently, and a policy sets clear rules on approved tools, patient data, human review and accountability.
Can staff use public chatbots at work?
Only for tasks with no patient or confidential data, using tools the policy allows. Patient-identifiable information should never go into an unapproved tool.
Who is responsible for AI-generated content?
The named person who reviews and uses it. The policy should say that they are accountable for it as if they had written it.
How often should the policy be reviewed?
At least yearly, and sooner when national guidance, your tools or an incident shows something needs to change.
References
Nuffield Trust (2025). How are GPs using AI? Insights from the front line (opens in a new tab) ↑
GOV.UK (Department for Science, Innovation and Technology) (2024). Public attitudes to data and AI: Tracker survey (Wave 4) report (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) ↑
legislation.gov.uk (The National Archives) (2026). UK GDPR Article 9: Processing of special categories of personal data (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) ↑
GOV.UK (Medicines and Healthcare products Regulatory Agency) (2025). Software and artificial intelligence (AI) as a medical device (opens in a new tab) ↑