Key takeaways
- 28% of GPs already use AI tools in clinical practice[1], mostly for documentation and admin rather than clinical decisions[2].
- The strongest uses today are ambient note-taking, letter drafting and code suggestions with clinician review, plus answering routine practice questions.
- Anything that supports diagnosis, triage decisions or treatment is likely to be a regulated medical device and needs far more assurance[3][4].
- Start with one admin problem, measure time saved, and keep a clinician responsible for every output[3].
Where general practice is today
AI in general practice is no longer a future topic. In a 2025 Nuffield Trust and RCGP survey of 2,108 GPs, 28% said they currently use AI tools in clinical practice[1]. Among those users, 57% use AI for clinical documentation and note-taking and 44% for administrative tasks[2]. National policy is pushing in the same direction: the 10 Year Health Plan commits to rolling out ambient voice technology, digital triage and the single patient record to cut clinical admin within two years[5].
Source: Nuffield Trust, 2025[2]
Show the data
| Item | Value |
|---|---|
| Clinical documentation and note-taking | 57% |
| Administrative tasks | 44% |
The useful question for a practice manager is not "should we use AI?" but "which specific jobs can it do well today, under what controls, and what will it save?" The answer is narrower than the marketing suggests, and more useful.
Uses that work now, and their limits
| Area | What AI does today | Main limit or rule |
|---|---|---|
| Consultation notes | Ambient scribes transcribe and draft notes for the clinician to edit | Clinician must check every output; consent and DPIA needed |
| Letters and summaries | Drafts referral letters, patient summaries and replies | Not a medical device if solely drafting for review; sign-off required |
| Coding | Suggests codes from letters and notes | Suggestions only; a trained person confirms |
| Front door and admin | Answers routine practice questions, explains how to get help, routes admin requests | Must not give clinical advice or triage unless assured as such |
| Online request routing | Sorts requests by type for the right team | Clinical prioritisation may make it a medical device |
| Population searches | Helps build searches and spot patients due for review | Outputs checked against clinical criteria |
| Clinical decision support | Suggests diagnoses or treatment | Regulated medical device territory; rarely a first project |
Documentation: the biggest early win
Clinical documentation is where most GPs who use AI start, and where the time savings are easiest to see. The 10 Year Health Plan estimates that saving just 90 seconds per GP appointment would generate over 2,000 full-time-equivalent GPs' worth of capacity[6]. Ambient scribes and letter drafting tools aim at exactly that kind of saving.
MHRA guidance on ambient voice technology from July 2026 says products intended solely for transcription, summarising clinical conversations, drafting letters or suggesting clinical codes for a clinician to review are not regulated as medical devices[3]. It also says clinicians remain responsible for reviewing and verifying AI-generated transcripts, summaries and other outputs[3]. That is the core rule for documentation tools: they draft, people decide.
57%
of GPs using AI use it for clinical documentation and note-taking[2]
The front door and admin
A large share of practice workload is administrative: callers asking about opening hours, how to order a repeat prescription, whether test results are back, how to register. The 10 Year Health Plan says the government will use digital telephony to ensure all GP practice phones are answered quickly and will build AI-powered online advice into the NHS App to end the 8am scramble[7].
Within a practice, AI can take the routine part of that load. An AI Employee such as Ava answers practice information in the practice's approved words, explains how to use the online form and the NHS App, signposts 999 and NHS 111, and passes anything about a patient to reception without taking health details. That boundary matters: answering questions about the practice is admin, while deciding how urgent someone's symptoms are is triage.
Triage and decision support: proceed carefully
Tools that assess symptoms, prioritise clinical requests or suggest diagnoses are a different category. The MHRA says many software and AI products used in health and social care are regulated in the UK as medical devices[4], and its 2026 guidance says AI voice products that support diagnosis, treatment or prevention, or take automated action without clinician review, are regulated as medical devices[3]. These products may still be valuable, but they need evidence, regulatory compliance, a thorough DCB0160 safety case and close monitoring. They are rarely the right first AI project for a practice.
A simple test: if the tool's output could change what happens clinically to a patient without a clinician deciding, treat it as high risk.
What AI still does badly
- Knowing what it does not know. Drafts can sound confident and be wrong, including invented details.
- Local context. It may not know your services, rotas or pathways unless you give it approved content.
- Nuance in conversation, such as safeguarding cues, distress or a patient minimising symptoms.
- Rare and complex cases, where training data is thin.
- Staying in scope. Products change with updates, and new features can move a tool into regulated territory.
What patients and GPs think
The public is cautiously positive. 52% of UK adults expect AI to have a positive impact on healthcare[8], but healthcare is also the area where people most often say AI needs careful management to prevent negative outcomes[9]. GPs who do not use AI are more cautious still: 89% were concerned about professional liability and medico-legal issues, and 88% about a lack of regulatory oversight[10]. Good governance, clear information for patients and clinicians staying in control are the answers to both groups.
Data protection basics for any AI tool
Every AI tool that touches patient information brings the same set of questions, whatever it does. Answer them before the first patient is involved, and write the answers into a data protection impact assessment:
- What personal data does the tool receive, and is all of it necessary for the job?
- Where is it processed and stored, and by which companies, including subcontractors?
- Is any practice data used to train or improve the supplier's products, and can that be switched off?
- How long are recordings, transcripts and outputs kept, and who can delete them?
- What do patients need to be told, and how can they object?
- What happens to the data when the contract ends?
Free consumer AI tools usually fail several of these questions at once. That is the main reason staff should not paste patient details into them, however convenient it seems.
Training and culture
Tools succeed or fail on how people use them. Clinicians need to know how to check AI output efficiently and what errors to look for. Admin staff need to know where the tool's job ends and theirs begins. Everyone needs permission to report problems without feeling they are blocking progress.
- Include approved AI tools, and the rule against unapproved ones, in induction for all staff and locums.
- Share examples of real errors found in audits, anonymised, so people learn what to watch for.
- Name one person in the practice as the go-to for questions about AI tools.
- Review the practice's AI use at a partners' meeting at least twice a year.
How to start well
Find out what is in use
Ask staff, without blame, which AI tools they already use. Pause unapproved ones until assessed.
Pick one problem
Choose a measurable admin burden, such as letter backlog, note-writing time or routine phone questions.
Check the ICB offer
Many ICBs have assured products, funding or shared clinical safety work. Start there.
Assure before go-live
Confirm intended purpose and regulatory status, complete DCB0160 and a DPIA, and agree data processing terms.
Pilot and measure
Run with a few staff for four to six weeks, measure time saved and errors found, and ask patients for feedback.
Decide and govern
Scale only if the numbers and the safety review support it, and set a review date.
Example: one practice's first year
Example: a fictional practice of 9,000 patients started with letter drafting for two GPs, then added an ambient scribe pilot after its CSO completed a safety case. It also used an AI assistant to answer routine phone questions at the morning peak, with patient matters passed to reception. After a year, the partners kept the scribe and the phone assistant, dropped a coding tool that saved little time, and decided not to adopt a symptom-checker until its evidence and regulatory status were clearer.
Measuring whether it is worth it
- Time: minutes saved per consultation, per letter or per call, measured before and after.
- Quality: errors found in audits, corrections needed, complaints.
- Experience: clinician and staff feedback, patient feedback on calls and consultations.
- Safety: incidents and near misses logged against the hazard log.
- Scope: any supplier updates that change what the tool does.
Frequently asked questions
How many GPs use AI?
In a 2025 survey of 2,108 GPs, 28% said they currently use AI tools in clinical practice[1].
Is AI triage allowed in general practice?
Tools that support clinical decisions or prioritise by clinical need are likely to be regulated medical devices[4]. They need regulatory compliance, a safety case and close monitoring.
Who is responsible if an AI-drafted note is wrong?
The clinician who approves it. MHRA guidance says clinicians remain responsible for reviewing and verifying AI-generated outputs[3].
What is the best first AI project for a practice?
A measurable admin problem with low clinical risk, such as letter drafting, routine phone questions or note-taking with clinician review.
References
Nuffield Trust (2025). How are GPs using AI? Insights from the front line (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 (Medicines and Healthcare products Regulatory Agency) (2026). MHRA clarifies regulatory status of ambient voice technologies used in the NHS (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) ↑
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 (Department of Health and Social Care) (2025). Fit for the future: 10 Year Health Plan for England (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 (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 for Science, Innovation and Technology) (2024). Public attitudes to data and AI: Tracker survey (Wave 4) report (opens in a new tab) ↑
Nuffield Trust (2025). How are GPs using AI? Insights from the front line (opens in a new tab) ↑