A production checklist for ElevenLabs voice agents

Eleven v3 provides expressive text-to-speech, while a production voice service also needs conversation behavior, knowledge, business tools, authentication, session state, and a transport such as web audio, Twilio, or SIP. Decide which provider owns each layer before implementation.

Document the latency budget across speech recognition, reasoning, tool calls, synthesis, and the carrier. A beautiful voice that routinely leaves three seconds of silence will feel broken to callers.

Request short-lived signed conversation URLs or session tokens from the ConveRact backend. The browser or embedded client should receive only temporary session material, never the ElevenLabs API key.

Scope credentials to minimum resources, encrypt them at rest, mask them in every response, and record validation and rotation timestamps. Reject cross-tenant voice configuration access at the API boundary.

Evaluate clarity, pacing, pronunciation, emotional range, and language support for the actual customer population. Use Eleven v3 where expressive synthesis fits, while choosing realtime settings based on latency and stability requirements.

Maintain a pronunciation dictionary for product names, people, addresses, currencies, and domain terms. Test numbers, dates, acronyms, mixed-language sentences, and poor telephone audio before publishing a voice.

Tell callers when they are interacting with AI where law, policy, or customer expectations require it. Obtain consent before recording, disclose how recordings and transcripts are used, and provide a human path.

Apply regional quiet hours, suppression lists, retention rules, and outbound-call restrictions. Production policy must be reviewed for every region in which calls operate.

Test natural barge-in so the agent stops speaking when the caller begins. Tune endpointing for different accents and environments, handle long silence without looping, and confirm recovery from an unavailable tool or synthesis request.

Use short acknowledgements before slow tools, set a maximum call duration, and provide a deterministic fallback when the AI cannot hear, understand, or complete the task.

Warm transfer should send the caller, verified identity, reason, conversation summary, completed authentication steps, and tool state to the receiving team. Confirm what happens if no human answers.

Test transfer during speech, after a failed tool, on explicit request, and during elevated sentiment. Record the transfer event and final disposition in the shared timeline.

Validate browser permission, inbound and outbound calls, caller ID, interruption, silence, tool use, transfer, hangup, provider errors, transcript delivery, recording policy, cost attribution, and post-call analytics.

Launch with sampled human review and dashboards for latency, containment, transfer, tool failure, outcome, opt-out, and cost per completed outcome. Natural speech is only one part of production readiness.