Voice AI in India: Language, Trust and the Economics of a Resolved Call

Voice AI becomes an enterprise capability when it completes an authorised customer task reliably. Natural speech is necessary for many use cases, but it is not enough. The system must understand the request, apply policy, use the correct operational tools, confirm important information and transfer the customer when automation is unsuitable.

This report merges two August 2026 studies of Indian enterprise voice agents. Company, client, institution and personal names have been removed. Non-comparable speech benchmarks and current-price claims have been removed; remaining scenarios are explicitly illustrative or dated forecasts. The main call is that buyers should measure resolved outcomes and full operating economics rather than select a supplier on a polished demonstration or a low per-minute price.

The public-service opportunity and its limits

MeitY’s 1 June 2026 BHASHINI–DPIIT release reports support for 36 Indian text languages,23 Indian voice languages and 35 international languages; more than 800 government websites and 15 million daily inferences were reported. The partnership aims at access across 22 Scheduled languages. These categories are not contradictory: Scheduled-language coverage, voice-service support and all supported text languages have different boundaries.

Those figures demonstrate programme reach as reported by government, not accuracy on a pension query, a disputed payment or an unfamiliar name. The India policy opportunity is to make important services usable without forcing the citizen to master a particular language or interface. That requires testing completion by language, accent, age and connection quality—not merely translating a landing page.

Public buyers should publish the fraction of eligible requests completed correctly, repeat-contact rate and the performance of the human fallback. A service that transfers its hardest users cannot claim its automated completion rate as the result for all citizens. The transfer should preserve context and avoid making the person repeat sensitive information. Cost savings become a public benefit only when the service remains accessible and the cost of those exceptions is included.

The useful architecture is often hybrid

A voice agent combines telephony, speech recognition, conversational reasoning, policy, tool execution, speech generation, logging and human support. Different suppliers package these components differently. A single “voice AI” price can conceal several pass-through services and separate operating responsibilities.

A cascaded architecture converts speech to text, processes the text and generates spoken output. It can make transcripts and intermediate decisions easier to inspect. A native speech-to-speech model can reduce some latency and preserve conversational cues, but the enterprise still needs evidence, control and reliable tool use.

The historical study cited cascaded latency of roughly 600 milliseconds to 1.2 seconds and native speech-to-speech latency of approximately 250–500 milliseconds. Those are source ranges, not comparable guarantees. End-to-end performance also includes telephony, endpointing, tool calls and policy checks.

The hybrid conclusion is therefore practical: use conversational models where they improve interaction, deterministic rules where policy must be exact and human support where judgement or recovery is required.

Figure 01 · Conceptual framework

From spoken request to completed service

How the parts connect

1. AAuthorise contactauthorisedUnderstand intent
2. BUnderstand intentunderstoodApply policy
3. CApply policypermittedExecute and verify
4. DExecute and verifyconfirmed successConfirm resolution
5. EUnderstand intentuncertainHuman handover
6. FApply policyoutside policyHuman handover
7. GExecute and verifyfailureHuman handover

Indian speech requires representative testing

Language switching, regional accents, background noise, interruptions and spoken numbers create a difficult operating environment. A system that performs well on clean English recordings may fail on the calls that dominate a particular customer population.

The earlier report compared error-rate ranges taken from different evaluations. Those figures have been removed from the public comparison: different languages, microphones, task mixes and scoring rules cannot establish one supplier’s superiority. A valid benchmark uses the same representative audio, transcription standard and outcome test.

Word error rate is useful but incomplete. Misrecognising a conversational filler differs from misrecognising an amount, date, address or account instruction. Test critical-field accuracy separately. Require explicit confirmation before consequential actions, and measure whether the customer understands what has been agreed.

Include real variation in the evaluation: mixed languages, unfamiliar names, elderly speakers, low-quality connections, repeated corrections and interruptions. A single headline language score can hide a failure in a customer segment the institution has an obligation to serve.

Choose tasks that can be bounded

Good early use cases have clear completion criteria and accessible data. Examples include appointment confirmation, document reminders, straightforward service requests and status updates. More complex tasks may be suitable with a human approval step or an assisted-agent design.

Collections, complaints, vulnerable-customer situations and consequential financial instructions require particular care in policy and escalation. The system should not improvise promises, pressure a customer or infer authority from conversational fluency. The permitted action set needs to be explicit.

The source’s use-case framework weighted eligible volume and task boundedness at 20% each, with data readiness, measurability, customer benefit and reversibility each at 15%. Those are historical procurement suggestions. Their purpose is to identify tasks where success can be measured without concealing risk behind a large call-volume opportunity.

A call is a controlled sequence

The agent first establishes that the call and contact are authorised. It listens, identifies intent and extracts relevant information. A policy layer determines what can be done. The system then uses a typed tool or workflow, checks the result, confirms critical details and records the outcome.

Each stage needs a failure path. If identity is uncertain, the system should not disclose protected information. If an API is unavailable, it should not claim that the task was completed. If the customer requests something outside policy, it should explain the limit and transfer appropriately.

Warm transfer is part of the product. The human agent should receive the relevant context and completed checks without forcing the customer to repeat everything. At the same time, the transcript should not be treated as proof that every statement in it is correct. Important facts may need confirmation.

Compare the full price stack

Historical price components must be normalised to a common service scope

These are historical source estimates, not current offers. They should not be added mechanically because some supplier packages include components that others charge separately. Messaging, support, analytics, storage and human fallback can create further costs.

The source gave effective-rate illustrations of roughly INR 9.9 per minute at 25,000 monthly minutes and INR 6.0 at 25 lakh minutes. The declining unit cost reflects fixed-cost absorption and commercial scale in the source example. It is not a guaranteed volume discount.

Figure 02 · Qualitative comparison

Test the population, not the demonstration

Test the population, not the demonstration

Test dimension	Include in evaluation	Measure
Language and speech	Include in evaluationAccents, code switching and spoken numbers	MeasureTask and critical-field accuracy
Environment	Include in evaluationNoise, interruption, silence and poor audio	MeasureRecovery and appropriate escalation
Authority	Include in evaluationWrong identity and out-of-policy requests	MeasureProtected information and action controls
Operational failure	Include in evaluationUnavailable APIs or incomplete downstream work	MeasureTruthful status and successful handover

Cost per resolved contact is the better denominator

Let $C$ be total operating cost over a period and $R$ the number of contacts whose intended task was correctly completed. Then

$$\text{Cost per resolved contact}=\frac{C}{R}.$$

The numerator should include platform, usage, telephony, integration amortisation, tuning, human escalation and rework. The denominator should exclude calls that sound successful but fail downstream or require the customer to contact the institution again for the same issue.

The source’s INR 12–25 AI and INR 40–120 human comparisons are not used as current procurement quotes. Rebuild both costs on the same task population, success definition and cost boundary. An explicitly illustrative test of ₹1,00,000 total cost and 5,000 accepted resolutions yields ₹20 per resolution; if only 4,000 qualify after repeat-contact checks, the true figure becomes ₹25.

A low-cost agent that creates more repeat calls can be more expensive overall. Measure repeat contact and eventual resolution, not just containment. Containment is only useful if the customer actually receives a correct outcome.

Staffing savings require a realistic transition

The historical study used a 300-seat operation at INR 32,000 fully loaded monthly cost per agent, producing INR 96 lakh monthly cost. That arithmetic is retained as a source illustration. It does not establish that all of that cost can be removed by automation.

Historical staffing forecasts concern different benefits and eligible volumes

These are historical forecasts and scenario assumptions. They depend on task mix and the retained human operation.

Some costs remain fixed, while difficult calls become a larger share of human workload. Training, supervision, exception handling and quality assurance do not disappear. A workforce business case should distinguish released capacity from cash savings and specify the operating changes needed to realise either.

Safety and quality must be tested together

Run a representative trial with documented outcomes. Include wrong customer details, ambiguous instructions, distress, anger, background voices, silence, interruptions and requests outside policy. Test whether the system recognises its limits and preserves the customer experience during escalation.

Evaluate critical-field confirmation, unsupported claims, unauthorised actions, policy adherence and tool-call accuracy. Measure latency and conversational quality, but do not let a natural voice compensate for an incorrect action. Test the workflow after every material model or policy change.

Disclosed buyer-scorecard weights prioritise verified outcomes

This is a useful emphasis, not a universal formula. Certain failures should block deployment regardless of aggregate score.

Figure 03 · Conceptual framework

The economics of a resolved contact

How the parts connect

1. ATechnology costsincludeTotal scoped cost
2. BOperating costsincludeTotal scoped cost
3. CTotal scoped costnumeratorCost per resolution
4. DVerified resolutionsdenominatorCost per resolution
5. EQuality guardrailsqualify successVerified resolutions

Commercial and operational dependence

A supplier may rely on external speech, language-model and telephony services. The buyer should understand those dependencies, especially where outages, data location or pricing changes can affect the service. A branded platform can be thin orchestration around several external components or a more integrated capability; the operating evidence should reveal which.

Contracts should address recordings, transcripts, retention, access, model training use, incident support and exit. The institution needs usable conversation and outcome data to assess performance independently. It should also know how a model update is tested and whether it can retain or roll back a configuration.

The historical forecast expected further price compression of 15–25% by end-2027 after earlier declines. That forecast is preserved without claiming it occurred. Falling inference prices can improve economics, but integration, telephony and human support may not decline at the same rate.

Distinguish a promise from a completed outcome

Collections and service workflows often use intermediate measures such as a promise to pay, a booking intention or a document request. These can be useful, but they are not final outcomes. A customer agreeing verbally to an action does not establish that payment arrived or the document was accepted.

The historical study included different cost-per-promise ranges for early and later delinquency. Those comparisons depend on case difficulty and contactability. A cheaper promise can have little economic value if fewer promises are kept. Link the call outcome to the subsequent business event, with an appropriate observation period and clear attribution limits.

The same principle applies to customer service. A call marked resolved should be checked against repeat contact and downstream completion. Review a sample of both successful and escalated calls, including critical-field confirmation and tool results. This detects the failure mode in which the agent ends the conversation smoothly while the task remains unfinished.

Quality monitoring should also examine which customers are excluded from automation. If difficult accents, older customers or complex requests are consistently transferred, the retained human operation needs suitable capacity. Automation performance should be reported on the eligible population and alongside the exclusions, rather than presented as a result for all incoming calls.

The deployment decision

Start with a bounded task, establish the existing cost and quality baseline, and test on representative calls. Approve production only when the agent completes the task, stays within policy and hands over effectively. Expand by task and customer population rather than extrapolating from one successful demonstration.

The strong voice-AI service combines conversation with controlled action and measurable resolution. That is the capability worth buying. A convincing voice, a large language list or a low minute price is evidence to examine, not the conclusion of the evaluation.

Source record and update scope

• MeitY / PIB — BHASHINI 36 Indian text/23 Indian voice/35 international languages; >800 websites; >15 million daily inferences; 22 Scheduled-language partnership aim. Observation/publication: 2026-06-01. Verified primary source; observed date stated.