This is the first in what'll be an occasional series of field notes — short write-ups of what we've learned working alongside specific behavioral health practices. Names redacted, lessons not.
The practice in this one is a 22-clinician intensive outpatient program (IOP) in central Texas, mixed-population (mood disorders + substance use), in operation for nine years. They serve roughly 180 patients a week across two locations.
The situation when we walked in
Like most IOPs, their growth bottleneck was not clinical capacity — it was the front-desk pipeline that fed clinical capacity.
A typical week looked like this:
- ~80 new inquiries (phone, web form, payer referrals, partner referrals)
- ~62 of those got a callback within 24 hours
- ~38 of those eventually scheduled an intake call
- ~22 of those completed an intake and started programming
The drop-off they were most frustrated by wasn't the back half (intake → started). It was the very front: inquiry → callback. They knew, in the abstract, that response time correlates with conversion. They didn't know how steep the curve was.
Their measured median response time, when we sampled the prior 30 days of inquiries, was 6 hours and 12 minutes. Their answer rate during business hours was ~64%. After hours and weekends, it was effectively zero.
What we deployed
We rolled out Supadesk in a narrow scope: it answered every inbound call and every web-form submission, 24/7. It did not yet handle benefits verification or scheduling — only intake conversation, qualification, and warm handoff to a human callback queue.
The rollout was three days of configuration (their referral criteria, their three primary payer mixes, the language and tone they wanted), one day of staff training on the new handoff workflow, and a one-week pilot before going live across both locations.
We chose to scope it narrowly on purpose. The team was skeptical about AI talking to prospective patients — reasonably so — and we wanted them to be able to listen to every interaction during the pilot before broadening scope.
What changed in the first 30 days
Three numbers:
| Metric | Before | After (day 1–30) |
|---|---|---|
| Median time to first response | 6h 12m | 0m 89s |
| Inquiry-to-callback rate | 78% | 99% |
| Inquiry-to-intake-call rate | 47% | 63% |
The first two were the ones we predicted. The third was the surprise — a 34% relative lift in inquiry-to-intake conversion, driven almost entirely by capturing inquiries that previously went stale before anyone could get to them. Specifically, weekend inquiries (which had been getting roughly a 12% conversion rate) jumped to ~58%, in line with the weekday rate.
The clinical director's reaction, paraphrased:
"We thought we had a referral-volume problem. Turns out we had a response-time problem disguised as a referral-volume problem."
What we didn't change
A few things we deliberately did not automate in the pilot:
- Clinical screening decisions. Supadesk identifies whether a caller fits the program's criteria, but the actual fit decision and treatment recommendation stays with the human intake team.
- Sensitive disclosures. When a caller discloses suicidal ideation or active substance use crisis, the system warm-transfers to the on-call clinical team immediately.
- The first scheduled intake call. That's still done by a person, even though the agent could technically book it. We wanted prospective patients to talk to a human before committing to the program.
This scope is a starting point, not a ceiling. The team is now considering bringing benefits verification and scheduling into Supadesk's scope in the next quarter.
Why the gains were bigger than we predicted
A few hypotheses, in rough order of how much we believe each one:
-
Time-to-response curves are steeper than people think. The data on lead-conversion-by-speed in B2B sales is well-known. The equivalent data in behavioral health intake is scarcer but, based on what we're seeing across customers, looks even steeper. A 90-second response converts roughly 5x better than a 6-hour response in this segment.
-
Off-hours inquiries are not the same population as business-hours inquiries. The people calling at 11pm on a Sunday are often in a moment of acute decision — by Monday morning that moment may have passed. We were essentially talking to a different (and more motivated) population that the practice had been missing entirely.
-
The conversation quality was higher than expected, not lower. We expected the AI vs. human comparison to favor humans on warmth and to favor the AI on completeness. The completeness was unambiguous. The warmth was — based on the practice's own listening sessions — roughly a wash, with the AI being noticeably more patient and slower-paced than rushed front-desk staff during a busy hour.
What we're going to test next at this practice
Now that the baseline is established, we want to test:
- Bringing benefits verification into the same first conversation. Currently the agent qualifies, hands off, and a human verifies benefits in a follow-up. Consolidating that into one interaction should compress the funnel further.
- Active outreach to lapsed inquiries. The practice has a backlog of ~400 inquiries from the prior 12 months that went cold. The agent can re-engage them on a rolling basis with a tailored message — separate experiment, low risk, potentially high upside.
Lessons for practices considering this
Three things we'd tell a similar IOP considering this kind of rollout:
- Start narrow. We saw a meaningful conversion lift before we touched scheduling, benefits, or any of the downstream work. The first 90 seconds is where the leverage is, and isolating it makes the pilot easier to evaluate.
- Listen to every interaction for the first two weeks. It's the fastest way to build internal confidence and to surface things the agent should be saying differently for your specific population.
- Measure conversion, not satisfaction. Staff and patient satisfaction matter, but they lag conversion. Conversion is the metric that tells you whether the system is working in week one.
If you're running a behavioral health practice and curious whether this would work in your context, we're happy to walk through your specific numbers and give you a no-nonsense read.