By Soham Moore, CEO of Rely Health

Most health systems we talk with are just beginning to evaluate voice AI agents. A January 2026 study by Microsoft and The Health Management Academy points to a similar pattern across agentic AI: 43% of health systems were piloting it, while only 3% had it live in a workflow.

What separates a successful pilot from one that stalls has less to do with how natural the agent sounds and more to do with what it is designed to prove and whether it gives your clinical team more moments of human connection with patients.

Here are five things I would recommend to any healthcare team evaluating its next voice AI pilot.

1 Be clear on the problem you need to solve

Discharge is a good example. A post-discharge agent can reach more patients, but will that help prevent readmissions? Or will it surface more concerns for your team to review? If those concerns go unresolved, the additional engagement may do little to change the outcome.

Before you choose a partner or even define the scope of the pilot, name the problem you need solved and ask how the program will get from the first call to that result. The pilot needs to account for all the work required after a patient engages.

2 Prepare for the new queue of work that will surface

Delivering care to patients is never a straight line. An agent will reach patients your team never had time to reach, and those conversations will surface new work: a ride that doesn’t accommodate a wheelchair, a coverage problem, an unscheduled referral, someone who just needs to talk.

Finding those needs is valuable, but someone has to resolve them. If that queue lands back on an already stretched team, the pilot has added work instead of removing it.

Before go-live, decide who owns that work and how quickly it needs to be resolved. Clinical concerns belong with clinicians, but take a hard look at which tasks actually require their expertise. Some have landed on the clinical team simply because that is how the process has always worked. A pilot is an opportunity to change that. In my view, whoever runs the program should own the non-clinical work through resolution, giving clinicians more time for the care that requires their license and judgment.

3 Judge success by the patient outcome

A completed call, a confirmed appointment and a booked ride are steps toward an outcome. Our partners want to know whether a heart failure patient actually saw a specialist within 7 to 14 days of discharge. The number of agents or people involved matters less than whether the patient received that care.

Let the outcome you want to achieve drive innovation and your evaluation criteria. Choose the measure first, whether that is readmissions, time to specialty care or completed referrals. Agree on the baseline before launch, then put it on a dashboard both sides can see.

In one regional health system’s discharge program, patients we engaged had a readmission rate of 14.2%, compared with 22.6% among patients routed to the program whom we could not reach. The comparison covered 1,657 EHR-matched patients in Q1 2026. 

4 Treat day one as the starting line

Every community runs on its own rules: which plan needs rides booked days ahead, which clinic holds same-day slots, who actually answers at the skilled nursing facility. Often, nobody has written those rules down.

That knowledge is learned by people doing the work. At Rely, we begin each engagement with an accelerated learning loop led directly by our expert human navigators. Their insights guide how our AI agents engage. AI engages first, and our humans own the exceptions. As navigators resolve unfamiliar barriers, that learning informs what AI can take on next.

Each interaction also helps us understand where the broader patient journey breaks down and how to improve it. We feed those insights back into the program, continually refining how we engage patients, resolve barriers and support clinical teams. That is how our programs evolve so that day 100 looks nothing like day one.

5 Start narrow and give it time to produce useful data

The pilots that expand usually start with a defined population, often the highest-risk group, and run long enough to show results. In our experience, leaders want three to six months of data before deciding to expand.

Keep the IT lift low so that clock starts sooner. Our navigators, for example, onboard into the existing EHR the way a PRN hire would, so there is no integration project before the first patient call.

My recommendation is to keep the barriers to starting low and share what you learn early. You need time to measure outcomes, but you can begin learning about patient needs and workflow barriers from the first call. Give other departments access to those insights as they emerge. A solution developed for one population may help another team rethink its approach. That cross-pollination is how a focused pilot can contribute to care transformation across the health system.

The question underneath all five

When the pilot ends, does your team have more time for patients, or less?

That is the standard I would use. Automation should create more room for human connection in care. A pilot earns the case for expansion when it shows how it can help your clinical team spend more time with the people who need them.

If you’re scoping a pilot now, or revisiting one that fell short, I’m glad to compare notes. Grab time here