Why healthcare AI stalls after the pilot
Pilots die in boardrooms
Most healthcare AI pilots die in boardrooms, not emergency departments.
The technology works. Clinical teams want it. But when it reaches the capital allocation committee, the conversation stalls.
McKinsey & Company latest survey of 150 US healthcare leaders makes this visible. 47 percent have already implemented generative AI. Another 38 percent are pursuing proofs of concept. Yet only 64 percent of those who deployed are seeing positive returns.A third of implementations are not delivering. The gap between pilot and impact keeps growing.
This is not a technology problem. It is a governance and capital architecture problem.The organisations seeing returns are not the ones with the best algorithms. They are the ones who designed procurement, governance and payment alignment for scale from day one. McKinsey’s data confirms it. 61 percent are choosing to partner rather than build or buy. They want infrastructure relationships not software licenses.
I sit between European and Nordic innovators building these tools and Gulf institutions deciding where to deploy capital. I also work as an NHS England GP, so I see the clinical reality daily. Where time disappears into documentation. Where referrals wait because capacity is trapped in processes not patients. The pattern from both sides is consistent. The technology is rarely the bottleneck. The bottleneck is whether governance was designed for a pilot or designed for a health system.
Sovereign capital has a structural advantage here that most investors cannot match. The mandate to think in decades not quarters. The scale to demand enterprise grade governance before writing a cheque. The convening power to align ministries, providers, regulators and technology partners before deployment begins.
Healthcare AI will not transform systems through pilots. It requires patient capital and cross stakeholder coordination that sovereign wealth funds are built for.The question for Gulf investment leaders is no longer whether healthcare AI works. It is whether your governance model will let it.
The Tuesday afternoon after the project ends
A promising pilot still leaves me with a question: what has to stay in place for patients to keep benefiting?
Imagine a Tuesday afternoon after the project ends. The clinic is short-staffed, a locum is using the tool for the first time and the person who used to sort out every problem has moved on.
That is a situation I'd want the pilot to prepare for.
Extra attention during a pilot is sensible. Protected staff time, training and quick access to the supplier may be exactly what a new service needs. The problem comes if the plan assumes those things can disappear without changing the result.
Before expanding, I'd put the support used during the pilot on one page. Mark what was temporary, what needs to continue and who will pay for it. Include the informal work, such as the nurse chasing missing information or the founder answering troubleshooting calls.
Then ask the team expected to run the service whether that plan is workable. A name beside a responsibility means little if that person has neither the time nor the authority to carry it out.
A pilot can provide real evidence of benefit. How that benefit will be sustained deserves attention before the project ends.
I'd be more persuaded by a modest result with a credible plan for continuing care than an impressive result whose support costs have quietly been left out.
Friction, not technology
The biggest barrier to AI adoption in healthcare isn’t technology. It’s friction.
But a new trial in JAMA Network Open shows that we can overcome this more easily than many think.
The study combined a short patient completed assessment with an AI review of routine primary care records. The outcome was striking: substantially more cases of dementia were identified across several thousand older adults and this happened without extra appointments, new tests, or added workload for clinicians.
This is the kind of progress that matters.
It isn’t “high-tech disruption”; it’s a smarter, more responsible use of information that already exists. For overstretched health services, that combination , meaningful clinical impact with minimal operational friction , is exactly what scalable clinical AI should look like.
For NHS leaders and policymakers
Early adopters of approaches like this won’t just improve the timeliness and equity of cognitive care. They will shape national expectations for what safe, evidence based AI integration looks like. The next steps are clear: governance pathways, pilot sites and embedding tools like this directly into existing digital workflows.
For innovators and investors
This study signals where real HealthTech value now lies. Health systems will reward workflow-native solutions that:
- scale easily
- use existing data
- respect clinical workload
- deliver measurable impact without increasing burden
These are the tools that succeed ,not because of hype, but because they solve real problems in a way that fits how care is actually delivered.
Reference
What it means
Taken together, these pieces give a simple test for any health AI proposal: was the governance designed for a pilot or for a health system, who pays for the support once the project team leaves, and does the tool add work or remove it?