Sarvam Voice Agent Handoff to Human Escalation (Delivery)

Sarvam Voice Agent Handoff to Human Escalation (Delivery)

1. 01

   Register the Sarvam API subscription key and install the Python SDK with 'pip install sarvamai'

   Confirm the key is scoped to the client workspace before any agent build begins, so usage credits bill to the correct retainer.

2. 02

   Validate the Text to Speech and Speech to Text endpoints with a small test payload

   Sarvam covers text-to-speech across 11 Indic languages and speech-to-text with diarization across 12. Confirm the client's chosen language pair returns audio and transcript in the expected format before wiring the agent.

3. 03

   Configure the confidence threshold that triggers fallback routing to a human agent

   Set the threshold in the voice agent configuration so low-confidence turns exit the bot instead of guessing. Document the chosen value in the client delivery file.

4. 04

   Connect the agent to the client's WhatsApp or phone number through the Sarvam API

   This is the same integration path used in the Sarvam Voice Agent Starter offer. Verify inbound and outbound routing on both channels before handoff.

5. 05

   Map the client's top 10 FAQ scenarios to agent intents and mark which ones must escalate

   Payment reminders and appointment booking can stay automated. Complaints, refunds, and anything touching account changes route to a named human queue.

6. 06

   Define the human handoff payload: transcript, detected language, confidence score, and caller number

   Sarvam's speech-to-text diarization output feeds this payload. The receiving agent needs the transcript and language tag to continue without asking the caller to repeat themselves.

7. 07

   Run a live escalation drill with the client's team and record the response time

   Place test calls in each supported language, force a low-confidence turn, and time how long the human queue takes to pick up. Anything above the client's stated SLA goes back to configuration.

8. 08

   Log the handoff rules, threshold value, and escalation contacts in the client workspace

   Retainer renewals depend on this record. When the client adds a language or a new queue, the next delivery sprint starts from this file rather than a blank build.