Primary care AI must prove it improves the next decision

New trials are testing AI at the first point of care. Their results show why answering a medical question, directing a patient and improving treatment are different achievements.

The first conversation in healthcare often determines what happens next. A patient describes a problem, somebody decides how urgently it needs attention, and information moves towards the professional expected to act. Artificial intelligence is entering each part of that process, from public chatbots to systems that advise clinicians or prepare referral information.

The evidence now emerging is more useful than a simple contest between doctors and machines. Trials are beginning to distinguish what a system knows from what people can do with it. For European primary care, the central question is whether the technology helps the right patient reach the right assessment, without creating an additional obstacle along the way.

A knowledgeable model can still be difficult to use

A UK study published in Nature Medicine in February 2026 tested this distinction with 1,298 participants. People worked through medical scenarios with assistance from one of three language models or through sources they would ordinarily use at home. Those using the models did not perform better than controls in identifying relevant conditions or choosing an appropriate course of action.

The models performed much better when researchers supplied the complete scenarios directly. The study identified problems in the interaction: participants sometimes omitted important information, models misinterpreted questions, and users did not consistently act on useful suggestions. These were controlled scenarios, not a measurement of outcomes among people seeking care for their own illnesses.

That finding changes the development task. A patient-facing service needs to help people express uncertainty, describe symptoms in their own words and understand what to do next. Improving a model’s examination score does not necessarily improve those abilities. Testing the whole conversation with intended users is therefore a different exercise from testing its medical knowledge.

Triage is a decision about urgency and destination

The distinction predates generative AI. A 2015 BMJ audit examined 23 symptom checkers using standardised patient descriptions. Correct diagnoses appeared first in 34% of evaluations, while appropriate triage advice was given in 57%. The historical results should not be treated as a score for today’s products, but the study separated two tasks that remain easy to confuse.

A tool could suggest an incorrect diagnosis while recommending an appropriate clinical assessment. Conversely, naming a plausible condition is of limited value if the accompanying advice sends a patient to an unsuitable service. Evaluation needs to follow the intended decision rather than combine every output into a single accuracy figure.

For a hypothetical primary-care pilot, researchers could examine the consequences of both excessive escalation and insufficient escalation. The first may add pressure to urgent services; the second may delay necessary assessment. Those errors have different implications, and an average score can conceal a particularly important weakness.

The receiving service belongs in the evaluation too. A recommendation to seek an appointment only helps if the route exists, accepts the patient and receives enough information. Otherwise, an apparently successful digital encounter can end with a patient repeating the same account elsewhere. Recording that repetition would help distinguish genuine coordination from a digital process that merely transfers the same work to a different team. It would also identify where information is lost between services.

Simulated clinical dialogue establishes a narrower result

Research on AMIE illustrates the capabilities being developed. A 2026 Nature study compared the conversational AI system with 21 primary-care physicians across 100 simulated cases involving multiple visits. The scenarios reflected UK NICE and BMJ Best Practice guidance, allowing assessment of disease-management conversations over time. The authors reported strong performance but explicitly identified the need for further research before real-world translation.

Simulation provides a controlled setting for examining how a system asks questions, revises a plan and responds to additional information. It can expose weaknesses before patients encounter them. It cannot reproduce every interruption, incomplete record or competing priority in routine practice, and it does not demonstrate improved health outcomes.

For commissioners, the useful next question is which parts of the simulated task will actually occur in the proposed service. A system evaluated through text may enter a setting involving interpreters, telephone calls or examination findings that never reach its input. The intended deployment should determine what additional evidence is needed.

A pragmatic trial tests whether patients benefit

A cluster-randomised trial published in June 2026 moved closer to that question. Across 16 Kenyan primary-care facilities, 103 clinical officers used an electronic record with or without language-model assistance. The primary analysis included 9,347 patient encounters and examined expert-adjudicated treatment failure within 14 days.

Treatment failure occurred in 2.2% of intervention encounters and 2.0% of controls. After adjustment, the odds ratio was 0.77, with a 95% confidence interval of 0.55–1.08; the difference was not statistically significant. No serious adverse events were judged related to the intervention. This supports a bounded conclusion: the trial did not establish a reduction in its primary outcome, and it did not identify a serious safety signal under its study conditions.

European services should neither dismiss the findings because the setting differs nor assume that the result transfers unchanged. Clinical roles, common conditions, access to tests and continuity of care all affect how decision support is used. The study offers a model for asking a clinically meaningful question, rather than a ready-made forecast for a European practice.

Shorter consultations can move work elsewhere

A Chinese trial of the PreA chatbot examined the transition to specialist care. Its final analysis included 2,069 participants across two centres. People either used the tool independently, used it with staff support or proceeded without it before seeing a specialist. In the independent-use group, mean physician consultation time fell from 4.41 to 3.14 minutes, a reduction of 28.7%.

The patients spent approximately three and a half minutes interacting with PreA, and specialists also reviewed its reports. The reduction therefore describes the consultation itself, rather than the removal of all preparatory work. The trial also examined communication and coordination; it was not evidence that shorter appointments necessarily improve diagnosis or longer-term outcomes.

For a European referral service, a useful evaluation would count the complete journey. It could record preparation time, repeated questions, information corrected by clinicians and any additional tests or appointments. A task may become easier for a specialist while requiring more effort from patients, reception staff or referring practices.

That redistribution is not automatically undesirable. Preparing information in advance may be worthwhile if it improves the encounter. The important point is to establish who performs the work and whether the resulting information is reliable enough to support the next decision.

Keep access and follow-up inside the service design

NHS England’s guidance on online consultations makes clear that digital access sits within a practice’s operating model. Its January 2026 update describes the minimum online offer for non-urgent appointment requests, medication queries and administrative requests, while recognising that some practices choose to accept urgent requests through their systems. An online form is not, by itself, a clinical triage service or an AI system.

An AI pilot should make the boundary equally clear. Patients need to know whether they are submitting information for later review, receiving an immediate recommendation or booking an assessment. The service needs a way to identify failed submissions, unanswered requests and conversations abandoned before the necessary information is collected.

Participation figures alone would miss those experiences. Follow-up could examine who completes the process, who switches to another route and who cannot use it. Language, disability, confidence and access to a private place to communicate may affect the answer. Alternative contact routes should be evaluated as part of the same service.

The strongest primary-care AI evidence will connect three stages: the information a person supplies, the action the service takes and the outcome that follows. Better dialogue and faster administration can contribute, but neither settles the clinical question. A useful system makes the next decision more dependable and leaves a clear route for correcting it when the patient’s situation changes.