Smarter referrals: why AI-assisted triage should begin in primary care

Every day, general practitioners (GPs) are asked to decide: does this patient need acute care? To which specialty and with what information? These questions, often made under time pressure and with limited access to diagnostics or specialist input, shape patient care trajectories. Chronic care decisions with uncertainty such as strange semi-vertiginous episodes falling between neurology and otolaryngology also lack structured support. If artificial intelligence (AI) is to fulfil its promise of transforming triage and acute care, its most impactful point of entry is not the emergency department or outpatient clinic; it is primary care. We argue that AI-based decision support should be piloted in general practice, where the gains in efficiency, effectiveness and positive patient impact are highest.

Rethinking triage: start where decisions begin

AI conversations on triage largely focus on hospital settings. Studies show models like GPT-4 (a Large Language Model from OpenAI) outperform junior doctors in triage accuracy and can efficiently summarise clinical presentations.^1–3 However, these capabilities are often applied after referral decisions, too late to influence the most critical point of uncertainty.

That uncertainty lies in the community: the urgent care centre, the after-hours general practice clinic, the rural general practice. An overstretched medical system has also increased the number of acutely ill people presenting to general practice rather than urgent care clinics. This is where decisions are made under constrained time and limited resources. Integrating AI in this space would reduce emergency department work and enhance the quality of early care decisions.

The cost of getting it wrong

Misjudging triage carries a high cost both financially and in terms of human health. Over-referral results in unnecessary ambulance transfers, hospital stays and consults that strain public health budgets. Under-referral poses greater risk: missed diagnoses, avoidable deterioration and deaths. Patients also suffer from misdirected referrals. Waiting in an emergency department for review that results in discharge is frustrating and anxiety-inducing. For rural or marginalised communities, hospital outpatient visits and admissions can mean family separation, lost income and disrupted support systems. Community pathways, a supposed solution to decision-making at the interface of primary and secondary care, have become overly complex and vary from region to region. Although AI has shown promise in improving hospital triage accuracy, the missed opportunity lies in its limited use upstream – in primary care, where these decisions originate.^4–6

What AI could do in general practice

Imagine a clinical decision support tool embedded within the primary care workflow, retrieving past medical history, medications, allergies and previous correspondence while clinicians enter symptoms and key findings. Trained on large datasets, AI returns:

This is not speculative. AI models perform well across clinical vignettes, show acceptable triage calibration, and have proven useful in managing inbox tasks, summarising notes and histories.^3,7–9 AI tools customised for primary care are already in use and showing promise.¹⁰

These tools do not replace clinical reasoning, they augment it. A calibrated risk score or differential diagnosis support clinicians in navigating ambiguity, validating a referral, or choosing safe watchful waiting.

Why primary care is the right place

Primary care is uniquely placed to benefit from AI. As gatekeepers and care coordinators, general practice manages uncertainty without expensive infrastructure, making the return on investment higher at the start of a healthcare episode than later. Yet, primary care remains underrepresented in AI development. Although hospital-based systems receive investment and pilot programs, primary care adapts tools built for different contexts.¹¹ Fragmented funding via partial government subsidy and patient co-payment dilutes and obscures responsibility for expenditure on technology despite the clear and acknowledged benefits of such investment.^12–14 A well-equipped telemedicine cart for primary care needs costs between NZD 14,400 and 24,000. This contributes to scepticism and failed implementation.^12,13

What’s needed is AI development, co-designed with primary care clinicians and focused on real-world decision points like triage. Training frameworks already exist and are being adopted internationally.¹⁴ With the right investment and policy direction, primary care could be where safe, equitable and clinically effective AI integration begins.

Safeguarding equity and trust

Key stakeholders in AI integration within primary health care in New Zealand include the Medical Council of New Zealand as the regulator of registered medical practitioners, the government (the Ministry of Health as the governance arm and Te Whatu Ora as the operational arm) and the Royal New Zealand College of General Practitioners as the body to set standards for general practice and provision of education. Each need to develop awareness of, and respond to rapid changes in the AI environment.

However, introducing AI into the health system in Aotearoa New Zealand is not without risk. Poorly trained models can reinforce biases, especially against Māori, Pacific and rural populations. A recent scoping review emphasised that AI’s effect on inequity is dependent on implementation design.¹⁴

AI systems must be built with diverse data and voices. Transparency, explainability and robust clinician training, both technical and ethical, are essential.^15–17 Community involvement in design, testing and rollout is essential.

A smarter future for triage

Importantly, GPs are not resistant to AI. Surveys show strong interest among clinicians – particularly when systems are safe, trusted and transparent.¹⁸ They seek tools that support rather than replace human judgement and relationships. Accurate information on limitations and uses, regulatory clarity, the development of local tools that reflect local clinical practise and reassurance on privacy concerns all require further development.^19,20

We propose that AI-based triage should be piloted and scaled in general practise focused on:

GPs make these decisions daily. AI can make those decisions safer, more equitable and more consistent, improving patient outcomes and reducing strain on secondary care.

Delaying AI integration at the front-line perpetuates inefficiency and inequity in the health-care system. To relieve downstream pressure, we must start upstream. The future of smarter triage begins at the primary interface of care: general practice.