Beyond the Pilot: a practical framework for building a scalable AI roadmap in your health system

Most health systems have dipped a toe in the AI water by now. A pilot is launched, it works technically in its contained corner, a slide is presented – and then nothing scales. The frustration echoes through boardrooms and hospital corridors alike: why do so many promising AI initiatives die in the "pilot trap," never delivering system-wide value?

The short answer is that moving AI from contained experiment to enterprise-wide change takes more than an effective algorithm. It takes a shift from tactical trials to strategic integration. HFS Research has recently described widespread "pilot fatigue" in exactly these terms: the value a pilot demonstrates is never tied to the organisation's core strategic measures, so leadership files AI under minor efficiency gains rather than genuine transformation. From what I have seen navigating AI adoption inside complex healthcare environments, escaping the trap comes down to four moves.

First, tie every initiative to a strategic imperative. Pilots stall when the link between the project and the organisation's mission is missing. So redefine what "value" means before you start. Not task efficiency in a sandbox but measurable movement on the goals the board already cares about (a specific clinical outcome, patient throughput, readmissions, patient experience scores, a new revenue stream). When AI's contribution is stated in the language of core objectives, the case for scaling makes itself. And before scaling any pilot, ask the strategic-fit questions honestly. Does it address a critical bottleneck? Does it serve a growth priority? If the answer is unclear, scaling is premature – however well the pilot ran.

Second, build the foundation before the ambition. A brilliant algorithm running on inadequate infrastructure, fed by poor-quality data, will fail at scale every time. Messy data and disconnected systems are the silent killers of AI ambition – HFS Research and others keep finding exactly this. So begin with an honest audit of the current state, of the kind the World Economic Forum has proposed. Is your data accessible, interoperable (FHIR standards help), and governed? Is it accurate, complete, and representative enough to avoid bias? Can the technology stack actually carry AI tools at scale? And do your clinical, IT, and operational teams have the skills to implement and manage them? Then invest accordingly. Scaling AI is not a software purchase – it usually means building out the digital core itself (cloud integration, federated data architectures, data quality pipelines). That spending looks like overhead. It is the plumbing for every ambition that follows.

Third, decide what to scale, and govern it. With alignment and foundations in place, resist the pull of scaling whatever pilot happens to be easiest. Rank initiatives by system-wide impact, strategic fit, technical feasibility, and manageable risk. Treat that ranking as a leadership decision, not an IT one. Put the governance in place before wide deployment, not after the first incident. Who is accountable for an AI output? How will bias be monitored? How will transparency be preserved when the algorithm is a black box? How will patient privacy be protected? Legal, clinical, and IT voices all belong in that room. And bring the users in early – clinicians, administrators, IT teams, even patients (the WEF makes the same recommendation). Understand their workflows. Hear their concerns. Co-design the integration. A tool forced onto unwilling or unprepared users rarely achieves anything.

Fourth, face the human factor. The tallest barrier is usually not technological. Fear of disruption, unclear ownership, siloed decisions, and plain resistance to change can cripple the best-laid roadmap. The counters are unglamorous. Visible C-suite sponsorship so the organisation knows this matters. Clear accountability (who owns the strategy, who delivers it, how decisions get made), because ambiguity breeds inaction. And real investment in change management: training, communication, support, and honest answers to staff worried about their roles.

None of this is theoretical for me. I have lived these challenges from three seats (as a clinician, as a hospital system CEO, and now as an AI builder). The pattern repeats across all of them: the algorithm is never the hard part. The hard part is the organisation around it. A framework points the way. Applying it means translating it into concrete moves inside your specific environment – your data, your people, your constraints.

If your leadership team is wrestling with that translation, I am happy to compare notes – I can be reached at drajaybakshi@gmail.com.