AdventHealth deploys ChatGPT across nine states by treating adoption as the product

The frontend is a blank text box, and that was the hard part

The product AdventHealth deployed is ChatGPT — first ChatGPT Enterprise, then ChatGPT for Healthcare. From a frontend perspective, that means the interface handed to physician advisors, finance staff and HR teams was a chat window. There is no bespoke UI, no purpose-built form for utilization management. The surface is deliberately general.

That generality is exactly what stalls adoption. Chief AI Officer Rob Purinton describes a workforce split between eager early users and a large group sitting out — not because the tool was blocked, but because they didn't know what to type. A blank prompt offers no affordances. It tells a clinician nothing about what a good input looks like or which of their daily tasks it can shorten.

We had folks who were eager to start, but there were a very large number of people who were on the sidelines. They weren't sure how to use AI effectively in their daily jobs.Montana Labs

AdventHealth's response was to stop treating the interface as self-explanatory. Purinton says leadership decided to "treat adoption as the product" — an admission that the shipped software and the used software are different things when the frontend is an open-ended chat.

Solving the empty-prompt problem with peer groups, not training programs

Rather than run centralized training on a general-purpose tool, AdventHealth organized domain-based peer groups: finance teams sharing prompts with finance teams, HR with HR. This is a frontend design decision made socially instead of in code. The prompts, workflows and examples that a dedicated UI would normally encode as buttons and templates were instead distributed as shared practice within each function.

The organization also instrumented usage directly. It tracks messages per user per business day — excluding weekends and holidays to hold the baseline steady — and manages that number like any other KPI, with targets and trends reviewed regularly. For a chat interface with no fixed workflow, message volume becomes the proxy for whether the tool has actually entered daily work.

Framing mattered too. Leadership refused to present the interface as automation. Purinton says they talk about "time back" instead — compressing a ten-minute review while preserving quality, and returning that capacity to clinicians. The language shapes how people approach the empty prompt: as an assistant to hand a task to, not a system that replaces them.

Utilization management: the interface drafts, the clinician decides

The clearest use case was utilization management, where physician advisors spent roughly ten minutes per case reading charts, identifying relevant details, checking criteria and drafting structured rationales. In the ChatGPT workflow, the model generates structured chart summaries, surfaces relevant clinical details and produces an initial rationale draft. The clinician keeps final judgment; the assembly time shrinks.

Across departments the same pattern recurs — a first-pass draft instead of a blank page, policies converted into structured formats, unstructured notes summarized into action steps. AdventHealth reports an 80% reduction in time spent on administrative tasks, and a physician who had done evening "pajama time" documentation was able to finish within regular hours.

Crucially, AdventHealth refused to measure this with self-reported estimates. It reads impact from EHR timestamps baked into the process, so it can see exactly how many minutes changed and whether the shift is statistically significant. When the frontend is a generic chat box, this kind of external instrumentation is the only honest way to know if it is doing anything.

What a generic interface demands of the organization behind it

The specific implication of AdventHealth's rollout is that shipping a general-purpose chat interface pushes the entire burden of product design onto the deploying organization. There is no workflow encoded in the tool; the workflow lives in peer-shared prompts, a usage metric, and EHR timestamps that verify the effect.

Purinton's closing framing — that adoption is "change leadership," not "go use the product" — is the honest version of this. When your frontend is a text box that can do almost anything, the work of deciding what it should do, and confirming that people do it, falls entirely on you. AdventHealth's 80% figure is less a statement about the model than about how much scaffolding a company must build around an open-ended interface before it produces measurable results.