Implementing Enterprise AI in Healthcare: Some Thoughts

AI in healthcare is happening, but slower than it should. COVID offered a useful precedent: when the threat was mortal and immediate, health systems stood up virtual visits and testing stations and doubled ICU capacity in weeks. The sector can move quickly when the stakes are unambiguous.

A comparable threat is building: up to fifteen million more uninsured, material cuts to Medicare and Medicaid hospital payments, and state-level price caps blocking cost-shifting to commercial payers. Unlike with COVID, the erosion is gradual—spread over years rather than weeks—which makes it harder to mobilize against. Closing that gap likely will require 15 to 20 percent reductions in delivered cost of care.

Agentic AI is the first technology that can plausibly deliver that scale of efficiency—not by replacing people but by making each person materially more productive, particularly in administrative workflows. Agentic AI is qualitatively different from prior technology waves: it learns, adapts, and becomes more capable over time. But the same structural barriers stall adoption: governance friction, professional-managerial mistrust, and insufficient C-suite urgency.

The pattern is familiar: diffuse committees, a flood of point-solution vendors, pilots launched without strategic prioritization, budget constraints, and workforce disagreement about whether fundamental change is even necessary. These same forces stalled digital health adoption pre-COVID. The current wave of uncoordinated AI pilots invites a backlash—led by nursing leadership and cautious CFOs—that leaves the field disillusioned and far short of the results needed to justify scaled investment.

There is an alternative, and the current approach is unlikely to succeed without it.

Here are the ingredients of a successful enterprise AI deployment that delivers results on the timeline this moment demands:

1. Platform first. AI reduces uncertainty by applying learning algorithms to large, well-structured datasets. Rather than assembling a strategy from dozens of pilots, health enterprises should build an ontology: a unified, accessible data layer that describes the enterprise and serves as the foundation for every AI deployment. A fragmented vendor landscape—much of it venture-funded and transient—will not produce coherent infrastructure or the operating efficiency.

2. Data ownership. Your data—patient records, financial and contract data, staffing information—is both legally protected (PHI) and commercially proprietary. Letting model providers (OpenAI, Anthropic, Google) absorb it to train their models, then sell it back to you, is not sustainable. You own this data. It must be secure, sovereign, and protected by encryption and robust cybersecurity infrastructure.

3. Model-agnostic. It is not clear whether the general-purpose LLMs driving the AI surge since 2022 will scale economically. The risk: compute costs rise as model owners chase revenue, leaving enterprises locked into toolsets whose costs they cannot control and that erode any operating gains. The answer: do not marry a model. Build a platform that routes across models, allows for open-source alternatives, and captures the price competition that inevitably follows. Most vendor-led AI strategies have married a model—and inherited its cost structure.

4. From platform to operating system. The objective is not deploying technology—it is redesigning work itself, incorporating machine intelligence and care orchestration. To build an effective operating system, leadership must set priorities: the problems worth solving first to improve the workforce experience and that patients notice: safety, clinical outcomes, access. The CEO must lead this with full support of the direct reports. It cannot be delegated to a committee or a portfolio of pilots.

We know this approach works because we did it ourselves. BRG built and deployed its own AI operating system in under three months on our own ontology—the core databases that run our advisory business. Over 1,000 of our experts use it daily. It learns what we need, automates prompt workflows, and surfaces the inquiries that save time and produce results fastest.

Now we’re deploying it for our clients. Agentic AI is not another tool. It is a partner in the work and the foundation of their future man + machine operating systems.