Wrong Pilot

Somewhere in a health system right now, an AI pilot is dying. Not because the technology failed. Because nobody designed it to survive contact with the organisation running it.

This is the defining AI problem for healthcare leaders. Not the models. Not the vendors. The mismatch between the pilots organisations choose to run and the structures they have in place to support them.

The failure picture is well-documented across enterprise AI broadly. According to S&P Global's Voice of the Enterprise survey of more than 1,000 business and technology leaders, the share of companies abandoning most of their AI initiatives before they reach production jumped from 17% in 2024 to 42% in 2025. Nearly half of all AI proof-of-concept projects are scrapped before anyone outside the project team ever uses them. In healthcare, where the stakes of getting it wrong extend beyond a bad quarter, the picture is no better.

The standard remedies are better planning, earlier clinical buy-in, and cleaner data. All correct. All aimed at the wrong problem.

The Selection Mistake

When a health system decides to run an AI pilot, the instinct is to start with the most valuable problem. Where is the biggest clinical need? Where is administrative cost highest? Where could AI make the most impact? That logic produces ambitious use cases. It reliably produces stalled projects.

Here is why. In healthcare, deploying any tool that influences patient care requires working through a set of approval and accountability processes before it can go live. Who authorises the technology? Who is liable if the output harms a patient? Who has reviewed it against regulatory requirements? Who signs off that it is safe to use in a clinical setting? These processes exist for good reason. They are also slow, complex, and in most organisations, not built with AI in mind.

High-ambition AI applications, such as a system that helps a doctor decide on a treatment plan or one that predicts which patients are likely to deteriorate overnight, require entirely new versions of those approval processes to be constructed before they can operate safely. That means two significant organisational changes running simultaneously: building new approval infrastructure while deploying new technology. One consistently loses. Usually the technology.

The organisations making genuine progress with AI in healthcare solved this not by moving faster through the approval process. They solved it by choosing pilots that fit inside the processes they already had.

The Example That Proves It

Picture a consultation room. A doctor is with a patient. A phone sits on the desk, recording the conversation. When the consultation ends, the AI has already drafted the clinical notes. The doctor reads them, edits where needed, and approves them before they go into the patient's record. No draft enters the permanent record without the doctor's sign-off.

This is ambient scribing, and it is healthcare AI's standout success story of the past 18 months. It is not the most consequential AI application in healthcare. Systems that assist with diagnosis, predict patient deterioration, or accelerate drug discovery are all more ambitious. Ambient scribing scaled when those did not, and the reason is exactly what that consultation room picture suggests.

The technology required no new approval pathways. No new liability framework. No new regulatory clearance. The doctor reviewing and approving every output before it entered the patient record was a process that already existed. The AI slotted directly into a structure that was already there.

The results reflect that. The ambient scribing category generated $600 million in revenue in 2025, more than doubling the prior year, according to Menlo Ventures research. Mass General Brigham began with a rollout to 20 clinicians. More than 2,500 now use the tool daily, according to Becker's Hospital Review. The US Department of Veterans Affairs piloted the technology across 10 medical centres in October 2025 and contracted its expansion to all 170 VA medical centres throughout 2026, covering a network that serves more than 9 million veterans annually.

The contrast with high-ambition AI tells the other side of the story. IBM's Watson Health was built to transform clinical decision-making at scale. It attracted substantial investment over a decade, employed thousands of people, and was ultimately sold at a significant loss relative to what was spent building it. The technology was sophisticated. The approval and accountability pathway for deploying it safely at scale did not exist. There was no point at which a clinician could review and approve each output before it influenced a patient decision. It never found a structure it could live inside.

The Question That Changes The Decision

This is not an argument for modest ambition. A technology category generating $600 million and expanding across national health systems is commercially serious. The point is that ambition aimed at the wrong target produces activity without progress.

Before committing to an AI pilot, the more useful question is not "what is the most valuable thing AI could do for us?" It is "what AI application fits inside the approval and accountability processes we already have?" Where does a human already review work before a decision is made? Where is sign-off already built into the workflow? Where is liability already established and understood? Those are the points where AI can move from test to production without requiring the organisation to first build an entirely new infrastructure around it.

The approval process is not the obstacle to your AI pilot. It is the blueprint for where to start.

What To Do

Before selecting your next AI use case, map your existing review and sign-off processes rather than your wish list. Identify the points where a human already checks work before it proceeds. Start there. As those early deployments prove their value, the organisation builds both the confidence and the structural capability to take on more complex applications over time.

The healthcare organisations making the most progress with AI are not running the most ambitious programmes. They are the ones honest enough to identify where they are ready to absorb the change and disciplined enough to start there rather than where they wish they were.

𝛀

If you found this Independent A.I. Think Tank release useful, consider subscribing to D!sruptive newsletter. Built for business leaders who need to cut through the noise and understand what AI means for their organisation.

If you need the help discussed in this article, get in touch to see how we can work with your business 🇺🇸 +1 512 518 4196, 🇦🇪+971 585 629 975 or hello@byDisrupt.com