What Should Diagnostic Labs Pay for Voice AI?

By Peush Bery, Xtreme Gen AI

Diagnostic labs should not buy the cheapest AI minute. They should price the complete workflow and measure cost per completed patient action.

Highlights

• Per-minute pricing is useful for measuring infrastructure consumption, but it does not prove patient or business value.

• Cost per call is misleading when a five-second wrong number and a four-minute completed booking count equally.

• Pure outcome pricing becomes difficult when outcome definitions, duplicate leads, slot capacity and patient decisions are outside the agent’s control.

• Diagnostic labs should use different outcome denominators for bookings, report support, reminders and home collection.

• Managed pricing can be rational when it includes integrations, prompt and tool logic, retries, WhatsApp continuity, QA and ongoing changes.

• The commercial comparison should calculate cost per reliable patient action while retaining transparent usage and quality controls.

A diagnostic lab is not buying minutes

A patient calls to ask whether fasting is required, another wants a home collection slot, and a third cannot open a report link. All three conversations consume Voice AI minutes, but they do not create the same operational value. One prevents a failed visit, one may create revenue, and one resolves a post-service support need.

This is why Voice AI pricing for diagnostic labs in India should not be reduced to the lowest rate per minute. The minute is a useful accounting unit because telephony, speech recognition, language processing and voice generation operate during the call. The buying decision, however, should ask what reliable patient action the complete workflow produces.

Why diagnostics needs its own pricing logic

Diagnostic operations combine commercial, service and patient-experience calls. Package enquiries and home collection requests may influence revenue. Appointment reminders and preparation instructions protect capacity. Report-status calls reduce support load. Sensitive questions require human escalation rather than automation. A single “conversion” number cannot describe all of them.

India’s National Health Accounts estimated household out-of-pocket health expenditure at ₹3,56,254 crore in 2021-22, equal to 39.4% of total health expenditure. This does not measure diagnostic Voice AI demand, but it reinforces an important operating reality: patients frequently experience healthcare purchases directly, so clarity, trust and service reliability matter alongside efficiency.

Model one: per-minute pricing

Per-minute pricing is straightforward when inclusions are clear. The lab can forecast connected usage and compare campaigns over time. It also reflects a real cost structure: telephony carries the call, speech-to-text processes audio, the model reasons and calls tools, text-to-speech generates the response, and systems create recordings, transcripts, summaries and CRM updates.

The quoted minute still needs qualification. Ask whether duration is aggregated to the second or rounded by call, whether voicemail and silence are billable, whether telephony is included, which speech and language models are used, whether concurrency costs extra, and whether transfers, tool calls, recordings, WhatsApp and support are included. Two ₹3 minutes can contain very different products.

Per-minute pricing works well as the consumption layer, particularly when the lab runs several workflows. Its weakness is that it can reward long conversations even when they fail. The commercial dashboard must therefore connect minutes with outcome quality.

Model two: per-call pricing

A per-call model looks simpler, but the denominator is unstable. A wrong number, a five-second disconnect, a voicemail encounter, a two-minute report-link resolution and a six-minute booking may all count as one call. Vendors can protect themselves with minimum durations or connected-call definitions, but every rule changes the effective rate.

Per-call pricing can fit narrow reminder flows when calls have similar length and one clearly defined message. It is weaker for open diagnostic enquiries where conversation length depends on test requirements, pincode, slot search, language, pricing questions and escalation. Labs should simulate their actual duration distribution before accepting an apparently low call price.

Model three: outcome pricing

Outcome pricing sounds perfectly aligned: pay for a booked test, resolved report query or confirmed appointment. The difficulty is defining an outcome the agent can control and both parties can verify. If a patient is qualified but no home-collection slot exists, did the agent fail? If the patient books after receiving a WhatsApp link, which channel receives credit? If a duplicate lead was already in CRM, should it be billable?

Outcome contracts need rules for attribution windows, duplicates, cancellations, patient eligibility, branch capacity, manual overrides, transferred calls and CRM evidence. The more valuable the outcome, the more likely it depends on pricing, service availability and human action outside the Voice AI system.

Outcome pricing can work for tightly bounded actions such as a confirmed reminder response or a completed callback request. For complex journeys, a hybrid model is usually more honest: transparent usage or capacity pricing plus agreed operational outcome targets.

Use a different outcome for each diagnostic workflow

For new enquiries, measure cost per qualified booking opportunity and cost per completed booking. For home collection, measure cost per serviceable request with pincode, test list and preferred slot captured, then cost per completed collection. For report support, measure cost per resolved access issue or correctly routed query. For reminders, measure cost per confirmed, rescheduled or cancelled appointment.

For pre-test instructions, the useful outcome may be confirmed understanding, a reschedule, or a human callback for an exception. For membership renewals, measure a qualified renewal response rather than forcing every conversation into an immediate payment outcome. Each denominator should represent a patient action that the workflow can reliably create and the lab can verify.

A worked calculation for a diagnostic campaign

Assume a lab spends ₹90,000 on a month of Voice AI usage and managed operations. The campaign produces 6,000 eligible connected conversations, 1,800 qualified patient actions and 900 completed bookings or resolved service cases. The usage cost is ₹15 per eligible conversation, ₹50 per qualified action and ₹100 per completed outcome.

Those figures are illustrative, not market benchmarks. Their purpose is to show why one invoice produces several cost views. A competing quote with a lower minute rate may be worse if it creates fewer verified actions, requires more internal staff or writes unreliable CRM dispositions. A higher managed rate may be justified if it reduces operational cleanup and produces more completed patient journeys.

The lab should compare the same workflow, lead set, time window and outcome definition across vendors. Changing the denominator halfway through the comparison makes every price claim meaningless.

What managed service changes in the price

A self-serve platform gives the lab tools to create and run agents, but someone inside the organisation must own prompt design, API connections, telephony, retries, dispositions, reporting, quality review and ongoing changes. Those internal costs may not appear on the vendor invoice.

Xtreme Gen AI’s managed model includes more of that operating layer: implementation, prompt and tool logic, callback and retry rules, CRM or LIS-facing API workflows, WhatsApp memory, reporting, QA and maintained changes. The lab should not assume managed is automatically cheaper. It should calculate whether transferred ownership, faster implementation and lower internal burden justify the commercial difference.

The right comparison is total quarterly ownership. Include vendor charges, telephony, numbers, concurrency, messaging, internal engineering, operations, QA, integration maintenance and the cost of unresolved or incorrectly handled patient actions.

Quality changes the economic result

A cheaper voice or speech model may be entirely appropriate for a short English reminder. The same stack may struggle with noisy mobile audio, regional-language switching, medicine or test names, addresses and interruptions during a home-collection enquiry. The least expensive stack that reliably produces the required outcome is the rational choice.

Labs should test real calls and score intent accuracy, critical field capture, tool success, safe boundaries, CRM disposition, callback accuracy and human handoff. Conversation naturalness matters, but a beautiful voice that books the wrong slot or loses the pincode is commercially expensive.

Do not turn outcome pressure into unsafe automation

A vendor paid only for bookings may be incentivised to push conversations further than the lab wants. Diagnostic workflows need boundaries. Voice AI should not interpret medical results, provide unapproved clinical advice or hide uncertainty to preserve conversion. Sensitive questions, complaints and exceptions should move to qualified human processes.

The pricing agreement should protect quality with exclusion and escalation rules. A safely routed medical question may be a successful workflow outcome even though it is not a booking. A patient opt-out should be honoured even though it reduces campaign output.

Compliance and data handling belong in the commercial scope

TRAI’s framework makes consent, preferences and registered commercial communication practices relevant to calling campaigns. The DPDP Act also matters when phone numbers, recordings, transcripts, addresses, report-related requests and other digital personal data are processed. The lab should validate its obligations with appropriate legal and compliance advisers.

Commercial proposals should specify calling infrastructure, consent and suppression responsibilities, data access, retention, security review, recording policy and incident handling. A price that excludes essential governance is not the complete production price.

The buying scorecard

Ask every vendor for the same four numbers: expected cost per connected minute, expected cost per eligible conversation, expected cost per qualified patient action and expected cost per completed outcome. Then ask which assumptions create each number and which costs sit outside the quote.

Review the ownership model, language and call quality, telephony, integrations, concurrency, WhatsApp continuity, human handoff, QA, reporting and change turnaround. Run a controlled pilot and inspect failed calls as carefully as successful ones. The objective is not to prove that AI can talk; it is to prove that the workflow moves patients correctly and economically.

Conclusion

Diagnostic labs should not abandon per-minute pricing, because the minute remains a practical way to account for variable infrastructure. They should stop treating it as the final measure of value. Per-call pricing can hide duration differences, while pure outcome pricing can hide attribution and operational dependencies.

The more useful model is transparent consumption plus workflow-specific cost per reliable action. Price bookings as bookings, report support as resolved or correctly routed queries, reminders as confirmed actions, and home collection as serviceable and completed requests. That is how a lab can compare Voice AI as an efficiency and capacity system rather than a cheap calling trick. To experience the Xtreme Gen AI Voice AI Agent, call 9228034172.

Sources and further reading

National Health Accounts 2021-22: https://www.mohfw.gov.in/sites/default/files/NHA%202021-22_up.pdf

TRAI: What is Spam or UCC: https://www.trai.gov.in/what-spam-or-ucc

TRAI TCCCPR framework: https://www.trai.gov.in/tcccpr

Digital Personal Data Protection Act, 2023: https://www.meity.gov.in/writereaddata/files/Digital%20Personal%20Data%20Protection%20Act%202023.pdf

Xtreme Gen AI pricing-unit guide: https://www.xtremegenai.com/blogs/voice-ai-pricing-per-second-minute-outcome-india

Xtreme Gen AI diagnostic report-ready calls: https://xtremegenai.com/blogs/voice-ai-diagnostic-report-ready-calls