By Hitesh Ganjoo, Founder
To audit a healthcare AI agent, you need to pull up any single call and see what the agent heard, what it understood, which of your rules it applied, what it did in your systems and when a person took over. To monitor it, you review a planned sample of calls every week against your policies, and check that the outcomes the agent reports match what actually happened.
The tempting answer is "we record every call".
A recording alone isn't enough.
Picture a Monday quality meeting at a hospital. A complaint has come in. A son says the "AI receptionist" told his mother her knee procedure would be painless and she'd be walking the same evening. The quality lead pulls the call. Twelve minutes of audio, half in Hindi. Someone sits through all of it, and even then the real question has no answer: why did the agent say that?
(An illustration, not a real call.)
What should you be able to see for every call?
Five things. If a vendor can't show all five for a call you pick, you don't have an audit trail. You have a pile of audio.
- What the agent heard. The transcript, and the audio where your clinic has chosen to record.
- What it understood. The intent, the patient or enquiry it matched, the language it switched to.
- Which rule it applied. The approved content or policy behind each answer. A price should trace back to the approved price sheet, not to "the model said so".
- What it did. Every action in your systems: slot booked, CRM updated, follow-up scheduled.
- Who took over, and why. Every handover, refusal and human override, with the reason.
Two, three and five are the ones I'd ask for first. They answer "why".
On our side, we review what the agent understood, the policy it applied, the action it took and when a person overrode it. That's the bar I'd hold any vendor to. Including us.
Is a call recording enough for an audit?
No. A recording tells you the words. It doesn't, by itself, verify the system's decisions.
Back to the knee call. With audio alone, the quality lead can confirm the sentence was said. With a trail, she can see the answer came from a patient-education script someone uploaded months ago, that nobody had marked it "clinical", and that no rule sent recovery questions to a nurse. Now the fix is obvious. Change the content, add the rule, re-test.
Without the trail you end up "fixing the prompt" and hoping.
Recordings carry their own risk too. A recording that identifies a patient is personal data; protect it and set retention and erasure in line with applicable requirements. More recording is not more safety. Set recording and retention per clinic, and make sure the trail is still useful on the days recording is off.
Which calls should your team actually review?
Not all of them. I'd start with the risky calls, not random happy-path ones:
- Every handover. Was it needed? Did the person who picked it up get the context, or ask the patient everything again?
- Every refusal or clinical question. Did the agent stay administrative? Did the patient reach a clinician?
- Every complaint, and every callback within a day. A patient ringing back that quickly may mean the first call didn't do its job.
- Every failed or mismatched write. The call says "booked". The calendar says nothing.
- A small random sample in each language. Hindi, Marathi, Tamil, code-mixed. I'd lean the sample towards the language your reviewers speak least.
Then get the right people in one room. Ask the front desk to check whether a booking was right. Ask a clinician to judge whether an answer drifted into advice. Ask compliance to look at consent and disclosure. I'd rather have one review with three lenses than three reviews that each assume someone else checked.
That's the shape of it on a typical deployment. The review has a regular cadence. Weekly is what I'd set as the default. A quality or operations lead from the clinic runs it. Someone from the front desk sits in to check whether bookings and follow-ups were right. A clinician joins when the sample includes clinical or policy-sensitive calls. Compliance looks at consent and disclosure.
The calls that come up first are the risky ones: handovers, complaints, low-confidence or policy-sensitive calls. Then a small random sample, so the quiet calls get a look too. When a reviewer finds an answer that shouldn't have been given, or a rule that didn't fire, that finding goes back into the agent's rules and the hospital's approved content. The next review checks that the change held.
How do you check policy adherence, not just call quality?
Call quality asks "did it sound good?" Policy adherence asks "did it stay inside the lines we agreed?"
Different scorecards. A warm, fluent call that quotes a price nobody approved is a failure. A slightly robotic "I can't answer that, let me connect you to the nurse" may be a perfect call.
So write the policies down before go-live, as things a reviewer can check:
- Cost answers come only from the approved price sheet.
- Symptom, report and medicine questions are handed over, not answered.
- The agent says it's an AI on every call.
- No appointment is confirmed until the calendar confirms it.
Then review call by call against those lines. If a reviewer can't answer yes or no from the trail, it isn't a policy yet. It's a hope.
In India there's one line I'd always add. The Telemedicine Practice Guidelines (2020) say platforms based on artificial intelligence or machine learning "are not allowed to counsel the patients or prescribe any medicines to a patient" (section 5.4). Your counsel can tell you how that applies to a phone agent. I'd still put "did it drift into counselling?" on the review sheet. We've written separately about keeping voice agents out of medical advice.
What outcomes should you monitor after go-live?
The ones the agent exists to produce. Krisper, our healthcare voice agent, works through patient enquiries already in a hospital's CRM: it answers questions on cost, process and timeline using the hospital's approved information, books a consult for patients who are ready, and writes the outcome back to the CRM.
So I'd monitor outcomes, not minutes:
- Task completion. Did each enquiry reach a clear next step?
- Handovers. How many, why, and did they land with a person?
- Policy adherence. What share of reviewed calls passed every line on your sheet?
- Operational outcomes. Consults booked, follow-ups due, enquiries closed.
And the check I'd never skip: does the CRM agree with the calls? Reconcile what the agent wrote back against your calendar and records. If the agent says forty consults were booked and the calendar shows thirty-six, someone has to go and find the four.
(Those numbers are made up. The habit isn't.)
Our buyer's guide makes the case for outcomes over cheap minutes before you buy. This is the version you run every week after.
Monitor across releases, too. Re-test every change to the model, content, rules or integrations before it reaches patients, and record which version handled each call. When a pattern shifts on a Thursday, the first question is: what changed on Wednesday?
What do Indian data rules say about logs and review?
The Digital Personal Data Protection Rules, 2025 set minimum security safeguards for a Data Fiduciary (the hospital, if it determines the purpose and means of processing). Rule 6(1)(c) includes "visibility on the accessing of such personal data, through appropriate logs, monitoring and review". Rule 6(1)(e) requires those logs and personal data to be retained for one year, to detect and investigate unauthorised access, prevent recurrence and support continued processing after a compromise, unless another law requires otherwise.
Two caveats. Rule 6 comes into force eighteen months after publication in the Official Gazette dated 13 November 2025, which is 13 May 2027 (rule 1(4), as corrected). And it's a security-safeguards rule. It doesn't, on its own, require you to review what the agent said.
I'd still build both on the same base. The trail that shows who opened a transcript should be the one that shows why the agent answered the way it did.
This is not legal advice. Our DPDP post for hospital voice AI goes deeper on consent, retention and roles.
Where does audit sit in the safety layer?
Underneath everything.
At Iksha we describe four layers. The first three, with observability and audit underneath, are the safety and policy layer. Applications, including Krisper, run on it:
- 01 Custom healthcare models. Specialised models for healthcare: audio, vision, text.
- 02 Hardening and secure protocol. Identity, consent, data boundaries and audit, in line with compliance and the law.
- 03 Business policy and guardrails. The clinic's rules: permitted actions, escalation and human review, enforced while the agent is working.
- 04 Applications. Healthcare agents ready to deploy, including Krisper.
Under every layer sits observability and audit: every interaction, decision and handover recorded and reviewable. Recording lives here, not in the guardrails. Guardrails decide what the agent may do. Observability and audit is how you show what it did.
That placement matters to a buyer. If audit is bolted on to one product, you see only what that product chooses to log. If it sits under every layer, ask to follow the model's understanding, the policy check and the system action on one timeline.
This is where the trail earns its keep. The recording tells you the words. The trail tells you why: the transcript, what the agent understood, which rule it applied, the actions it took, the handover, and the CRM entry, on one timeline.
The kinds of things it shows are ordinary, and that's the point. An answer drawn from content nobody had updated. A handover whose reason only appears in the trail, not in the audio. A CRM entry that doesn't match what happened on the call. The team then updates the approved content, tightens the rule, or fixes the write-back, and the next sample checks that it held.
The checklist I'd take into a vendor review
Pick the call yourself. Then ask the vendor to show you:
- What the agent heard, understood, applied, did and handed over on that call.
- The rule or approved content behind one specific answer.
- A handover from the receiving person's side. What did they see?
- Your clinic's recording and retention settings, and the trail with recording off.
- How you filter for handovers, refusals, complaints and failed writes.
- Last week's policy-adherence results.
- Agent outcomes reconciled against your CRM and calendar.
- Which version of the agent, content and rules handled a given call, and what changed in the last release.
If the answer to any of these is a slide, ask again next week.
FAQ
How do you audit conversations from a healthcare AI agent?
Pull any single call and check five things: what the agent heard, what it understood, which approved rule or content it applied, what it did in your systems, and whether and why a person took over. A recording alone cannot verify all five.
Is recording every call required to audit an AI voice agent?
Not necessarily. Recording is one input, and recordings that identify a patient contain personal data. Set protection, retention and erasure per clinic in line with applicable requirements. The audit trail should still show decisions, actions and handovers when recording is off.
Which AI agent calls should a hospital review?
Review by risk: every handover, refusal or clinical question, every complaint or callback within a day, every failed or mismatched system update, and a small random sample in each language your patients use.
What is the difference between call quality and policy adherence?
Call quality asks whether the call sounded good. Policy adherence asks whether the agent stayed inside the rules the clinic approved, such as quoting only approved prices and handing over symptom questions. A fluent call can still fail policy.
Do India's DPDP Rules require logs for AI voice agents?
Rule 6 of the DPDP Rules, 2025 lists security safeguards that include appropriate logs, monitoring and review of access to personal data, with those logs and personal data retained for one year unless another law requires otherwise. It comes into force on 13 May 2027. It is a security rule, not a rule about reviewing conversations. This is not legal advice.
Source notes
Reviewed 17 November 2026 (Asia/Calcutta). Legal references are to the official documents listed below. Iksha product statements in this article describe Iksha's own product and stated design, not an independent audit of every deployment. This is not legal advice.
- Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), 13 November 2025, rules 1(4) and 6(1)(c), (e): https://www.meity.gov.in/static/uploads/2025/11/53450e6e5dc0bfa85ebd78686cadad39.pdf (read with the corrigenda: https://www.meity.gov.in/static/uploads/2025/12/3c7ebbae0e5456f493f486e6845df86b.pdf)
- Telemedicine Practice Guidelines, 25 March 2020, section 5.4: https://esanjeevani.mohfw.gov.in/assets/guidelines/Telemedicine_Practice_Guidelines.pdf
- Iksha product statements describe Iksha's own product and stated design, not an independent audit of every deployment.
