Why most enterprise AI programmes never leave the lab. And the three shifts that get them out.

Walk into almost any health system in the US right now, and you will find an AI strategy. You will probably also find a steering committee, a vendor shortlist, and a handful of pilots in various states of stalled. What you will rarely find is AI running in production, embedded in clinical or revenue workflows, generating measurable returns at the scale the business case promised.

The numbers tell the story. 70% of providers and 80% of payers now have an AI strategy in place or in development, up sharply from 60% the year before, according to the 2025 Bain and KLAS Healthcare IT Spending study. 

Yet a separate KLAS report covering 1,742 healthcare organizations found that while nearly every organisation interviewed is piloting AI, very few have scaled it across departments or functions. Of more than 3,000 respondents, only 17 mentioned agentic AI by name, and exactly one had it in production.

This is the pilot trap, and it has less to do with technology and everything to do with the foundation it sits on. 

Why good strategies stall in execution

The conventional explanation for stalled healthcare AI is that the technology is too new, the regulators too cautious, or the use cases too risky. None of those is wrong. They are just not the binding constraint.

The binding constraint is that healthcare enterprises are trying to deploy AI on top of operating environments that were never built to support it. KLAS describes the current state of the industry as «early scale-up». Essentially, active but not yet durable. 

Without a foundation, organizations risk stalled pilots, suboptimal results, and unrealised ROI. Deloitte makes the same point at the enterprise level: Legacy data and infrastructure architectures cannot power real-time, autonomous AI, and 35% of AI leaders cite infrastructure integration as their single biggest barrier.

Translated for healthcare, that means four familiar problems are now blocking AI value creation, not just operational efficiency:

• Fragmented patient journeys. Care happens across primary care, specialists, hospitals, post-acute settings, pharmacies, and increasingly the home. Each touchpoint runs on a different system, with a different data model, and a different definition of the same patient.
• Data silos and weak interoperability: Roughly 70% of providers still struggle with seamless data exchange across platforms, per HIMSS data. AI models are only as useful as the data they sit on top of, and most enterprise data is locked in EHRs, claims systems, and departmental tools that were never designed to talk to each other.
• Regulatory and compliance overhead: HIPAA, state privacy laws, the new CMS interoperability and prior authorisation rule (CMS-0057-F), and the emerging patchwork of governance requirements specific to AI all add real cost to every deployment. KLAS reports that the conversation among healthcare leaders has shifted from «what can we buy?» to «how do we control it?»
• Legacy tech debt: Decades-old core systems, custom integrations, and brittle middleware mean that even the most valuable AI use cases get stuck in months-long integration cycles. The pilot works; the path to production does not.

Add these together, and you get the plateau. Pilots are cheap because they bypass the foundation. Production is expensive because it cannot.

Why the standard playbook makes healthcare AI worse

Most enterprises respond to the plateau by doing more of what is not working: more pilots, more vendors, more proofs of concept. The logic is reasonable. If one experiment did not scale, run five more and pick the winner. The problem is that none of those experiments were ever going to scale, because none of them addressed the system underneath.

There is also a category error baked into how many programmes are structured. Too much of the work is treated as advisory: strategy decks, target operating models, governance frameworks, and vendor selection. All useful. None of it ships software into clinical or revenue workflows. The AI sits in the proof-of-concept environment, the slides get presented, and the budget cycle resets.

The pilot trap is not a technology problem. It is a problem of treating AI as a project when it needs to be treated as infrastructure.

Three shifts to get healthcare AI into production

Across the healthcare organizations Wizeline works with, the ones that move from pilot to production make three changes early. They are not glamorous, but they are the changes that compound.

1. Shift from pilots to platforms

Stop treating AI as a series of isolated experiments. Start treating it as a shared production platform that multiple use cases plug into. Say, a common data layer, a common model serving and monitoring layer, a common compliance and audit trail. Each new use case then becomes a configuration on top of the platform, not something that’s built from scratch.

This is how organizations get from one stuck pilot to ten production deployments without ten times the cost. It is also how they make compliance scalable. Audit trails, model versioning, and data lineage live in the platform, not in each project.

2. Shift from advisory to engineering ownership

Healthcare AI does not fail at the strategy stage. It fails between the strategy and the running system. That gap is closed by engineers who own the outcome, not consultants who hand off the deck.

In practice, this means embedding product-minded teams who can take an AI use case from problem definition through to a deployed, monitored, integrated system inside the EHR or the revenue cycle workflow. It means measuring the partner on what is in production, instead of what is in the roadmap.

3. Shift from greenfield AI to workflow-embedded AI

The AI use cases that scale in healthcare are the ones that disappear into the workflow. Ambient documentation works because clinicians do not have to change what they do. Coding automation works because it lives inside the existing revenue cycle management (RCM) process. The pattern is consistent: AI delivers ROI when, instead of asking the organization to adopt a new one, it transforms a workflow that already exists.

The implication for technology leaders is to start with the workflow, map where AI removes a constraint, and engineer the integration end-to-end. Greenfield AI products are easier to demo. Workflow-embedded AI is what actually moves the operating margin.

What Healthcare ^ AI looks like in practice

Three patterns from recent work illustrate the model.

Medecision: Re-architecting a population health platform so that AI-driven care management could be embedded directly into care manager workflows, rather than running as a parallel analytics layer. The platform shift unlocked use cases that had been blocked by data and integration constraints for years.

Waystar: Engineering AI built for production into AI into revenue cycle workflows where every claim, code, and denial has hard-dollar consequences. RCM is exactly the workflow KLAS identifies as the highest-priority AI investment area for providers in 2025. And exactly the workflow where pilot-grade AI cannot survive contact with real volume.

Azenta: Building data and AI infrastructure for life sciences sample management, where compliance, traceability, and integration with lab systems are not features bolted on at the end but the core engineering problem from day one.

None of these are pilots. They are systems running in production, embedded in workflows that the business already depends on. That is the shift.

The bottom line

Healthcare AI is not failing because the models are not good enough. They are. It is failing because most enterprises are trying to deploy AI built for production on foundations that were only ever built for pilots, with consultants leading delivery rather than engineers, on workflows that were never re-engineered to absorb it.

In practice, a health system deploys an AI-powered prior authorization tool. The model is accurate. But it sits outside the EHR, requires case managers to re-enter data by hand, and was stood up by a consulting team that handed off a deck and disengaged. Six months later, adoption is low, and the business case has evaporated. Not because the model failed, but because nobody owned the workflow it was supposed to live inside.

The organizations that escape the trap in the next 12 to 24 months will be the ones that stop running more pilots and start building the infrastructure, the engineering capacity, and the workflow integration that makes AI a permanent part of how the business runs.

The technology is ready. The question is whether the operating model is.

If your AI roadmap looks ambitious on paper but your production deployments are thin, it is worth a conversation.

Wizeline partners with healthcare and life sciences enterprises to engineer AI into production, from foundational platform to systems embedded in existing workflows. Get in touch to talk through where you are stuck and what it would take to get unstuck.