Healthcare support is one of the harder problems in AI voice. Patients calling about appointments, prescriptions, and test results are often anxious. They want to speak in their first language. They need to trust that the system understood them correctly.
And the margin for error — a wrong appointment time, a misheard prescription detail — is genuinely narrower than in most other industries.
Lumen Labs runs a network of diagnostic labs and outpatient clinics across Maharashtra and Gujarat: 18 facilities, 400,000 patient contacts annually, and a support team of 40 agents stretched across a 16-hour call window.
When they deployed AIVA Voice in November, they were handling 12,000 calls a month at an average wait time of 6.4 minutes.
Illustrative example. Lumen Labs is a composite scenario built to show how AIVA works for this kind of business — not a verified customer account.
The problem with IVR
Lumen's original system was a DTMF tree — press 1 for appointments, press 2 for test results, press 3 for billing. It was universally despised. Patients calling about results navigated four levels of menu before reaching a human.
Patients who spoke Marathi or Gujarati and weren't confident with the English-language options often hung up and called back, hoping for a different agent.
The IVR was also creating work rather than saving it. Agents spent the first 90 seconds of every call reconstructing what the patient had already done in the menu system and what they actually needed. Information the IVR had gathered wasn't usable, because it was never surfaced to the agent in any structured form.
Lumen's head of operations, Vikram Shah, put it plainly: "Our IVR was a waiting room, not a support system. It held patients in a queue but didn't help anyone."
This is the structural failure of menu trees generally, and it's worth understanding before replacing one — AI receptionist vs IVR covers why the two are different categories of thing rather than two versions of the same thing, and our multilingual IVR guide covers the specific ways menu trees fail multilingual callers.
The integration
Deploying AIVA at Lumen meant connecting three systems: their appointment scheduling platform, their lab results database (with strict access controls and PHI handling), and their billing system for payment queries.
The PHI access was the most careful part. We worked with Lumen's data security team to implement field-level access controls: AIVA can confirm whether results are ready and direct a patient to the right portal to view them, but it never reads or speaks the results themselves.
Anything requiring actual discussion of a result routes to a medical professional. That boundary is hard-coded and not configurable.
This is a deliberate design position rather than a technical limitation. Indian health data handling falls under the Digital Personal Data Protection framework, and the safest architecture for sensitive fields is one where the agent's capability is bounded rather than its instructions.
A prompt telling a model not to read results is a request. An access control that never returns the field is a guarantee.
We've written about that distinction more generally in how AI phone agents handle data privacy, and the platform-level controls are documented on the security page.
Setup took eight days — longer than a typical AIVA deployment, mostly due to access-control configuration and the testing required to verify the PHI boundaries held in every scenario we could construct.
If you're deploying AIVA in healthcare, medical, or financial contexts: we require a security review before go-live. The PHI boundary configuration is standard for any deployment involving health or financial data, not an enterprise add-on.
Where escalation lines sit in healthcare
The escalation configuration at Lumen is stricter than a typical commercial deployment, and deliberately so.
Beyond our three standard triggers — emotional distress, compliance and account security, and any explicit request for a person — Lumen added clinical categories: anything touching symptoms, medication, result interpretation, or urgency assessment routes to a human immediately, regardless of how confidently the agent could answer.
That's a wider escalation net than we'd normally recommend, and it costs them some efficiency. It's the right trade for the sector.
The general framework for drawing this line is in our escalation rules post; the healthcare version of it starts from the same place and then adds categories where being right isn't sufficient, because the decision itself belongs to a clinician.
The results
After 90 days of operation:
| Metric | Before | After |
|---|---|---|
| Average wait time | 6.4 minutes | 52 seconds |
| First-contact resolution | 38% | 89% |
| Calls requiring human escalation | 62% | 11% |
| Patient satisfaction (post-call) | 3.1 / 5 | 4.4 / 5 |
Agent utilisation shifted rather than shrank: the same 40 people moved from tier-one volume to complex patient queries and callbacks.
The most striking finding came from Lumen's own patient survey data: 71% of patients who interacted with AIVA preferred it to their previous IVR experience. Patients didn't care whether they were talking to an AI.
They cared whether they got a fast, accurate answer in a language they were comfortable in — which, across the twelve languages we support, they now did.
The two weeks that decided it
Lumen's outcome looks like a deployment story. It was mostly a calibration story, and the calibration was unusually specific to healthcare.
The first week of live traffic surfaced a category we hadn't planned for: callers who ask an administrative question that is actually a clinical one. "Is my report ready?" is administrative.
"Is my report ready, because the doctor said to call if the numbers were high" is not — the caller is asking for reassurance, and answering the literal question well is the wrong outcome.
We spent most of week one building triggers for that shape: any administrative query carrying clinical context routes to a person, even though the agent could answer the surface question accurately. It costs Lumen efficiency on a real slice of calls.
It's the correct trade, and it's not a rule we'd have written from a requirements document — it came out of reading transcripts.
Week two was the opposite problem. The initial configuration escalated far too eagerly on anything containing a medical noun, which meant callers asking where the Andheri lab is were being transferred because they'd mentioned a blood test.
Loosening that took daily log review, and it's the reason we now insist on a two-week window rather than a go-live date.
The general shape holds across sectors: week one you find what you should have escalated, week two you find what you shouldn't. Teams that budget for one week get only the first half of that lesson.
What transfers to a smaller clinic
Lumen is an 18-facility network, but almost nothing about the outcome depended on that scale.
The three things that made it work — connecting to the live scheduling system rather than a static FAQ, bounding what the agent can access rather than what it's told, and setting escalation by who the patient needs rather than what they asked — apply identically to a single-location practice with 40 calls a day.
If anything they matter more, because a small clinic has no queue to hide behind when the phone goes unanswered.
We've written up the smaller-practice version on our clinics page, and the lab-specific considerations in AI phone agents for diagnostic labs.
Vikram Shah's summary is the line we've quoted most often since: "We thought the barrier to AI in healthcare was trust. It turns out the barrier was IVR. Patients will happily talk to an AI that listens. They won't tolerate a menu tree that doesn't."
If you run a clinic or lab and want to test that on your own line, start free with ₹500 of credit — no card required.