Scaling AI in health: aligning patient, regulatory, and commercial incentives
Open sourcing notes from our latest sitrep, with founders, policymakers and VCs
Breakthrough progress in regulated markets requires a rare alignment between the interests of startups, investors, policymakers and the public. Take health: AI can transform triage, diagnosis, treatment and therapeutics, delivering personalised care at population scale. But there’s work to do to make that a reality.
So last week we hosted another sitrep breakfast, again with our friends at Eka, convening founders, VCs, regulators and policymakers to discuss the state of play for scaling AI in health.
Just like our last event on energy tech and policy, here we’re opening up our notes from the conversation, featuring perspectives from startups, early and late stage investors, and policy perspectives from folk (formerly) at MHRA, NICE, and DHSC.
Read on for the readout, and hit reply if you’d like to join the network we’re building at Form convening the frontier of technology and policy. And thanks to Molten for hosting us!
Full readout:
End-to-end, vertical integration of care pathways can solve the challenging interplay of product, commercial and regulatory strategy: Startups have to decide between building broad versus building deep, or building regulated versus unregulated, which in turn shapes their commercial opportunity. Building a fully regulated, end-to-end, vertically-integrated care journey is hard — like Flok Health’s for musculoskeletal — but signals a path for how to scale AI in an NHS setting so long as founders have strong regulatory capability as a core competency. It also reduces the reliance on clinicians to ‘use’ technology correctly and retains control of the user experience, as well as having commercial benefits to manage the full pathway.
Product velocity is a function of regulatory flex and speed: In contrast to hardware, software requires continuous iteration and therefore continuous compliance. But historically, shipping products in ‘v1, v2, vN’ methods isn’t always possible when each product upgrade needs to be certified, but there are ways to start to think about this in a more systematic way. THE MHRA is working on ‘predetermined change control plans’ to enable this, though they’re not yet operational, and many Approved Bodies have huge backlogs. Credit also to Scarlet Comply, which is building a new model of Approved Body which enables continuous deployment and iteration.
Evolving reimbursement & risk tolerance: There’s work remaining to develop strong reimbursement models to offset the up front costs required to build fully-regulated software medical devices, particularly for smaller companies. Laddering up distribution, validation and certification can be easier for larger companies, but if founders with strong regulatory capability can help shortcut some of this process to retain the advantage of early stage agility.
Breakthrough potential: We also discussed the possibility of breakthrough AI — e.g. ‘AI doctors’ autonomously managing diagnosis, or agents that might even automate clinical consultations — and whether regulatory and reimbursement structures, inside the NHS, limit future possibilities. It’s hard to acquire and train on enough quality data outside of clinical settings to push at the frontier, but also hard to deploy in the first place. Taking this one step on, there is a need to rethink & reframe how we think about ‘ability to fail’ within the healthcare system while remaining true to the tenets of patient safety.
Go-to-market incentives x NHS absorptive capacity: Startups want to build, DHSC policymakers want to improve absorptive capacity, and investors want to participate where markets are viable, but it takes work to align these incentives. Evidence collection and procurement is often siloed and needs to be consolidated into larger learning loops. Improving patient EHR systems — which may be a focus of a new government — could also reduce some of this friction. There are some positive cases to build on however, e.g. Stroke Delivery Networks which have created a network for scaling procurement of AI for stroke use cases.
Death by pilot? Or ring fenced NHS budgets for innovation: Tying back to Risk Tolerance, pilots can be few and far between, and even harder to scale from. There isn’t a natural network effect within the NHS which necessarily equates to 1) more (paid) pilots, 2) high utilization, and 3) long-term contracts. This links back to Learning Loops as we’re not able to build coherent & consistent bodies of information over time which create compounding data advantages. But national commissioning isn’t necessarily the answer, because it’s not always a viable commercial model and it can restrict competition and innovation for health services too. But there was a great suggestion to call the NHS to create ‘open innovation tenders’ where founders & innovators could be called in to solve the highest acuity problems within hospitals and PCNs.Â
Fixing regulators: Meanwhile structural barriers make it hard for policymakers to make necessary reforms, even when electoral and stakeholder incentives are aligned — particularly when not directly managing the NHS. Regulators have little incentive for risk, and are motivated more by downside protection than safely maximising the upside. The MHRA and NICE can struggle to 1) attract and 2) retain top talent, so delays have quickly become a rate limiter on startup progress.
AI-first services can improve health equity, not just risk it: There was a discussion around whether AI helps or entrenching inequalities and health disparities: how are models trained, using what data, measuring which health outcomes? But AI also has outsized potential to reduce bias provided it is built correctly, and new, tech-enabled care pathways can even be more inclusive for certain groups: e.g. the most popular time for Flok’s MSK patients is 8pm, which they can meet easily but no normal physio clinic can deliver.
Get in touch with any insights, suggestions or feedback, and if you’re building the future of regulated markets, in energy or elsewhere, pitch us here.
 | A guest post by
|