Every board now has an AI question, and every vendor now has an AI answer. I sit on both sides of that exchange. I help client teams choose and roll out AI, including voice agents that handle live patient calls. I advise on what to buy, what to ignore, and what their own staff should run. The gap between the question the board is asking and the answer the vendor is selling is where most organizations lose the next eighteen months.
The way out is not a platform decision. It is a discipline: tie every AI initiative to a specific, measurable operating outcome before you spend a dollar. If you cannot name the metric that moves and by how much, you do not have a project. You have a demo with a budget attached.
Start where the work is repetitive and lossy
I open every AI engagement with the same question: where is the work high-volume, repetitive, and currently leaking value? The honest answers are unglamorous. After-hours inquiry capture. Intake qualification. Insurance and eligibility triage. Documentation drafting. Content production at scale. These are places where a well-scoped model produces revenue or returns hours you can actually count on a timesheet.
I avoid the inverse: the broad "AI transformation" mandate with no number attached. Those initiatives fail slowly and expensively because nobody can say when they are done or whether they worked. A practical AI roadmap scores each candidate use case on two axes (revenue or hours at stake, and regulatory risk) and ships the top-right quadrant first. Everything else waits until the first win has paid for itself.
If you cannot name the metric that moves and by how much, you do not have a project. You have a demo with a budget attached.
How to tell real from vaporware
When a vendor demos, I run four questions before I get interested. What specific task does this replace or accelerate, in words a frontline manager would recognize? What does it cost fully loaded: licensing, integration, and the human review it still requires? How does it handle the case it gets wrong, and who catches that? And what happens to our data: where does it live, who can see it, and is there a signed agreement that says so?
A tool that survives those four questions is worth piloting. A tool that answers them with adjectives is a slide deck. I have watched organizations sign six-figure contracts on the strength of a demo that answered none of them, then spend a year discovering that the model was accurate in the showroom and unusable against their actual call volume and edge cases.
The demo is engineered to succeed. It runs on curated inputs, a clean happy path, and a presenter who steers around the model's weak spots. Your environment is the opposite: messy data, callers who interrupt, insurance questions the model was never trained on, and the 3am edge case nobody scripted. The only honest test is a scoped pilot against your real volume, measured against the number you named up front. If a vendor resists that pilot, they already know how it ends.
Governance is a design input, not a cleanup project
In a regulated environment, governance does not get added at the end. Data handling, vendor agreements, audit trails, and human-in-the-loop review belong in the architecture from the first sketch. This is not a tax on speed. The organizations that move fastest with AI are the ones that treated compliance as a constraint up front, because they never have to stop and rebuild after an incident. I advise teams to design every system on the assumption that someone will audit it, because eventually someone will. I go deeper on what that actually requires in compliance-aware AI.
Concretely, that means an AI voice system that captures a crisis call routes protected information only where a signed agreement allows it, logs every interaction for review, and hands off to a human the moment the conversation exceeds what it was scoped to handle. A behavioral health group deployed exactly that pattern to answer inquiries no human was available to take at 2am, with the guardrails built in from day one, not bolted on after.
Real AI in healthcare looks boring
The AI that works looks unremarkable from the outside: a few high-return use cases, measured carefully, governed tightly, expanded only after the previous one has proven out. It does not trend on a conference stage. It quietly answers calls, drafts documents, and qualifies leads while the metric it was scoped against moves in the right direction. That is the strategy I recommend, and the one I have seen work in production. Everything louder than that is usually a demo.
