Briefings for the people who sign off on healthcare AI

Choosing: nothing is signed yet

The job is separating a vendor from its demo. Fifteen security and compliance questions belong in any RFP for a tool that will touch protected health information, and a shorter checklist for a scribe vendor specifically narrows that list to what should already be documented before the call. Running fewer than ten providers changes the answer: buying at that size means the self-serve end of the market and a monthly term, not an enterprise sales process built for a health system. the integration questions that decide the timeline cover which API a vendor actually uses to read and write, which is the detail that turns a two week go-live into a two month one. And before any of that, the build-versus-buy math names the narrow set of organisations for whom building is genuinely the right call.

Deploying: the contract is signed

Define the metrics before the pilot starts, not after: a baseline held for two weeks, one outcome metric that would justify the spend, and a stopping rule written down while everyone is still being reasonable. Knowing whether the thing you bought is an agent, a copilot or plain automation matters at the same stage, because that distinction is what decides the controls the rollout needs, not the vendor’s marketing name for it. And once it is live, adoption is won or lost in the eight weeks after go-live, with a champion model and a note quality feedback loop doing more than a mandate ever does.

Governing: more than one tool is live

A governance committee earns its place by returning decisions on a clock: a written charter, a named decision owner, and risk tiers that give a low-stakes tool a lighter review than a high-stakes one. A smaller practice rarely needs a committee at all; a two-page acceptable use policy covers approved tools, what staff may never paste into an unapproved one, and who reviews AI-drafted output before it is used. On the regulatory side, a year-by-year account of what actually moved tracks ONC, CMS and FDA in one place, a comparison of the state AI laws with healthcare duties covers a group operating across state lines, and the CMS-0057-F provider checklist turns a rule that binds payers, not providers, into the handful of things a provider organisation should actually check this quarter.

Fixing something that already went wrong

Sometimes the question is not what to buy or how to run it, but why a live tool is not working, or why a number being quoted in a meeting does not hold up. Pilots rarely stall because the model was bad: integration work that got deferred, a baseline nobody captured, and no governance route from pilot to production do most of the damage. The evidence behind the documentation burden numbers is narrower than the marketing built on top of it, and this piece says exactly where the two studies carrying most of that weight stop supporting the claim.