I've been in health IT long enough to have lived through several waves of "this will transform everything." Electronic health records. Telehealth. Patient portals. Each one delivered value eventually, but not in the way the early hype suggested. AI in primary health is following the same pattern, and I think it's worth being honest about where we actually are.
The Honest Picture
- AI in primary health has real, demonstrated value in specific areas: clinical documentation, triage support, and administrative automation. These aren't glamorous, but they address genuine pain points.
- The harder applications, clinical decision support, predictive analytics, population health management, are limited by data quality problems that existed long before AI arrived.
- New Zealand's primary care sector faces specific challenges that global AI solutions don't address: small practice sizes, diverse populations, and health system structures that differ from the US and UK models most AI tools are built for.
- The biggest risk isn't that AI won't work. It's that organisations will invest in the wrong applications while ignoring the data foundations they need.
What's Actually Working
I'll start with the good news. There are areas where AI is delivering measurable value in primary care right now.
Clinical documentation. This is the most immediately useful application I've seen. GPs spend a disproportionate amount of time on documentation. AI tools that can draft consultation notes from recorded conversations, suggest coding, and pre-populate referral letters are saving real time. Not transformative time, but meaningful time. Thirty minutes a day back for a GP who's already stretched is significant.
11hrs
per week spent by NZ GPs on administrative tasks including documentation
Source: RNZCGP Workforce Survey, 2023
Triage and prioritisation. AI-assisted triage tools that help reception staff and nurses prioritise patient contacts are working in practices that have the right workflows around them. The key phrase is "AI-assisted." These tools support clinical judgement. They don't replace it. The ones that try to replace it fail, because primary care triage requires context that no model has.
Administrative automation. Appointment reminders, prescription renewals, recall management, reporting. These are high-volume, low-complexity tasks where automation delivers consistent returns. Not exciting. But reliable.
What's Still Hype
Diagnostic AI. The demos are impressive. The reality in a NZ GP practice is different. Diagnostic AI works well with clean, structured data in controlled settings. A GP consultation generates messy, contextual, relationship-based information that current models struggle with. And the liability question remains unanswered: if an AI-assisted diagnosis is wrong, where does responsibility sit?
Predictive analytics for population health. In theory, AI should be able to identify which patients are at risk of hospitalisation, which populations need targeted interventions, which chronic conditions are trending. In practice, the data quality in most NZ primary care systems isn't sufficient for reliable predictions. You can't predict outcomes from data that's incomplete, inconsistently coded, and structured around billing rather than health.
Personalised care plans. AI-generated care plans sound compelling. But primary care is inherently personal. A GP who's known a patient for ten years has contextual knowledge that no model can replicate. AI can support that relationship with better information. It can't substitute for it.
The Data Quality Problem
This is the issue I keep coming back to, and it's not new. I was dealing with data quality problems at RAPHS years before anyone was talking about AI in primary care. But AI makes the problem more visible, because AI requires good data to function, and good data is exactly what most primary care systems don't have.
~40%
of clinical data in NZ primary care systems contains inconsistencies in coding or classification
Source: Health Informatics New Zealand, Data Quality in Primary Care Report, 2022
The reasons are structural. GPs are under time pressure, so they use shortcuts in documentation. Different practices use different coding conventions. Patient management systems allow free-text where structured data would be more useful. Legacy data from system migrations is often incomplete or incorrectly mapped.
None of this is anyone's fault. It's the accumulated result of building systems under constant time and resource pressure. But it means that any AI application that depends on historical data quality, which is most of the interesting ones, will underperform until the underlying data is addressed.
What NZ Primary Care Should Do
I'm not going to suggest that every practice needs an AI strategy. That's premature for most. But there are practical steps that make sense now.
Fix your data first. Before investing in AI tools, invest in data quality. Clean up coding inconsistencies. Implement structured data capture where you're currently using free text. Establish baseline data quality metrics. This isn't glamorous work, but it's the foundation for everything that comes later.
Start with administrative automation. The ROI is clear, the risk is low, and it frees up time for clinical work. Don't try to solve diagnosis or clinical decision support first. Solve the paperwork problem.
Be sceptical of vendor claims. If a vendor tells you their AI tool will reduce GP workload by 50%, ask for evidence from New Zealand practices. Not US hospitals. Not UK trusts. NZ general practice, with our funding models, our patient management systems, and our workforce constraints. The context matters more than the technology.
Think about governance early. AI in health generates new questions about data use, patient consent, clinical responsibility, and algorithmic transparency. Most practices aren't thinking about this yet. The ones that start early will be better positioned when regulation catches up, which it will.
Don't wait for perfect. None of this means primary care should ignore AI. The tools are getting better rapidly. But adopt them for what they actually do well today, not for what demos suggest they might do tomorrow. Start small. Measure outcomes. Scale what works.
Primary health is where AI will eventually make the biggest difference in New Zealand. It's where the data lives and the decisions happen. But "eventually" is doing a lot of work in that sentence. The organisations that succeed will be the ones that take the boring steps now, data quality, workflow integration, governance, so they're ready when the technology catches up to the promise.