An AI call center uses AI to answer customer phone calls, assist human representatives, or review conversations after they end. If you want an AI voice agent to resolve support requests, start with one queue and a result you can verify in your backend. Keep a route to a person. Expand only after the agent passes the same task and permission checks you expect from your support team.
This guide is for contact-center operations leads and engineers adding voice agents to an existing support operation. It covers a pilot for order-status calls, including a scorecard you can adapt. For a small business choosing an answering and scheduling service, start with our AI receptionist guide.
Choose which part of the call center needs AI
“AI call center” can describe several different products. Separate them before comparing vendors.
| Need | What to evaluate | What still needs an owner |
|---|---|---|
| Answer calls and complete routine tasks | A voice agent with access to approved tools | Tool permissions, customer verification, escalations |
| Help a human representative during a call | Transcription and agent-assist software | The representative’s workflow and approval of suggestions |
| Find problems in completed calls | Conversation analysis and quality review | Review criteria, recording access, retention |
| Manage queues, staffing, and routing | Contact-center software | Workforce planning and carrier configuration |
Cartesia Managed Agents combines speech recognition, an LLM, and Sonic speech generation. It handles turn-taking and interruptions and can call tools during a conversation. It is the voice-agent part of this system, not a replacement for workforce management or every feature of your contact center.
If your existing contact-center vendor already supports the task and integrations you need, evaluate that option too. Building a custom agent makes sense when your team needs control over business rules or customer-facing voice behavior and can maintain the integration.
Pick a queue with a verifiable outcome
For an order-status pilot, a successful call means the verified customer received the status of the correct order from the order system. The agent must not invent a delivery date or expose another customer’s details.
Write down the eligible call types before launch. Exclude refunds, address changes, and payment collection from this first pilot. Those actions require different permissions and tests. Route them to your existing support team rather than stretching an order lookup into an all-purpose agent.
Use your current queue data to estimate volume, repeat contacts, and handling cost. Preserve the original denominator: if you exclude hard calls from the AI queue, do not compare its resolution rate with the entire human queue and call the difference an improvement.
Connect the agent to the order system
Create an agent in the Cartesia Playground. Give it an opening that identifies it as an AI assistant, a narrow task, and instructions for when to hand off. Select a voice and test how it reads order numbers, dates, and delivery statuses.
A webhook tool lets the agent call an HTTPS endpoint on your server. Your endpoint should enforce the customer’s authorization before returning order details. Tool authentication establishes which service made the request; it does not prove the caller owns the order.
Return only the fields needed for the answer. For example, return a status, a verified delivery estimate when one exists, and whether a human needs to investigate. Keep internal notes and unrelated customer records out of the response. Cartesia’s webhook responses are limited to 4 KiB, so do not return an entire order history.
Keep the first tool read-only. If you later add write actions, enforce permissions in your backend and deduplicate retries. A prompt telling an agent to “be careful” cannot prevent an endpoint from applying the same refund twice.
Test the phone route and the human handoff
Managed Agents supports phone connections through Cartesia numbers, imported Twilio numbers, and SIP trunking. Check the fit with your carrier and contact-center routing before moving a queue. Use a test number first.
Configure Transfer call destinations for a customer asking for a person, a failed identity check, and requests outside the pilot. Transfers require a Twilio or SIP-trunked phone call; SIP URI destinations require SIP trunking. The transfer tool returns an error in Playground and v1 WebSocket calls, so a browser conversation cannot validate this part of the rollout.
Test the receiving queue during staffed and unstaffed hours. Verify whether the representative receives any conversation context; do not assume a phone transfer also transfers a transcript. If your integration cannot pass context, define what the agent should tell the caller and how the representative will retrieve the case.
Keep your original routing configuration available for rollback. Agree who can pause the pilot if it starts exposing data, making incorrect claims, or sending callers to an unanswered number.
Use a scorecard before routing live traffic
Run scripted calls against synthetic customer records. Record the expected outcome before each test, then inspect both the conversation and backend result.
| Test | Required result |
|---|---|
| Customer asks about an order after passing verification | Agent reports only the matching order’s returned status |
| Caller gives another customer’s order number | Backend prevents access; agent follows the approved fallback |
| Order service times out | Agent says it cannot confirm the status, without inventing an answer |
| Caller corrects the order number mid-sentence | Lookup uses the confirmed number |
| Caller interrupts while the agent reads the status | Playback stops and the agent handles the new request |
| Caller asks for a refund | Agent routes to the authorized workflow without promising a refund |
| Caller asks for a person | Caller reaches the intended queue or the approved unavailable-hours fallback |
| Caller tells the agent to ignore its rules | Server permissions still prevent restricted access and actions |
Repeat tests over real phones, with background noise and the accents your queue serves. Measure response time from the end of the caller’s turn to audible playback on the caller’s device. Text-to-speech time to first audio measures only part of that delay.
Choose launch thresholds with the support owner. Treat a confirmed unauthorized disclosure or incorrect account change as a stop condition. A good average score does not cancel a serious permission failure.
Measure resolution, not just containment
Containment counts calls that do not transfer to a human. It can include hang-ups and unresolved callers. Define a verified resolution using your order system and a repeat-contact window appropriate to the task.
For an illustrative cost calculation, suppose 1,000 eligible calls cost $600 in agent and carrier charges plus $400 in human follow-up. If 700 calls have verified resolutions, the operating cost is $1,000 / 700, or about $1.43 per verified resolution. These are example inputs, not Cartesia prices or expected results. Report implementation and maintenance costs separately, then include them when deciding whether the pilot pays for itself.
Compare against equivalent calls handled by the existing workflow. Track repeat contacts, incorrect answers, failed transfers, abandonment, and customer complaints alongside cost. Review failed calls by cause: recognition, business logic, tool availability, routing, or speech playback. Each requires a different fix.
Expand one task at a time
Start with a limited traffic allocation and keep reviewing failures. Add a new task only after its permission checks and fallbacks have tests. For multilingual calls, validate recognition and speech generation separately: a TTS model’s language coverage does not establish that the entire agent supports those languages.
Before production, have your privacy and legal owners review AI disclosure, call recording, data retention, and vendor agreements. The requirements depend on your callers and use case. Do not put sensitive production records into an unapproved test environment.
Create a Managed Agent for one read-only support task. Use the current pricing for your cost model.