Deploy an AI Voice Agent

AI Voice Agents are best for bounded conversations. Pick one concrete use case before you build. Common production patterns:

• Inbound qualification — greet, capture intent, route to a human rep.
• Outbound appointment confirmation — call to confirm / reschedule an existing booking.
• Payment reminder — notify a customer of an upcoming invoice, take basic questions, hand off to billing on request.

Avoid unbounded “talk to our AI about anything” deployments — they're harder to guardrail and higher-risk for regulatory exposure.

Open AI → Voice Agents → New agent. Pick a voice (10+ preset voices across US, UK, IN, AU accents), set a name, and choose a personality preset (Professional, Friendly, Warm).

Under Disclosure, choose when the agent states it's AI-powered. The FCC 2024 ruling and EU AI Act Article 50 both require this disclosure — Adoptiv defaults to “at call start”. Do not disable.

Adoptiv uses a “goal + guardrails + tools” model — you describe what success looks like, list what the agent must never do, and attach the CRM tools it can call.

Example goal: “Greet the caller, confirm their identity from the CRM, ask what they need, and either answer from the knowledge base or hand off to a human rep.”

Example guardrails:

• Never quote prices. Defer to a human rep for pricing.
• Never confirm, guess, or speculate on account information that isn't in the CRM tool response.
• If the caller asks to speak to a human, transfer immediately.
• If the caller says “stop calling me” or equivalent, record the revocation and end the call.

Under Tools, give the agent read access to CRM tools it needs. Common ones:

• lookup_contact(phone) — match the caller to a contact record
• get_recent_activity(contact_id) — last 5 calls / SMS / emails
• get_deal(contact_id) — current deal stage, owner, renewal date
• create_task(contact_id, text) — schedule a follow-up for the human rep
• transfer_to_human(reason) — warm-transfer to the queue matching the reason

Each tool runs under the agent's service account — permissions are scoped through RBAC, so the agent can never see data a human agent in the same role couldn't.

Under Escalation, define when the agent must hand off. Conservative defaults:

• Caller explicitly asks for a human (“agent”, “supervisor”, “representative”, “speak to a person”).
• Sentiment drops below -0.4 for two consecutive turns.
• Caller uses profanity or sounds distressed.
• Caller brings up a regulated topic the agent can't handle (medical, legal, billing disputes).
• Caller asks a factual question the agent can't answer confidently from a CRM tool or knowledge-base hit.

Click Simulate. A browser-based chat lets you role-play caller scenarios without spending minutes or dialing real numbers. Adoptiv ships with 20+ red-team prompts (angry caller, jailbreak attempts, regulatory tricks) — run them all before going live.

You can also upload a CSV of test scenarios and run them as a batch; Adoptiv scores each on goal-success, guardrail adherence, and escalation timing.

Go to Telephony → Numbers, pick the number you want the agent to answer on, and under Inbound routing choose AI Voice Agent → [your agent].

Start with a soft launch — route only 10% of inbound traffic to the agent for the first week, with the rest going to humans. Review the transcripts daily. Dial up the percentage once you're confident.

The AI → Voice Agents → Analytics dashboard shows, per agent: total calls, goal-success rate, average handle time, escalation rate, top escalation reasons, and every call's transcript + recording + AI summary.

Treat your agent like a new hire. Read 5 transcripts per day for the first month. Every time you find a failure mode, add a guardrail or a tool. The script is code — version it and change-manage it.