The first time you hear a good voice agent, it is easy to be impressed. The voice is warm. The answers are quick. The appointment gets booked.
But a demo is the easiest call the agent will ever handle.
Real patients speak over the agent. They mix languages. They forget a doctor’s full name. A family member may answer the phone. The calendar may stop responding. Someone may casually mention chest pain while asking about tomorrow’s appointment.
We have learned that the quality of the voice matters, but what happens around the voice matters much more. A healthcare agent must know what it can say, what it can do, when it must stop and when a person needs to step in.
If you are evaluating one, you do not need to become an AI expert. You simply need to ask better questions. These are the ten we would start with.
Follow the patient’s words.
Start with a very ordinary question: after a patient speaks, where do those words go?
The recording may move through a phone provider, a speech service, an AI model, a hospital system and an analytics tool. Each step may create its own copy. So “the data is encrypted” is not enough. You should be able to follow the complete journey in plain English.
“Show us every place a patient’s recording, transcript and details travel, how long each copy stays there, who can see it, and what happens when we ask you to delete it.”
Ask what applies in your market.
A global vendor should not give every healthcare organisation the same compliance slide.
In the United States, start by asking whether the vendor will handle protected health information, whether it will sign the right Business Associate Agreement, and which other companies will touch that information. In India, ask how the service maps to the IT Act and SPDI Rules where applicable, the phased DPDP framework, CERT-In directions and relevant healthcare guidance.
The point is not to turn every buyer into a lawyer. It is to see whether the vendor understands your market, can explain its responsibilities simply, and is willing to put them in writing.
“For this exact use case and country, which privacy, security and healthcare rules apply to us, to you and to your partners, and what have you built to meet them?”
Then have your own legal team check the answer.
A phone number is not a patient.
Phones are shared by families everywhere. A spouse may answer. A parent may call for a child. An old number may belong to somebody new.
Before sharing anything sensitive, the agent must know who it is speaking with. It should also introduce itself as AI and explain recording or data use in language a patient can understand.
Answer as a family member. Say it is the wrong number. Call on behalf of a patient. See what the agent asks, and more importantly, what it refuses to reveal.
“Supports your language” tells you very little.
In India, a patient may begin in English, move into Hindi and use a local name for a procedure, all in one sentence. In the US, the same challenge may appear across English, Spanish and dozens of accents. In either market, the agent must recognise unfamiliar doctor names, pronounce medicines correctly and hear the difference between fifteen and fifty on a noisy line.
Test with the people, places and words from your own organisation. Generic language scores will not tell you whether the agent works for your patients.
Give the vendor a list of your doctors, departments, medicines and locations. Test it in every language they promise. Include code-switching, dates, numbers and similar-sounding names.
A confident sentence is not a confirmed appointment.
This is one of the most important distinctions in voice AI. The agent can say “you are booked” even when the hospital calendar never accepted the booking.
Good systems wait for a real confirmation. If the calendar is unavailable, the agent says so honestly and offers the next sensible step.
Ask the vendor to make the calendar fail. Then repeat the booking request. Does the agent admit that it could not confirm the slot? Does the retry create a duplicate? This one test reveals a great deal.
Teach the agent where healthcare becomes clinical.
A scheduling call can quickly turn into: “Should I stop this medicine?” or “Does this result mean I have cancer?” A warm, fluent agent can sound authoritative even when it should not answer.
The safe boundary must be written with clinicians. The agent needs approved information, clear forbidden areas and a reliable way to bring in a qualified person.
Ask for a diagnosis. Ask the agent to interpret a report, change a dose or guarantee an outcome. Then ask the vendor what prevents an unsafe answer beyond a line in the prompt.
A good handover is a successful outcome.
Some conversations should not stay with AI. The agent may be unsure, the patient may be distressed, or the request may simply need human judgement.
The important question is not whether the agent can transfer a call. It is whether the right person answers with the context they need. “Transfer attempted” and “patient connected” are very different outcomes.
Ask for a person during working hours and after hours. Mention an urgent symptom. Make the staff member unavailable. Watch where the call goes, what the patient hears and what information reaches the human team.
Ask to see the calls that went wrong.
Every voice agent fails. The useful question is whether the team can find those failures, understand them and prevent them from returning.
A highlight reel tells you what is possible. A failure review tells you whether the system is ready.
“Show us your serious failures by language and call type. What happened, how did you fix it, and how do you know the fix worked?”
Include noisy calls, interruptions, older callers and attempts to reach another patient’s information.
Do not confuse a cheap call with a useful outcome.
Per-minute pricing sounds simple, but it can hide retries, failed bookings, transfers and the time staff spend correcting mistakes.
Measure what the organisation actually wanted: a correctly booked appointment, a completed follow-up, or a safe handover. Then count the full cost of getting there.
“For every hundred patient calls, how many end in a correct, checked outcome, and what is our total cost for each one?”
Discuss the ending before you begin.
A healthcare organisation should never feel trapped inside a voice AI platform. If you change vendors, or the company disappears, you should know what happens to your phone numbers, recordings, transcripts, settings and patient records.
“Show us a sample export. What belongs to us, what can move, what cannot move, what will migration cost, and how will you prove that remaining copies were deleted?”
Put the answer in the contract.
The ten questions, on one page.
Bring this to the vendor call. A good partner will have a written answer for every row.
| Question | What good looks like | Warning sign |
|---|---|---|
| 01Data journey | A plain-English map of every copy, its retention and its owner, including sub-processors. | Only the vendor’s own database is described. |
| 02Market rules | Obligations for you, the vendor and its partners, in writing, for your country and use case. | One compliance slide for every market. |
| 03Identity | Verifies the caller before sharing anything, discloses that it is AI, explains recording. | Reads details to whoever answers the phone. |
| 04Language | Tested on your doctors, drugs and departments, per language and accent. | One combined accuracy number. |
| 05Confirmation | Waits for your system to confirm before telling the patient; admits failure honestly. | Says “booked” before the calendar responds. |
| 06Clinical limits | Boundaries written with clinicians and enforced outside the prompt. | “We told it not to diagnose.” |
| 07Handover | Right person, with context, measured as “patient connected”. | Success measured as “transfer attempted”. |
| 08Failures | A failure review by language and call type, with fixes and evidence. | Only a highlight reel. |
| 09Cost | Priced per correct, checked outcome. | Priced per minute with no outcome data. |
| 10Exit | A sample export, ownership terms and a deletion proof in the contract. | “We have never had anyone leave.” |
Start with one job. Agree on what “good” means.
Pick one administrative workflow, such as booking or rescheduling. Decide the success rules and stop conditions before the first patient call.
- Real calls in every language and accent you plan to support
- Every completed action confirmed by your own system, not by the agent’s transcript
- Every serious safety, privacy or handover failure reviewed
- One named owner from your team and one from the vendor
- A rollback plan and a clear decision on whether to expand
The right partner will enjoy this conversation.
They will not rush you back to the polished demo. They will show you the awkward calls, the boundaries, the handovers and the proof behind each completed action.
A beautiful voice creates a good first impression. Trust comes from everything the system does after the patient speaks. That is the standard we believe healthcare voice AI should meet, and the standard we work toward while building Krisper and the safety layer around it at Iksha Labs.
Source notes
This is a practical evaluation guide, not legal advice. The rules depend on the organisation, workflow, location and data involved. Have qualified counsel confirm the current position for your use case. Sources were reviewed on 15 September 2026.
- United States: HHS guidance on HIPAA Business Associates and Business Associate Agreements, including AI services that handle patient information on behalf of a covered healthcare provider.
- India: IT Act SPDI Rules, 2011; DPDP Act commencement notification and DPDP Rules publication; CERT-In Directions; Telemedicine Practice Guidelines, 2020; and the TRAI commercial communications framework.
Iksha Labs
We build the safe agentic layer for healthcare. Krisper, our voice agent, turns patient conversations into completed actions, or a clear, contextual handover.