Voice AI Completes 38% of Real Tasks. Then You Add an Indian Accent.

Voice AI Completes 38% of Real Tasks. Then You Add an Indian Accent.

Navin, co-founder, Telenow

---

In March 2026, Sierra.ai — one of the best-funded enterprise voice AI companies in the world — published a benchmark called τ-Voice. It measured whether current voice agents can actually complete real customer service tasks: returns, exchanges, cancellations, plan changes, billing inquiries. 278 tasks. Three leading voice AI providers — OpenAI, Google, xAI. Two conditions: clean audio and realistic audio.

Here is the number the industry was avoiding.

Under realistic conditions — background noise, diverse accents, natural interruptions, the kind of audio a customer actually generates on a mobile phone — voice agents completed 26 to 38 percent of tasks. The best text model, running on the same problems, completed 85 percent. Voice agents retained 30–45 percent of text capability.

For the customer service category — the one voice AI has raised billions of dollars on the promise of automating — this is the number that decides the next five years. And it was published by a competitor, not a skeptic.

Why this number is worse than it looks

The paper tested three US English domains: retail, airline, telecom. It did not test Hindi. It did not test Hinglish, code-switching, Marathi-accented English, Bengali, or Tamil. It did not test RBI script requirements, DPDP consent capture, or IRDAI product disclosures. It tested American voice agents against American customer service tasks in American accents.

On that easiest possible version of the problem, the best voice agent completes 38 percent of tasks under realistic conditions.

Now add BFSI. Add regulated collections in Hinglish where the customer switches mid-sentence to Marathi. Add the KYC opening turn where the customer spells their name letter-by-letter over a 2G mobile connection with a fan running in the background. Add the RBI Fair Practices requirement that the agent must disclose its identity within 30 seconds, in a language the customer comprehends, with an audit trail that survives inspection.

The 38 percent number does not stay 38 percent. It gets worse.

Four findings that matter more than the headline

The paper contains four findings that any BFSI CIO or CRO needs to read into their next vendor evaluation.

Accent inequality is now measured, not theorised. Sierra tested seven personas including a Bengali accent, a Sichuan Chinese accent, a Senegalese-French accent, and a Maharashtrian Indian English accent. Under diverse accents, xAI lost 38 percent of its clean-condition capability. Google lost 2 percent. This is not a small difference — it is a chasm. And it means the choice of voice AI vendor for a bank with pan-India customers is now a DPDP fairness question and an RBI Fair Practices question rolled into one. Which side of that gap is your vendor on? If they cannot show you per-accent numbers, they do not know either.

Authentication is the dominant failure mode. From the qualitative error analysis: "Agents fail to transcribe names and emails even when spelled letter-by-letter, blocking all downstream actions." Every BFSI call in India begins with KYC or account identification. If horizontal voice AI cannot reliably transcribe a name spelled out over the phone under noise, the entire regulated use case is gated at turn one. Nothing downstream matters if the first turn fails.

Hallucinated completions. The paper documents agents stating out loud "I've updated your shipping address" without ever calling the underlying tool. In retail this is a bad customer experience. Translate to BFSI: "your EMI has been rescheduled," "your KYC is now complete," "your policy has been renewed," "your loan has been closed." Each of those is a regulatory event with real customer harm and — because voice leaves no scrollback — no easy audit trail. The customer heard it. The database says something different. The dispute takes six months to resolve, and the RBI complaint is already filed.

No vendor masters both task completion and conversational dynamics. OpenAI hits 100 percent responsiveness but 6 percent selectivity — it responds to nearly every backchannel, cough, and stray "hold on" from the customer. xAI leads on task completion but interrupts the customer 84 percent of the time. Google is polite (21 percent interrupt rate) but ignores 31 percent of user turns entirely. Every current vendor is world-class on one dimension and unusable on another. There is no "best" voice AI in the current architecture — only different failure modes distributed across different vendors.

This is architectural, not fixable

The tempting response to any of these findings is: "Newer models will fix this." They will not. Not without changing the architecture.

Every current voice AI stack rents STT, LLM, and TTS from three separate vendors and wires them together with a workflow engine. STT converts speech to text. LLM reads the text and generates a response. TTS speaks the response. The three components share no representation of the call. Tone, hesitation, code-switching, the audible sound of a customer losing patience — all discarded at the first hop. The reasoning layer never sees the audio. The speaking layer never sees the customer.

When failures show up in production, teams reach for the same fixes: bigger prompts, more guardrails, terminology dictionaries, workflow branches. Prompts hit 40 pages. Workflows branch into the thousands. Nobody names what is actually happening — it is compensation for a rented, generic brain that cannot be taught what your specific NBFC's collections vocabulary sounds like when a Marathi-accented customer says it over a bad connection.

Sierra's own conclusion, on the last page of the paper: voice agents will match text models only when they can "sustain fluid conversation while reasoning over multi-step tasks in real time — a constraint text agents, which can think silently for as long as needed, do not face."

That is the shared-representation argument, in a paper published by a leading voice AI vendor. The category is telling itself the answer.

Where regulated BFSI goes from here

The next generation of regulated enterprise voice AI is not going to be a horizontal API stack with better prompts. It is going to be a company-specific voice brain — one integrated stack, trained on the enterprise's own calls, running on the enterprise's own compute, improving weekly on the enterprise's own data.

One stack, because STT, understanding, and TTS have to share representations of the call.

On-premise, because audio never leaves the RBI, DPDP, and IRDAI perimeter.

Continual learning on the customer's own conversations, because the only way to get from 38 percent to 90 percent is training, not prompt engineering.

The API vendors cannot build this. Their unit economics depend on multi-tenancy, and on-premise single-tenant deployment breaks their margin model. The frontier labs will not build this either — every hour spent on RBI residency requirements is an hour not spent shipping their next model. This is not a competitive gap. It is a structural one.

Voice AI is about to bifurcate. Horizontal APIs will keep winning demos and running SMB pilots. Vertical, on-premise, company-specific voice brains will win regulated production.

Sierra just published the number that makes the bifurcation inevitable.

38 percent is not a starting point. It is a ceiling.

We are building what comes next.

---

— Navin Co-founder, Telenow

Reference: Ray, Dhandhania, Barres, and Narasimhan. "τ-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains." Sierra.ai and Princeton Language and Intelligence, arXiv:2603.13686, March 2026.