What an AI voice agent does on a live call

One call, five steps

A RUNAI VOICE agent handles a call in five steps, between the caller’s question and the answer they hear:

1. Listen (streaming speech): Any of 70+ languages, word by word, accents and code-switching included.
2. Recall (memory): Who’s calling, what happened last time, what they prefer.
3. Reason (grounded): Answers only from your approved knowledge — never from imagination.
4. Act (your systems): Books, checks, updates — calendar, CRM, order status, ticketing.
5. Speak (your brand voice): A designed voice, in about a second — not a robot reading a script.

The last step is the one a caller notices first. Measured on Run AI’s own production, the time to the first spoken word is <1.2s. Every step has its own budget: “Every hop has a millisecond budget, measured on every call — not a marketing claim.”

Where the steps run matters too. In the Voice page’s words: “Sandboxed · PII-scoped. Reasoning and actions run in isolated, permission-scoped environments. Your customer data never trains public models.”

What the agent is not allowed to do

On a live call there is no edit button, so the limits are designed before the first call. In the Voice page’s words: “The headline failures of voice AI share one root cause: agents deployed without guardrails and left without an operator. We engineered against every one of them.”

• It does not answer from imagination. “Answers come only from your approved knowledge base. On uncertainty, the agent says so and transfers — it throws an error instead of a wrong answer.”
• It does not improvise policy. “Every intent the agent may handle is explicitly defined — and everything outside that scope routes to a human. No improvised policy, no invented discounts.”

When it hands the call to a person

The handover is part of the design, not a failure of it: “Any caller can reach a human in one utterance — warm handoff, full context attached. AI takes the after-hours, the overflow and the routine. People keep the moments that need people.”

Callers don’t hate AI. They hate being trapped by it. Design the exit ramp well, and the agent becomes the fastest route to a great answer — not a wall between your customer and help.

The jobs a voice agent does on a phone line

• 24/7 service desk: Support, order status, FAQs, troubleshooting — resolved instantly, escalated warmly, logged completely.
• Speed-to-lead agent: Calls every new lead back in seconds, qualifies, and books the meeting.
• Reception & booking: Every call answered. Appointments booked, rescheduled and confirmed straight into your calendar and systems.

Start with one role. Every deployment is a single, well-defined job with its own success metric — not a vague “AI transformation.”

How it is tested before a caller hears it

A RUNAI VOICE agent passes six checks, before launch and after it:

1. Eval gauntlet: Every release passes a written quality bar before a caller ever hears it.
2. Red-team calls: We attack our own agents — trolling, jailbreaks, edge cases — before the public can.
3. Latency budget: Every hop has a millisecond budget, measured on every call — not a marketing claim.
4. Call QA: Conversations reviewed continuously; failures become fixes within days.
5. Drift watch: Agents degrade without care. Ours are monitored, tuned and re-evaluated weekly.
6. Staged rollout: New behavior reaches a small share of calls before it reaches all of them.

A typical deployment of one voice role goes live in ~4 weeks, week by week:

• Wk 1, discover: The role, the knowledge, the systems, the guardrails — scoped and signed.
• Wk 2, build: Persona and voice designed; knowledge grounded; calendar, CRM and phone lines connected.
• Wk 3, rehearse: Eval gauntlet and red-team calls. You hear it, test it, and sign off on every behavior.
• Wk 4, go live: Staged rollout under live monitoring — a small share of calls before all of them.
• Always, operate: Call QA, tuning, drift watch, monthly reporting. The agent compounds; a hire churns.

Does the caller know it is AI?

Where law or ethics require it, yes. In the Voice page’s words: “AI-voice disclosure rules are live across your major markets — US telemarketing law treats AI voices as regulated artificial calls, the EU requires people to know when they’re talking to AI, and synthetic-media labeling is law in India. We build compliance into the runtime, not the fine print.”

• Disclosure: “Callers are told they’re speaking with AI wherever law or ethics require it — configured per market, logged per call.”
• Data: “Client data never trains public models. PII-scoped, sandboxed, access-controlled — auditable end to end.”

When a voice agent is the wrong choice

• When most calls need a fresh judgement rather than an answer the business has already approved. An agent that answers only from approved knowledge has little to say on those calls, and it will hand most of them over.
• When nobody will own the knowledge it answers from. The agent is only as current as that knowledge; if prices, hours or policies change and the knowledge does not, the answers go stale.
• When the job has no clear edge. Each deployment is one defined role with its own success metric; “answer everything” is not a role.

If one role does fit, the Voice page shows the whole call, and a conversation starts with your call volume.

Questions

What happens when the agent doesn’t know the answer?

Answers come only from your approved knowledge base. On uncertainty, the agent says so and transfers — it throws an error instead of a wrong answer. No improvised policy, no invented discounts.

Can a caller always reach a human?

Any caller can reach a human in one utterance — warm handoff, full context attached. AI takes the after-hours, the overflow and the routine. People keep the moments that need people.

Do callers know they’re talking to AI?

Callers are told they’re speaking with AI wherever law or ethics require it — configured per market, logged per call. In regulated sectors the agent never forgets the script it’s required to say.

Who runs it after launch?

We do. Most voice pilots die after the demo because nobody operates them, so running the agent is our product, not your problem: call QA, tuning, drift watch, monthly reporting.