AI Is Everywhere in Hospitals-Now Comes the Proof and the Price

AI Is Already on the Floor: What OpenEvidence Signals for Healthcare

If you want proof that AI has moved from hype to the hospital floor, look at OpenEvidence. It's a medical search and Q&A tool grounded in peer-reviewed sources like The New England Journal of Medicine, JAMA, and others. Launched in 2023 via Mayo Clinic's Platform_Accelerate program, it has spread fast - you'll see it open at nursing stations across the US.

It's not an outlier. Hospitals are activating ambient AI documentation (scribes) because writing notes burns clinicians out. Revenue cycle automation is rolling, driven by prior auth and denial reduction. Next up: AI agents that can run multistep tasks without constant prompting.

The Bubble Talk - And Why Healthcare Feels the Rush

Major institutions have warned about an AI bubble. At the same time, tech firms have committed more to AI data centers than it cost to build the US interstate highway system over four decades. That spending needs payback.

• Industry math: AI needs roughly $2T in annual revenue by 2030 to justify infrastructure spend; some estimates say the gap could be ~$800B short.
• OpenEvidence raised ~$200M at a ~$6B valuation; projected revenue is ~$50M in 2025 and ~$100M in 2026. That urgency you feel? It's linked to revenue lagging valuations.
• Healthcare AI spend was ~$26-29B in 2024, projected ~$39B in 2025, with long-range forecasts into the hundreds of billions by 2032.
• Cost reality: Enterprise AI scribes often list around $600 per clinician per month, plus setup and tiers. A 500-clinician rollout lands in multimillion-dollar territory annually before services.
• Revenue cycle AI gets approved faster because it's sold on denial reduction and financial lift.

The result is a hard push to deploy now - not because all the tech is flawless, but because the business case needs to materialize.

What AI Does vs What Healthcare Needs

Platform lock-in is the quiet force shaping decisions. If a scribe or revenue tool is embedded in the EHR, it's easy to switch on - and hard to swap later after workflows, templates, and billing adjust around it.

Clinicians often like scribes because they feel the time savings and lower cognitive load. The risk is defaulting to "what's in the box" over better third-party tools. The bigger risk: letting documentation speed and coding improvements become the scoreboard while outcomes turn into a footnote.

A better bar is simple: Does the model do what the developer says it does on new, external data? Public benchmarks and outside testing should be the norm. And hospitals should track what matters for patients (usage patterns, clinician responses, safety parameters) - not just dashboard vanity metrics.

Keep Control of Your Stack

• Interrogate lock-in. If an EHR-native tool isn't best-in-class, negotiate swap rights and data portability up front.
• Pilot and compare. Run head-to-head pilots with clear metrics: time saved, note quality, error rates, user satisfaction, and downstream clinical signals.
• Demand external validation. Require performance on held-out, local data and published benchmarks.
• Align incentives. Favor vendors who tie success to patient outcomes and coordination - not just RVUs and speed.

Data, Contracts, and the Hidden Costs

"HIPAA compliant" is a comfort phrase. The real test is the business associate agreement (BAA) and the vendor's exact rights to use, de-identify, and reuse your data. If data are de-identified and leave HIPAA, they fall into a patchwork of state rules - and patient expectations may not align with commercial reuse.

Build protections up front: indemnity if the vendor de-identifies, explicit subvendor lists, downstream deletion duties with attestations, and real audit rights. Make it boring, precise, and enforceable.

Then there's the grid. AI isn't free to run. A single data center can draw enough power for 100,000 US homes, and bigger builds are projected to use far more, per the International Energy Agency. If your cloud model spikes compute at peak hours, your energy bill will feel it. On-prem? Cooling and water become part of total cost of ownership.

Also beware overbuilding and stranded assets. If you lock into long, fixed commitments without an exit, you'll lose leverage when capacity and pricing normalize.

A Simple Procurement Checklist

• BAA specifics: PHI scope, purpose limits, de-identification rights, and reuse prohibitions.
• Subprocessors: full registry, approval rights, and flow-down terms.
• Data lifecycle: retention, deletion timelines, and certified destruction.
• Validation: external benchmarks, local holdout testing, ongoing drift monitoring.
• Energy/TCO: peak usage modeling, cooling/water costs, and power price sensitivity.
• Exit: portability, conversion support, and termination-for-convenience without punitive fees.
• Governance: human-in-the-loop, bias/safety review, incident response, and audit access.

If the Bubble Deflates

A market correction would hurt, but it likely won't break care delivery. Expect a shakeout, not a shutdown. Scribes and revenue tools will probably persist because clinicians feel the benefit and finance teams see the return.

Hospitals could face paused rollouts, tougher privacy terms, higher energy costs, and cuts to tools that can't show outcome gains. Contracts will matter. Data trust will matter more.

What to Do Next Week

• Set outcome metrics now: documentation quality, patient safety signals, readmissions, and clinician time-on-task.
• Run 90-day scribe pilots with 2-3 competing vendors. Measure accuracy, edit time, note completeness, and downstream billing impacts.
• For clinical decision support, hold the line until there's clear guardrails and external validation on your population.
• Bake data terms into every deal: BAA clarity, de-identification indemnity, subvendor transparency, and deletion attestations.
• Model energy and cooling costs with facilities and finance before committing to cloud bursts or on-prem hardware.
• Negotiate exits and portability. Avoid long, fixed commitments without performance outs.
• Educate your teams. Consider focused AI training paths for clinical, IT, and finance leaders to make better procurement calls - see courses by job at Complete AI Training.

The Bottom Line

Money wants results now, but patient safety is the only metric that endures. If you demand evidence and protect your data, this wave becomes progress. If you chase promises, it becomes waste. The test is simple and unforgiving - and it's already underway.