Healthcare AI Investments Face Market Correction and ROI Crisis

Healthcare organizations racing to deploy artificial intelligence systems now confront uncomfortable evidence: the enterprise AI market exhibits characteristics of a speculative bubble approaching correction, and most institutions lack the infrastructure to weather what comes next.

The warning signs extend beyond isolated failures. According to IBM research cited in a new National Academy of Medicine commentary, only 25% of AI initiatives delivered expected returns over the past three years, with just 16% scaled enterprise-wide. More striking, the share of companies abandoning most AI projects jumped from 17% in 2024 to 42% in 2025, even as 91% of organizations plan increased AI spending.

Why it matters

The impending correction will expose which healthcare institutions built AI capabilities aligned with patient outcomes versus those that chased market narratives. Organizations unprepared for this shift face not just financial losses but opportunity costs—resources spent on unvalidated technologies represent dollars and clinician hours diverted from interventions with proven clinical impact.

Digital health funding masks structural weakness

Healthcare AI follows the broader enterprise pattern. Digital health funding appeared stable at $3.5 billion in Q3 2025, but megadeals now account for 39% of total funding while Series B deal flow—the critical stage where startups must demonstrate commercial viability—thinned to just 30 raises through Q3 2025, compared to more than 60 annually over the previous four years, according to Rock Health data.

The pharmaceutical sector reinforces the trend. When Recursion Pharmaceuticals and Exscientia merged in August 2024 following a 78.7% valuation collapse, the combined companies had consumed hundreds of millions without producing a single FDA-approved drug.

Seven diagnostic questions

The commentary, authored by a medical AI researcher whose work predates the current enthusiasm cycle, offers institutions seven questions to assess their readiness:

• Can you independently validate vendor claims on your own patient population?
• Do AI investments address clinical outcomes or primarily workflow optimization?
• Have you ever declined to deploy a technically functional AI system that lacked clinical evidence?
• When AI systems fail, can you identify the reason and determine accountability?
• Do investments address problems patients care about or problems administrators prioritize?
• Have you built institutional capacity to evaluate AI or just procurement relationships?
• Can you articulate stopping criteria for systems currently in production?

Organizations unable to answer affirmatively have built dependencies on vendor promises rather than sustainable capabilities, the analysis warns.

The workflow optimization trap

Studies of over 2.5 million uses of ambient documentation demonstrate measurable value—approximately 20 minutes daily time savings and reduced burnout. Yet this represents automation of clerical tasks, not advancement of clinical mission. Organizations built exclusively for workflow optimization helped staff efficiency but did not fundamentally improve patient care.

The correction will not primarily harm institutions financially. The deeper cost is opportunity: every dollar spent on AI that fails to improve patient outcomes is a dollar not spent on interventions that do. Healthcare resources are finite, and their allocation reflects institutional values.

Organizations that demanded evidence, validated independently, monitored rigorously, and focused on outcome-relevant problems will emerge with infrastructure that continues serving their mission regardless of market conditions. Those that chased promises will face disappointed clinicians, patients, and the realization that deployment is not validation.

The analysis was first published by the National Academy of Medicine in its Perspectives series.