GLG Frontier
>
Research

Inside the Hospital AI Approval Process: What Gets Through and What Doesn't

Hospitals are buying AI faster than they can govern it. Five practitioners – including a sitting chief medical informatics officer and a former hospital COO – explain who signs off, what gets checked, and why some tools scale while others stall after the contract is signed.
Hospitals
Frontier Briefing
GLG Frontier
November 2026
CONTRIBUTORS
GLG Frontier Research Team

For hospital operators who hope to deploy AI tools, procurement is just the tip of the iceberg. Getting staff to keep using the tool, proving the return, and deciding who answers for it when something goes wrong represent the more critical hurdles today.

The practitioners we spoke to agree on broad patterns. Administrative and operational AI – documentation, coding, prior authorization, scheduling – moves fastest, because the risk is bounded and the benefit is easy to measure. Clinical AI moves more slowly, because it affects patients and needs stronger evidence.

When it comes to the machinery approving AI deployment, the practitioners together provide a more complicated picture.

According to the CMIO, governance is trailing adoption in the U.S. – many systems that took on AI tools quickly over the past year or two are only now building oversight around them. “I would contend that we’re still trying to play a little catch up here,” he said.

Who Decides: A Life Cycle, Not a Vote

The practitioners describe a standard approval life cycle run by a cross-functional group. The former information services strategy executive outlined the stages a mature process covers:

  • Use-case intake, starting from a costly workflow problem rather than a product.
  • Risk classification, which separates bedside tools from back-office ones.
  • Validation and approval against the system’s own data and standards.
  • Deployment monitoring once the tool is live.
  • Incident response when outputs go wrong.
  • Retirement, so tools don’t pile up indefinitely.

Governance has to spell out where automation stops and a human takes over, and who exactly is accountable for the workflow. A use case with no accountable owner is, in his account, one of the most common reasons a pilot never scales.

The pace of change adds pressure to this process. Each new feature, which can ship monthly, reopens the question of whether the tool still meets the system’s clinical efficacy standards, the CMIO noted. A newer category of AI governance vendors has appeared to help systems harmonize oversight across many point tools without relitigating previous decisions.

The Criteria: Workflow Fit, Security Before Everything

The experts highlighted many of the same evaluation criteria, including workflow fit, clinical or operational evidence, integration with the electronic health record (EHR), security and privacy, and total cost over the tool’s life. A polished feature list ranks well below how well the tool fits the actual work.

They all highlighted security as a non-negotiable. The former information services strategy executive put it most plainly: “security is an architectural requirement and vendor selection criteria, not a final approval step,” he said.

Cost goes well beyond the license. Buyers count integration, local validation, training, workflow redesign, cloud and token consumption, security monitoring, and ongoing support – all while weighing a vendor’s ability to survive in the long term.

Evidence: Nobody Wants to Go First

For clinical tools, evidence comes in sequenced order. The former health system digital leader described what hospitals look for: a reference site that went first and produced real-world results, published retrospective validation, access to the vendor’s clinicians and engineers, an FDA clearance or breakthrough designation, and a credible return case. The last step is the hospital’s own sandbox test on its retrospective data, then a small prospective cohort.

Who produces the evidence depends on the tool. The former medtech governance leader said health systems often run their own trials of ambient documentation but rely on vendor evidence reviewed by regulators for diagnostic AI, because local validation is costly and needs data traceability many systems lack. The CMIO added that even early wins need validating before a program expands.

Liability Lives in the Contract

Practitioners agree that liability remains with the clinician. “These tools only can help you influence decisions. These do not replace decisions,” the CMIO said. The former medtech governance leader noted that liability for AI-assisted care still runs through existing professional standards and device law, not AI-specific statute.

A vaguer liability question stalls many hospital AI contracts, the former COO said: after a data breach or technology outage, who owns the response? Larger systems have more to protect, so they ask for insurance limits that small vendors may not be able or willing to carry. The former medtech governance leader described other terms buyers write in, such as notice before the model changes, performance obligations, and clear conditions for switching suppliers, ideally from a panel of comparable vendors.

What Gets Approved: Ambient Documentation and Bounded Back-Office Work

The former COO ranked ambient documentation as the top area of hospital AI adoption. Likewise, the former digital leader, who led an early rollout at a health system, said any remaining pilots should convert. “That’s the ROI layup of all ROI layups.”

It clears approval for structural reasons. The clinician who reviews and signs the note also gets the benefit, spending less time charting and more time face-to-face with patients. Ambient scribing has also become a recruiting and retention issue, as clinicians increasingly expect it.

The bar to pass approval continues to rise. The CMIO said the baseline now includes deep EHR integration, specialty-specific workflows, and reach beyond outpatient visits into acute care and nursing. Vendors limited to outpatient physicians, he expects, will hit a ceiling. Vendors that stay hands-on past go-live, when clinicians start asking for more, differentiate themselves.

The tools getting through tend to share a profile: bounded, measurable results, and mostly away from the bedside. The CMIO pointed to outpatient autonomous coding, prior authorization, after-hours patient calls, and scheduling. For patient-facing agents, the former digital leader said one design choice matters most: an easy, visible way to reach a human.

What Stalls, and Why

According to the interviews, approved tools fail to scale for five common reasons:

  • Pilots and production come from different budgets. Pilots run on innovation budgets, while enterprise rollout needs recurring funding. The former medtech governance leader said a business case should identify who will pay over at least three years.
  • Pilots skip the data work. Data quality and governance get deferred during a pilot, and the rebuild costs surface at scale.
  • Return cases assume clean automation. The former COO said pro formas often count staff savings while ignoring the exceptions that fall out of AI workflows, which still need a human owner.
  • Use drops after go-live. Training tends to end at launch. Coders who hit early errors stop trusting a tool, and the CMIO said clinicians who stumble once may walk away unless someone follows up quickly.
  • The payment model works against it. The former digital leader said AI that prevents downstream procedures can cut fee-for-service revenue, so value-based care organizations adopt it far faster.

Clinician resistance follows a similar logic. The former medtech governance leader said pushback concentrates where the physician carries the risk but not the benefit, which helps explain why diagnostic AI moves slower than ambient scribing. The former information services strategy executive summed up the gap: “many pilots prove that the technology works without proving that the organization can absorb it,” he said.

Our Take: What This Means for Teams Building AI for Hospitals

The approval process rewards products built to be governed, not just products that score well. A few questions worth testing before a hospital deployment:

  • Can a governance team see exactly what changed in each release and rerun its own checks before clinicians see the change?
  • Can a hospital evaluate the product in a sandbox on its own retrospective data, with results a named clinical owner can defend?
  • Is the point where automation stops visible in the interface, logged, and easy for a user to trigger?
  • Do your evals measure sustained use at day 90, including how users respond after the first visible error?
  • Does every exception land in a queue with a clear owner, so the return case survives real workflows?
  • Does the value reach the clinician who carries the risk and the organization paying under its own payment model?

None of these can be answered from model metrics alone. The thresholds, failure cases, and escalation rules come from people who have sat on hospital AI committees, run go-lives, and negotiated vendor contracts. That experience-based judgment is what gives hospital leadership and staff trust and confidence in an evaluation.

Go Deeper

This Briefing draws on five GLG Library interviews: AI in Health Systems (September 29, 2026), AI in Healthcare IT: Health System Perspective (August 19, 2026), Trends in Hospital Operations (August 18, 2026), AI in Healthcare (September 28, 2026), and AI Governance and Enterprise Adoption Dynamics in Health and MedTech (September 28, 2026). GLG clients can read the full transcripts in myGLG. To pressure-test a hospital AI product, GLG can connect you with CMIOs, health system IT leaders, and hospital operators through a one-on-one expert call. Teams building evaluations for clinical or operational AI can work with GLG Frontier to design benchmarks grounded in real practitioner judgment.

About this piece: This Frontier Briefing draws on five GLG Library interviews recorded in August and September 2026. The experts are the chief medical informatics officer (CMIO) of a large multistate U.S. hospital operator; a former information services strategy executive at a large U.S. nonprofit health system; a former hospital chief operating officer and regional health system executive; a former health system digital leader who later worked in virtual care; and a former enterprise data governance leader at a global medical technology company, who spoke mainly about Asia-Pacific markets. Their accounts of how hospitals evaluate and govern AI, and any recommendations to health systems, are theirs. The builder takeaways are GLG Frontier’s.

Get connected to the world's leading platform for trusted human expertise

Enter your information below to get in touch with our team.