Key takeaways
- Online request volumes jumped after practices had to keep online consultation tools open in core hours, which is why AI-assisted sorting and triage are being pitched to practices[1].
- Software that prioritises or directs clinical care may be a medical device, and many AI products in health and care are regulated as such[2].
- Before go-live you need a named clinical safety lead, a hazard log, a DPIA and a clear rule on who makes the final clinical decision.
- GPs who do not use AI worry most about liability (89%) and lack of regulatory oversight (88%); answer those concerns in your governance pack[3].
- Run any pilot in parallel with your current process and audit disagreements before switching.
Why AI triage is on the agenda
Since 1 October 2025, practices in England have had to keep online consultation systems open throughout core hours, 8am to 6.30pm, for routine requests[4]. Demand followed. In March 2026 practices received 9,617,628 online requests, 69% more than in August 2025 before the requirement[1]. Every request still needs someone to read it, decide how urgent it is and route it to the right person, and that reading time is now one of the largest hidden workloads in general practice.
The government's 10 Year Health Plan commits to rolling out digital triage alongside ambient voice technology and the single patient record, so suppliers are moving quickly[5]. The governance questions are the same whatever the product: what exactly does it decide, how do you know it is safe, and who is accountable when it is wrong?
What 'AI triage' can mean
Products sold under the same label do very different jobs. Before any procurement conversation, decide which of these you are actually buying, because the regulatory and safety burden rises sharply as you move down the list.
| Function | What it does | Clinical risk | Typical governance |
|---|---|---|---|
| Admin sorting | Separates admin requests such as fit notes or letters from clinical ones | Low, if clinical requests are never auto-closed | DPIA, audit of misfiled requests |
| Summarising | Condenses free text for the reviewing clinician | Moderate: omissions can mislead | Clinical safety case, sampling against originals |
| Prioritising | Flags requests as urgent or routine | High: missed urgency is the main hazard | Medical device status check, hazard log, parallel running |
| Directing care | Suggests or books the clinician, service or timeframe | High | As above, plus clear sign-off rules |
Most practices start with admin sorting and summarising, where a human still reads every clinical request. That is a sensible first step: it saves time without handing urgency decisions to software.
The regulatory duties you cannot skip
The MHRA says many software and AI products used in health and social care are regulated in the UK as medical devices[2]. If a tool prioritises or directs clinical care, ask the supplier for its device classification and registration in writing, and do not accept a general statement that it is only decision support.
- Clinical safety: suppliers work to DCB0129 and deploying organisations to DCB0160. Your practice needs a clinical safety officer, a hazard log for your own use of the tool and a safety case before go-live.
- Data protection: health data is special category data under UK GDPR Article 9, so you need a lawful basis and an Article 9 condition[6]. Complete a DPIA before processing starts.
- Confidentiality: apply the Caldicott Principles, especially using the minimum necessary data and being transparent with patients[7].
- Contract: confirm where data is processed, whether it is used to train models, retention periods and how you can audit outputs.
No contract clause moves clinical responsibility for a patient away from the practice. Decide in writing which decisions the tool can make, which it can suggest and which only a clinician can take.
How GPs are using AI and what worries them
A 2025 Nuffield Trust and RCGP survey of 2,108 GPs found 28% currently use AI tools in clinical practice[8]. Only 13% of all GPs use tools provided by their practice, while 11% use tools they obtained themselves[9]. That second group matters for governance: tools adopted informally sit outside your DPIA, your safety case and your contracts. Among GPs who use AI, 57% use it for documentation and 44% for administrative tasks[10].
Among GPs not using AI, 89% were concerned about professional liability and medico-legal issues and 88% about lack of regulatory oversight[3], while 82% worried about patient privacy and data security[11]. These are the concerns your partners and salaried GPs will raise, so build the governance pack around them rather than around supplier marketing.
Questions to ask a supplier before you sign
- Exactly which decisions does the product make, and which does it only suggest?
- Is it a registered medical device in the UK, and for what intended purpose?
- Can we see your DCB0129 clinical safety case and hazard log, and the name of your clinical safety officer?
- What evidence do you have on missed urgent cases, and how was it collected?
- How are requests handled when the system is unavailable or uncertain?
- Where is data stored and processed, and is any of it used to train or improve models?
- What audit trail can we export, and how quickly?
- Which practices of similar size and patient mix can we speak to?
If a supplier cannot answer the first three questions in writing, stop there. Everything else depends on them.
A governance plan from decision to go-live
Define the scope
Write one paragraph on what the tool will do, what it will not do and which requests are out of scope, such as anything from a child or a patient flagged as vulnerable.
Appoint a clinical safety lead
A trained clinician who owns the hazard log and the safety case for your deployment, with time allocated to the role.
Complete the DPIA
Map the data flow, the lawful basis, retention and sharing, and consult your data protection officer.
Run in parallel
For at least four weeks, let the tool classify requests while clinicians triage as normal, then compare. Investigate every case where the tool rated a request less urgent than the clinician did.
Agree go-live criteria
Set thresholds in advance, such as zero missed red-flag cases in the parallel run, and a sign-off by the clinical safety lead and partners.
Tell patients
Update your privacy notice and website so patients know how requests are processed and who reads them.
Audit monthly
Sample requests, review incidents and keep the hazard log live. Report results at partners' meetings.
Example: a fictional practice of 12,000 patients pilots summarising and admin sorting only. Over four weeks of parallel running, clinicians disagree with the tool's admin and clinical split on 3% of requests, all of them misfiled admin items rather than missed clinical ones. The practice goes live for sorting, keeps clinician review of every clinical request, and schedules a second review before considering prioritisation.
Measuring whether it actually helps
A tool that saves reading time but slows the response to patients is not an improvement. Under Modern General Practice Access, patients should know on the day how their request will be handled[12]. Measure against that standard before and after any change, using the same report and the same weeks of the year where possible.
- Time from submission to first human review, by hour of the day.
- Share of requests given an outcome on the day they arrive.
- Clinician minutes spent reading and routing requests per session, from a two-week time study.
- Disagreement rate between the tool and the reviewing clinician, split by direction: rated too urgent or not urgent enough.
- Incidents and near misses logged against the hazard log.
- Patient complaints and compliments that mention online requests.
Agree in advance what would make you stop. A single missed red-flag case should trigger a review of the hazard log and a pause on any automated prioritisation until the cause is understood. Reporting these measures at every partners' meeting keeps the tool under clinical control and gives you the evidence to expand it, or to switch it off, with confidence.
Keep a short record of each decision, who took it and the evidence behind it. If a patient or regulator later asks why the practice used the tool in a particular way, that record is your answer.
Patients and the wider team
Public views are mixed. Healthcare is the area UK adults most often say needs careful management of AI to prevent negative outcomes, chosen by 29%[13]. Be open about what the tool does, keep a human route for anyone who wants one, and make sure reception can explain the process in two sentences.
Phone tools are a separate decision from triage. An AI Employee such as Ava answers practice information questions, signposts urgent help and passes patient matters to reception; it does not triage. Keep that boundary clear in your policies, and read common AI errors in healthcare admin and the clinical safety officer role before you expand any tool's scope.
Frequently asked questions
Is AI triage software a medical device?
It can be. The MHRA says many software and AI products used in health and care are regulated as medical devices, so ask the supplier for the device classification and intended purpose[2].
Do we need a DPIA for an AI triage tool?
Yes. It processes health data, which is special category data under UK GDPR, and is new technology with a high risk to individuals, so complete a DPIA before processing starts[6].
Can an AI tool close a clinical online request without a clinician?
Avoid it. Keep clinician review of every clinical request unless your safety case, device classification and parallel-run evidence clearly support a narrower scope.
How many GPs use AI tools?
A 2025 Nuffield Trust and RCGP survey found 28% of GPs use AI tools in clinical practice, mostly for documentation[8].
References
Care Quality Commission (2026). State of Care 2025/26: capacity in primary care (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) ↑
Nuffield Trust (2025). How are GPs using AI? Insights from the front line (opens in a new tab) ↑
British Medical Association (2025). Pressures in general practice data analysis (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) ↑
GOV.UK (National Data Guardian) (2026). The Caldicott Principles (opens in a new tab) ↑
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) ↑
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) ↑
NHS England and DHSC (2023). Delivery plan for recovering access to primary care (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) ↑