Two billion hours short: Why healthcare can’t afford another AI pilot

Reading Time: 8 minutes

 

It is 3 a.m. on a med-surg floor. The charge nurse has two call-outs for the morning shift and a printed roster she is rebuilding by hand, phone wedged against her shoulder. Downstairs, the emergency department is holding six patients because nobody can tell them which beds will be clean and free by seven. A supply tech is walking to a different floor to find the kit that the inventory system swears is in stock. And somewhere across town, a physician is finishing yesterday’s notes at the kitchen table.

Every one of those people is doing their job as well as the system allows. But the support systems they rely on are slowing them down.

Healthcare is running out of workforce hours

According to Lumeris, the United States is roughly two billion hours in time debt: the gap between the care people need and the hours the workforce can physically deliver. That is not a gap you can close with a recruitment campaign. Even if every open nursing and physician role were filled tomorrow, the debt would remain.

As Dr. John Halamka, president of Mayo Clinic Platform, puts it: “We aren’t going to have enough caregivers to deliver the care they need. We have to use AI.”

The spending is real. The returns are not.

Healthcare has heard the AI message and is spending accordingly. Menlo Ventures estimates U.S. healthcare AI spending reached $1.4 billion in 2025, nearly tripling from the prior year, yet much of the sector is still struggling to turn that investment into scaled, measurable workflow impact.

MIT’s Project NANDA studied enterprise AI across hundreds of deployments and found that 95% of organizations are getting zero measurable return. According to the report, only 5% of integrated AI pilots can show measurable value. Healthcare faces this same challenge.

Research from Microsoft and The Health Management Academy, published through NEJM, found that 43% of health system executives are piloting or testing agentic AI, while only 3% have agents running in live workflows. Roughly one in three organizations has no plans to explore agents at all in the next year or two.

Why healthcare gets stuck in AI pilot purgatory

Healthcare doesn’t lack ambition when it comes to implementing AI, but its unique environment has made it challenging:

• Fragmented data. EHR, claims, scheduling, HR, supply, and finance sit in separate systems with different keys, different owners, and different refresh rates. The nurse rebuilding the roster is doing manually what no single system can do for her.
• Systems built to bill, not to decide. Epic, Oracle Health, Meditech, and Athena were designed to store records and generate revenue. They were never designed to power real-time operational decisions.
• Governance debt. PHI, HIPAA, 42 CFR Part 2, and state Medicaid rules mean every AI project can end up re-arguing the same questions about access, lineage and auditability. When those controls are rebuilt from scratch each time, months disappear before anyone reaches the workflow.
• Poor pilot performance. Most AI pilots start with an impressive demo, but have no path into the workflow, no named owner, and no ROI conversation with the CFO. Most fail to get past renewal.
• The talent gap. Most health systems cannot hire the data engineering and ML talent required to build every AI capability from scratch.

Put together, it’s clear why so many health systems have a portfolio of promising pilots and not a single hour returned to the floor to reduce the time debt. It is why the next phase of healthcare AI will not be won by better demos. It will be won by organizations that can govern AI once, connect it to live workflows and scale it safely across use cases.

The issue of trust: why sovereign AI matters in healthcare

Beyond all the operational challenges, in healthcare, the question is not simply whether an AI tool works. It is whether the organization can trust where the data sits, who can access it, what the model is allowed to do, and whether every action can be explained later.  So many promising pilots slow down in legal, privacy and security review before they ever reach the workflow because of this lack of trust.

Sovereign AI is the answer to the problem. It means AI is  built on a foundation the health system controls: data remains in its own environment, access follows existing identity and device policies, PHI is protected at a granular level, and every model call, data access and agent action creates an auditable record. Instead of renegotiating governance for every use case, the controls are designed into the platform from the start.

That control is important because healthcare AI will not be won by a single model or vendor. The best model today may not be the best model in eighteen months. A sovereign foundation gives health systems the flexibility to use open, closed, cloud or on-premises models without surrendering control of the data, the workflow or the audit trail. Sovereignty is what lets healthcare move faster, because compliance, security and trust are no longer rebuilt from scratch every time.

How successful healthcare AI projects move from pilot to ROI

Successful organizations start with a solid data foundation and build an AI solution around one real problem rather than a roadmap of twenty.

In practice that means four things:

1. Data from live systems is unified into pipelines that are governed once, not per project.
2. Those pipelines are mapped to the objects the hospital runs on like patient, bed, claim, provider, or shift, with the rules and actions that govern them attached.
3. Agents are grounded in that layer, so every answer traces back to a source and every action leaves a record.
4. The whole thing is delivered continuously, so it improves after go-live instead of decaying.

The other difference is who does the building. Often, consultants are employed who spend six months talking to people, with the result being a PowerPoint deck and a 120-page recommendation deck that never reaches production.

Projects that work and deliver value take a different approach, like using Forward Deployed Engineers (FDEs). These people are part business analyst, part engineer. They sit beside your charge nurses and your revenue cycle team, find the real problem rather than the stated one, and build against it in the room.

Where healthcare organizations should deploy AI first

Start where the value is obvious. The first use case should be the one that pays for the second. In most health systems it falls into one of two buckets.

Optimization

• Dynamic scheduling that matches nurse staffing to real demand by unit and shift, then extends to appointments, resources, and patient flow.
• Demand forecasting that predicts census, ED surges, and OR volume, so beds and staff are ready before the wave arrives. The Fourth of July fireworks surge is predictable, and so are most others.
• Supply chain rebalancing that gives a live view of stock, demand, and spend, traces shortages, and redistributes across floors and sites.

Automation

• A patient service engine that handles inquiries and drafts responses grounded in the record and coverage.
• Claims and revenue cycle acceleration that harmonizes coding, claims, and contract data, cutting denials work from days to hours.
• Broader process automation that mines clinical and billing workflows and deploys agents with human-in-the-loop sign-off.

Whichever you pick, three people must be involved to ensure it happens fast:

• An executive sponsor who owns the budget and clears roadblocks.
• A process expert like the head of nursing or the revenue cycle lead, who defines the business rules and holds domain authority.
• A data expert who grants and validates access.

Projects that skip one of these three are the projects that end up as slideware.

 

How SHI can help

The challenges facing healthcare affect all of us, that is why SHI is helping healthcare organizations move from AI experimentation to governed, measurable implementations faster.

SHI brings a practical advantage to this work because we sit at the intersection of the Microsoft estate, licensing, migration, data foundations and AI operationalization. That means we can help health systems move beyond an Agentic AI build and connect the outcome to the infrastructure, governance, security, commercial model and operating practices needed to scale it.

Your next step: SHI’s Agentic Camp

SHI’s Agentic Camp has a simple premise: choose one live business problem. We use your data, on your systems, running on the Palantir platform to ensure sovereign AI, implemented by our FDEs, leaving you with a production-ready agentic workflow and a roadmap to scale. The total engagement takes four weeks. The build takes just four days.

• Your data. We ingest live systems, not sample files. EHR, claims, scheduling, HR, finance.
• Your problem. You bring the highest-value pain you have. We solve that, not a made up demo scenario.
• Your people. Your clinicians, ops leaders, and IT team co-build. This is not over-the-wall consulting.
• Your win. A production use case, measurable KPIs, and an adoption roadmap for use cases two through twenty.

Agentic Camp is the entry point, not the end state. It is designed to identify the first high-value workflow, prove the value case, and create a path into repeatable healthcare AI solutions that can be deployed across additional use cases.

Conclusion

The two-billion-hour time debt isn’t going to get any smaller unless action is taken fast. You don’t have time to conduct lengthy pilots that never go anywhere. Health organizations that win will be the ones that put a foundation under a single real problem, prove the return, and then do it repeatedly.

The charge nurse at 3 a.m. does not need another proof of concept. She needs the roster to build itself before her shift starts.

NEXT STEPS:

Speak to an SHI expert to learn more about our Agentic Camp. Let’s make your AI possibility a reality. Fast!

Learn more about our AI solutions here