Waiting for perfectly clean data is stalling healthcare AI, not protecting it

Ask a hospital CIO why the AI program has not started yet and one answer comes up more than any other: the data is not ready. Records are fragmented across electronic health records, claims platforms, clinical systems and departmental applications, and nobody wants to train a model on a mess. Deepak Mittal, founder and CEO of NextGen Invent, hears that reasoning constantly and considers it the single most costly assumption in healthcare AI right now. His response to executives is a question rather than an argument.

Nobody has. So a health system that waits for clean data is, in practice, deciding never to start. Mittal's point is that the assumption was true at one point and is not true today. A project can begin on a reliable subset of the data, lean on synthetic data, or put AI itself to work finding and fixing quality problems. Data integrity still matters, especially where clinical or financial decisions ride on the output. But the sequencing has flipped: a carefully chosen project can move forward while the data work continues. He says preparation that once took three months now takes three hours with his firm's accelerators. NextGen Invent, recently acquired by Straive, has built more than 500 AI models since its first in 2010, which is why the claim carries weight.

Start with the problem, not the tool

The clean-data excuse is one symptom of a broader habit: starting with the technology instead of the result. Mittal says the question he hears most from executives is "What can AI do for us?" and he treats it as a marker of low AI maturity. The better question is what success looks like for the organization, and only then how AI helps get there. Without that clarity, even an impressive pilot struggles to win adoption or show a return.

It also explains why so many pilots never scale. Mittal points to organizations that chose the wrong first problem: they lacked the data, or lacked the maturity, and still picked a mission-critical application to make a statement. Prioritizing by effort, value and likely adoption gets a program to its first win. Leadership commitment, training, change management and culture then have to develop together. As he puts it, strategy can't be strategic without commitment.

Governance moves to the front of the process

Most organizations treat governance as something to bolt on after selecting a product. Mittal argues it belongs in product strategy, before the system is designed or purchased. That matters more as healthcare moves from generative AI, which organizes information so a person can decide, to agentic AI, which acts on that information. In revenue cycle management, he describes an agent that can handle roughly 90% of denials and prepare the appeals, routing the remaining 10% to a human. The goal is for that 10% to shrink over time as the agent learns, which he compares to an intern becoming an associate. Deciding which actions can be automated, when a person must review, and who stays accountable cannot wait until after implementation.

The same discipline applies to the business case. Mittal names five levers AI can pull: increase revenue, reduce cost, create a competitive edge, build new products or services, and strengthen the organization's influence across the broader ecosystem, including how it shares data. Not every project hits all five, but leaders should know which one they are buying before the check is written.

None of this removes the clinician. Mittal is direct that he would not let an AI make a cancer decision for him ten years from now; he would want a doctor using AI as one more tool alongside imaging and records. Stevenson agrees that major clinical decisions need a human in the loop. The work ahead for health systems is less about buying more technology and more about breaking assumptions: that data must be perfect, that governance can come later, that the boldest use case should go first. Clear those, and adoption, scale and measurable return have a path to follow.