Why Voice Copilot

Why Voice Copilot β€” The Reasoning, The Logic, The Flow

For someone who caught the gist from the pitch deck and wants the deeper "why." This page explains the problem, why the obvious solutions don't solve it, how we're different from Sanas and the other voice-agent platforms, the system flow, and an FAQ.

The problem, precisely

India's contact-center industry is being reshaped by voice AI. BPOs are putting AI voice agents on real calls alongside human agents. That creates a question every BPO's enterprise client now asks, and none of them can answer today:

"Prove to me that quality doesn't drop when a bot takes the call."

The reason they can't answer it is structural: the tools that measure human agents and the tools that measure AI agents are two completely separate stacks, on two different rubrics.

  • Human-agent quality is measured by QA suites (Cresta, Observe.AI, Balto) β€” enterprise-priced, US-first.
  • AI voice agents are measured by pre-deployment eval tools (Coval, Cekura, Hamming) β€” a different rubric, and they test bots before they go live, not in production.

So a BPO running a hybrid floor β€” some calls human, some bot β€” has no single scorecard that grades both, in production, on the same scale. They literally cannot compare their humans and their bots, or prove parity to the client. That's the gap.

Layer on three India-specific pressures:

  1. Compliance is statutory and deadline-driven. RBI Fair Practices Code (identity disclosure within ~30s, collections calling hours 8am–7pm IST), TRAI DLT/DND, and the DPDP Act (full enforcement ~May 2027, penalties to β‚Ή250 crore). Every incumbent ships US-law compliance packs (Mini-Miranda, TCPA). Nobody ships RBI/TRAI/DPDP-aware call scoring.
  2. Mid-market is priced out. Cresta/Balto start at ~$40K–$150K/yr with 50-seat minimums. A 20–200 seat Indian BPO has zero real-time copilot + QA option below that.
  3. Vernacular is unserved. Calls run in Hindi/Hinglish/regional with mid-call code-switching. Accent-maskers are English-only.

The reframe (the core idea)

Most "AI for contact centers" pitches are one of: mask the accent, coach the human, or replace the human with a bot. Each shrinks or competes head-on with a funded incumbent.

We reframed the product as a referee, not a player:

The quality + compliance referee for the hybrid call floor β€” one rubric, in production, scoring every call regardless of whether a human or an AI agent handled it, with India-regulation evidence built in.

Why this position is durable:

  • It grows as automation grows. Every bot added to a floor is a new thing that needs scoring β€” not a lost seat. A "coach humans" product shrinks as bots take over; a "referee" product expands.
  • It answers the BPO's actual existential question (quality parity, above) β€” which no mid-market tool can answer today.
  • A referee doesn't need to own call volume. Long-term the buyer flips to the enterprise brand that runs 3 BPOs + 2 bot vendors and has no neutral scorecard across them. The BPO can't referee its own game; the bot vendor won't grade its own homework. Independent measurement is structurally a third party's to win β€” the credit-rating-agency dynamic.

How we're different from Sanas and the rest

There are three adjacent categories. We are none of them, and that's deliberate.

vs. Sanas / Krisp (accent maskers)

Sanas is a funded, patented (US and India) accent-conversion company; two of the largest BPOs on earth are investors. They change how an agent sounds. They do not measure quality and do not train anything.

We are different in two ways:

  1. We don't erase accents β€” we coach clarity. Accent "neutralization" is contested ground: Bloomberg ("AI to neutralize Indian accents"), UCLA ("AI white voice"), and Oxford applied-linguistics have all called it dehumanizing. We keep the agent's own voice. Our voice-twin tech (which can do a clean rendition) is repositioned as a coaching mirror β€” "hear how your best self handled that turn" β€” not a mask. "Clarity through coaching, not erasure through masking" is contested but unowned ground.
  2. We measure, not mask. Sanas's output is audio. Our output is a score and a coaching signal β€” the rubric that grades every call, and the training signal a self-improving AI lane is built on.

vs. Cresta / Observe.AI / Balto (human-QA suites)

They score human agents well β€” but US-first, $60K+/yr, 50-seat minimums. The Indian mid-market is priced out, and they don't grade AI agents on the same rubric in production.

We are different: mid-market-priced, India-compliance-native, and our rubric is agent-type-agnostic β€” it scores humans and bots on one scale.

vs. Coval / Cekura / Hamming (bot-eval tools)

They test AI voice agents β€” but pre-deployment, on a bot-specific rubric, disconnected from how humans are judged.

We are different: we evaluate bots in production, on the same rubric as humans β€” not just pre-deployment on a bot-specific rubric. Turning that production signal into an automatic training loop is our roadmap (see the flywheel below), and it's a path the pre-deployment tools structurally don't have.

The one-liner: The only system that coaches humans and scores humans and AI agents on one rubric, in production β€” with India compliance built in.

The logic / the flow

The two lanes and the shared rubric

PRODUCTION Β· HUMAN LANE                    AI-AGENT LANE
─────────────────────────                  ─────────────────────────
Customer β†’ Human Agent                     Scripted customer drives the call
   β”‚                                            β”‚  (rehearsed, no human mic needed)
   β”‚  Live Coach Agent                          β”‚
   β”‚  (real-time rubric + next-best-action,     β”‚  AI Voice Agent
   β”‚   grounded in handbook via RAG)            β”‚  (carries memory of the last
   β”‚                                            β”‚   call into the next)
   └──────────────┐              β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                  β–Ό              β–Ό
        Compliance & Self-Correcting Agent
        (live eval every turn: RBI/TRAI/DPDP,
         anti-hallucination, mid-call correction)
                  β”‚
                  β–Ό
             UNIFIED RUBRIC  ── one scale for BOTH lanes:
             empathy Β· accuracy Β· resolution Β· compliance Β· clarity Β· CSAT

Live today vs. the target. The solid lines above run in production: the live coach, the compliance/self-correcting agent, the unified rubric across both lanes, and a scripted customer that lets a whole AI-agent call run with no human on a mic. The self-improving loop below β€” where scores automatically tune the agent and a threshold gate promotes it to live calls β€” is the target architecture, not the current build. The AI agent today remembers across calls; it does not yet learn from its scores.

The eval-driven flywheel (where this compounds) β€” the target loop

The insight that makes it a product and not a feature: every real human call is training signal for the AI, graded on the same rubric as your best humans. Steps 1 and 3 run today; steps 2, 4 and 5 are the roadmap that turns measurement into a self-improving loop.

  1. Capture (live today) β€” a real human call + the live-coach nudges are recorded and scored.
  2. Simulate (partial: scripted today, adaptive replay is the target) β€” a scripted customer already drives a whole AI-agent call; replaying a similar captured scenario to the AI is the next step.
  3. Score (live today) β€” one rubric grades the human call and the AI's attempt, on the same scale.
  4. Improve (target) β€” low scores tune the AI's prompts, RAG context, and the coach's nudges (eval-driven, in-context β€” no fine-tuning required to start).
  5. Gate (target) β€” when the AI crosses the quality + compliance threshold, it's promoted to supervised live calls.

The result we're building toward: today's human-augmentation product becomes the training ground and governance layer for tomorrow's AI agents. Augmentation funds automation, on one measured, compliant rail. Today the measured, compliant rail is real; the automatic improvement on top of it is the roadmap.

Why compliance is a code-owned clock, not an LLM guess

Compliance must not hallucinate. So the system splits the work:

  • Code owns the clock and the deterministic facts β€” the 30-second disclosure timer, the 8am–7pm IST calling window (computed from timestamps). These are not LLM judgments.
  • The LLM only judges meaning β€” did this turn contain a required disclosure? β€” and must cite a verbatim quote.
  • The grade gate: only authoritative (verbatim-regulation) rules can cap a call's QA score. Industry-guidance rules (e.g. the exact 30s threshold) render as advisory β€” flagged, never used to fail a call. This is deliberate: over-claiming compliance is the exact failure the product exists to prevent, so the system refuses to hard-fail a call on a threshold that isn't literal regulation.

Why mid-call self-correction (the Live Agent Corrector)

A passive monitor that only reports a bot's mistakes after the fact isn't enough for a "governance layer." So the AI lane closes the loop: when the bot misinforms (contradicts its own earlier turn), skips a required compliance line, or sounds robotic, a supervisor injects a steering instruction and the bot corrects itself live, in its own voice β€” "I'm sorry, I misinformed you earlier…", "let me check that for you." Two guardrails keep it honest: it only auto-corrects on high-confidence, checkable errors (compliance gaps + self-contradiction β€” never open-domain fact-guessing without a knowledge base), and a turn-taking guard ensures it never talks over the customer.

Why GPT-S2S realtime today (and the vernacular plan)

The live product runs on OpenAI gpt-realtime speech-to-speech. This is a current-state engineering choice, and it's deliberate:

  • Latency drives experience. gpt-realtime gives ~400–500ms time-to-first-audio β€” a genuinely conversational feel with tuned VAD + barge-in.
  • The alternatives we evaluated:
    • Gnani Warp (their real-time S2S) is in beta with ~1.2s latency (per their own team) β€” too high for the richer live experience we want today.
    • Sarvam has no native S2S β€” it's STTβ†’TTS, which chains two hops and adds latency.
  • So we use GPT-S2S now for the richest, lowest-latency experience, while the NVIDIA-native target keeps Gnani Vachana (their STT/TTS speech models on NVIDIA Nemotron) as the path to Indic/Hinglish vernacular β€” a different need (language coverage + data residency) than the real-time S2S loop. Warp-the-S2S-product and Vachana-the-speech-models are not the same thing; we pick the best tool per requirement.

The target architecture (NeMo Agent Toolkit for orchestration, NeMo Evaluator for the rubric, NeMo Guardrails for compliance, NeMo Retriever for RAG, self-hosted NIM for DPDP data residency) is the migration plan, not the current build.

FAQ

Q: Isn't this just another AI-coaching or QA tool?

No. Coaching and QA are shipped, commoditized categories. Our wedge is the one thing none of them do: one rubric that scores humans and AI agents in production, with India compliance native. The coaching and QA capabilities are table stakes we already have; the referee position is the differentiator.

Q: Aren't you helping BPOs replace their own human agents?

The framing matters and we're deliberate about it: the product is "route work to whoever does it best, and prove quality to the client" β€” not "build the case to fire humans." In the near term it up-levels humans (live coaching) and governs the AI (eval-gated, never live until it matches human quality). The humans are how the AI learns what "good" is.

Q: How is this different from Sanas? They're also "voice AI for Indian contact centers."

Sanas changes how an agent sounds (accent conversion) and measures nothing. We don't touch the accent β€” we coach clarity in the agent's own voice, and our output is a score and a training signal. Different product, different value, and we sidestep both the accent-masking backlash and Sanas's India patent.

Q: Who evaluates the evaluator? An LLM judge can be wrong.

Three defenses. (1) Compliance is code-owned where it can be (timers, calling windows) β€” not an LLM guess. (2) The grade gate only lets verbatim-regulation rules affect a score; anything softer is advisory. (3) The roadmap includes a human-labeled calibration set for the LLM-as-judge rubric. We never market it as "guaranteed compliance" β€” it's evidence and monitoring.

Q: Why not just build a replacement AI agent like everyone else?

Because the replacement-bot market is crowded and undifferentiated, and because a bot vendor can't credibly grade its own bot. The neutral referee is a structurally defensible position that grows as bots proliferate, rather than competing to be one of them.

Q: Why does the AI need "training" if there's no fine-tuning?

"Training" here means eval-driven, in-context improvement β€” the design is that rubric scores from real calls tune the AI's prompts, its RAG grounding, and the coach's nudges, rather than retraining weights. That loop is the roadmap, not yet wired: today we produce the scores (the signal); automatically feeding them back is the next step. Fine-tuning (NeMo Customizer) is held further in reserve until enough consented, labeled call data has accrued. Starting inference-only means when the loop lands it improves without a training pipeline.

Q: Is the compliance layer real or a mock?

Real and live. India-compliance scoring (Layer 2) runs on production calls today β€” a live nudge during the call and a per-rule, transcript-anchored evidence section after it, surfaced in the review and on the manager dashboard's compliance column.

Q: What about data privacy β€” you're scoring calls?

Data residency is a first-class design goal, not an afterthought β€” the NVIDIA target self-hosts inference via NIM so call data stays in-country for DPDP. Our own pipeline follows consent/retention/erasure discipline: we score others on compliance, so our own handling must be clean.

Q: What's actually built vs. planned right now?

Live: the human copilot (voice twin + live coach + sentiment + post-call rubric), the live AI-agent studio (scored on the shared rubric), two-party live calls, the manager dashboard, India-compliance scoring, and mid-call AI self-correction. Built, not yet deployed: knowledge-base RAG + domain guardrails. Planned: the NVIDIA/Gnani vernacular migration, third-party bot-call ingestion, and multi-tenant scale.

Q: Where do I try it?

/demo (no password, zero-cost). For the live surfaces and how to use each, see the companion Resources guide β€” with where to enter the access password.