Look, right now almost every company selling into healthcare says the same three words: "we use AI." That tells you nothing. It is not a result. It does not move a single patient closer to treatment, and it does not take a single hour off your team's week.
The only thing that matters is what happens to the work. Did the administrative process get faster? Did it cost less? Did more patients get what they needed, sooner? Those are the metrics. Everything else is theater.
I want to be very clear and simple about how we think about this at Synthpop, because it is how we build and how we sell.
"We use AI" is not a metric
We build agents that automate administrative work in healthcare: intake, eligibility, prior authorization, resupply outreach, referral processing. The value is not that there is a model behind it. The value is that work which used to take a person twenty minutes to three hours now happens on its own, correctly, and hands your team only the cases that actually need a human.
So when someone asks whether our AI is "good," I don't answer with model names. I answer with the workflow: what it did, how many times, and what changed downstream. If we can't point to that, we haven't earned anything yet.
Time to treatment is the number that matters most
Start with payor calls, because that is where the pain is most obvious. A manual verification or authorization call runs anywhere from twenty minutes to three hours. Multiply that across a book of patients and you have a team that spends its day on hold instead of on care.
At one national sleep therapy and home sleep testing provider, our voice agents automated more than 55,000 payor calls over two years, with no added staff time. Time to therapy for PAP patients dropped by roughly half.
That last number is the one I care about most, because it is not an efficiency stat. It is a patient starting treatment sooner. That is the reframe: administrative automation is not a back-office cost play, it is a time-to-treatment play. When you take the phone work off a person, the patient moves through the journey faster. The business result and the patient result point in the same direction.
"The AI caller handles first-round communication with payors for authorization requirements on home sleep studies and PAP therapy, which frees our team for the calls that truly need a human touch."
- Insurance Benefits Manager, national sleep-care provider
Same team, more work, less cost
The second thing to measure is what happens to your team's capacity. The goal is simple: reduce hours, reduce cost, increase revenue, without asking anyone to work nights.
A few real numbers from deployed workflows:
- On resupply outreach, one provider cut staff hours on the workflow by about 77% and lifted order rate by 9.5%, reaching patients outside of nine to five without adding headcount.
- On referral processing, a diagnostic lab operator serving more than 100 health systems took referral handling from around twenty minutes to under a minute, roughly a 96% reduction, and moved cost per referral from the ten-to-twenty-dollar range down to two to four dollars.
- Across intake and onboarding, the same class of agents onboarded more than 100,000 patients, with intake completion above 90%.
Look at those together: fewer hours, lower cost per task, more orders captured, more patients onboarded. That is what "AI is working" actually looks like, on a P&L and in a clinic.
Be honest about the numbers
Here is the part most vendors skip. If the data is not clean enough to defend, do not over-report it. I would rather show you a smaller number I can stand behind than a big one I have to walk back next quarter.
The way we hold ourselves to that is simple: pick a source of truth, measure the same metric the same way, and report it consistently from that point on. When we say 55,000 calls or 96% faster, those are counted, not rounded up for a headline. If a number is still soft, we say so, and we tell you how we are going to firm it up.
Trust in healthcare AI is not won with a demo. It is won with numbers you can audit. That is a higher bar, and it should be.
How to tell if your AI is actually working
If you are evaluating this for your own operation, here is the short list I would measure. Not model benchmarks. Work.
- Time to treatment: how much sooner does the patient get care?
- Staff hours per workflow: how many hours came off the task, and where did that time go?
- Cost per task: what does one intake, one referral, one verification actually cost now?
- Throughput: how many more orders, referrals, or patients moved through with the same team?
Start with one workflow where you can measure all four today. Intake and payor calls are usually the fastest to show ROI, because the pain is concrete and the before-and-after is easy to count. Prove it there, then expand across the rest of the patient journey. That is the whole playbook: show ROI now on one workflow, earn the right to automate the next one.
If you want to see what these four numbers look like on your data, that is the conversation worth having. Bring one workflow. We will show you the before and after, honestly, and you can decide from there.
