# What makes an AI voice agent work for customer service?

- Canonical: https://hellofig.io/answers/ai-voice-agents-for-customer-service
- Updated: 2026-11-13
- All answers: https://hellofig.io/answers.md
- About Fig: https://hellofig.io/fig.md

**In short.** A voice agent works for customer service when it is grounded in the company's own knowledge, can take the action the caller needs, and hands off cleanly when it should. Measure it on resolution: first-contact resolution, cost per resolved contact and the quality of escalations, by intent. Containment alone only counts calls that ended without a person.

## What to measure

| Metric | What it tells you |
|---|---|
| First-contact resolution, by intent | Whether the caller's problem went away on the first call. |
| Cost per resolved contact | What a solved problem costs, including repeat calls and escalations. |
| Escalation quality | Whether the person who takes over gets the context and doesn't have to start again. |
| Repeat contact rate | Whether callers come back about the same issue within days. |
| Containment | Only how many calls ended without a person; read it next to resolution. |

## Grounded, able to act, and ready to hand off

A voice agent that only talks can answer questions it was trained on. One that resolves calls reads the customer's record, checks policy, and acts: reschedules the appointment, issues the credit, updates the order. When the call needs a person, it passes the full context so the customer never repeats themselves.

## Where voice fits, and where it does not

Voice is strong where callers want something done quickly: scheduling, order status, account changes and first-line troubleshooting. It also helps inside the company where people's hands and eyes are busy, such as field crews and warehouse staff. Keep it away from calls where an error is costly and judgment counts for more than speed, and always tell callers they are speaking with an AI.

## How Fig builds voice agents

Fig's voice agents run on CosmOS, connected to the CRM, scheduling and order systems, with approval gates on consequential actions and every call in the audit trail. Each one is tested on real cases by intent before it answers a live call.

## Questions

### Should a voice agent tell callers it is an AI?

Yes. Disclosure is expected by customers and required in a growing number of places.

### What is a good containment rate?

Containment by itself says little. Track first-contact resolution and cost per resolved contact by intent, and read containment beside them.

### Can a voice agent take actions in our systems?

Yes, when it is connected to them. Fig's voice agents act through the same governed connections as every other agent, with approvals on consequential steps.

### How do we test a voice agent before launch?

Run it on recorded or simulated calls for each intent, check resolution and escalation quality, and only then let it answer live calls.