Contact center automation is software that takes over work your agents and supervisors do by hand: answering and routing calls, resolving routine requests, helping agents mid-call, and writing up what happened afterward. Most of the savings come from automating the right calls in the right order, and most failed projects tried to automate the wrong ones first.
This guide is for contact center operations leads and the engineers who work with them. It maps the layers of automation, explains how to pick the first queue, and shows the cost and measurement math for voice agents. If you’ve already picked a queue and want a step-by-step pilot plan, read our AI call center pilot guide.
The five layers of contact center automation
“Automation” covers very different products. Separate them before comparing vendors, because they change different numbers.
| Layer | What it automates | Who it helps |
|---|---|---|
| Routing and self-service | Phone menus, intent detection, and conversational IVR that sends a caller to the right queue | Callers and the routing team |
| Voice agents | Answering the call and completing the request end to end, with a transfer when needed | Callers, and every agent who no longer takes that call |
| Agent assist | Live transcription, suggested answers, and knowledge lookup during a human-handled call | Human agents |
| After-call work | Summaries, dispositions, CRM notes, and quality scoring | Human agents and QA |
| Outbound and back office | Reminders, confirmations, callbacks, forecasting, and scheduling | Operations |
Agent assist and after-call work make each human-handled call cheaper. Routing makes the call reach the right person sooner. Voice agents are the only layer that removes the human-handled call entirely, which is why they carry most of the savings and most of the risk.
Call center automation vs contact center automation
A call center handles phone calls. A contact center adds chat, email, and messaging. The playbook is the same for both, but phone is the hardest channel. A chatbot can take three seconds to reply and nobody notices. On a call, a pause that long sounds like a dropped line. The audio is narrowband, callers talk over the agent, and an order number has to be heard correctly the first time. If you automate chat first, don’t expect the results to transfer to phone without new testing.
Which calls to automate first
Good first candidates share five traits:
- High volume, narrow goal. “Where’s my order?” has one goal. “I have a problem with my account” has dozens.
- An outcome your backend can confirm. The order system shows the address changed, or the calendar shows the new appointment.
- Data reachable through an API. If agents copy information between three screens, the automation will need the same access.
- A cheap or catchable mistake. Reading back a delivery date is low risk. Issuing a refund is not, unless your backend enforces the rules.
- A clean way out. The caller can reach a person, and the person sees what already happened.
A rough rule: if your agents could handle the call half asleep at 4:55 on a Friday, it’s a candidate.
| Usually good first | Usually later |
|---|---|
| Order and delivery status | Billing disputes |
| Booking, moving, or canceling appointments | Complaints and escalations |
| Payment reminders and balance questions | Cancellations with retention offers |
| After-hours intake and message taking | Anything that needs regulated advice |
| Store hours, policies, and simple FAQs | Calls where the caller’s identity is hard to verify |
| Outbound confirmations | Multi-issue calls that need judgment |
Pull a month of call reasons from your telephony or CRM data, sort by volume, and cross off anything that fails the five traits. The top survivor is your first queue.
What a voice agent needs to resolve a call
A voice agent that resolves calls, rather than just answering them, needs every one of these pieces working together:
- Speech to text that hears names, numbers, and IDs on a phone line. Average accuracy hides the errors that matter; our word error rate guide explains why.
- Turn detection that knows when the caller has finished, so the agent neither interrupts nor leaves dead air. See voice activity detection.
- An LLM with tools that can look up the record, make the change, and transfer the call.
- Text to speech that starts speaking quickly and reads dates, amounts, and confirmation codes the way a person would.
- Telephony: phone numbers, your carrier or SIP trunk, and transfers to a human queue.
- Evaluation on completed calls, so you can see what happened on calls nobody listened to.
Cartesia Managed Agents package these pieces: Ink for speech to text, the LLM you choose, and Sonic for text to speech, with turn-taking built in. System tools handle ending calls, sending keypad tones, and transferring to a phone number or SIP URI. Webhook tools call your own HTTPS endpoints for lookups and updates. Numbers can come from Cartesia, from your Twilio account, or from your carrier over SIP. If you already run your own voice stack, you can use Sonic and Ink through the API and keep the rest.
What contact center automation costs
Voice automation usually has three running costs: the platform’s per-minute rate, telephony, and LLM usage. Integration work and ongoing call review come on top.
Here’s the math at Cartesia’s listed rates, checked October 6, 2026: Managed Agents calls cost $0.06 per minute, and Cartesia-provisioned numbers add $0.014 per minute.
| Example workload | Monthly figure |
|---|---|
| Calls automated | 10,000 |
| Average call length | 3 minutes |
| Agent minutes | 30,000 |
| Agent calling at $0.06/min | $1,800 |
| Telephony at $0.014/min (Cartesia numbers) | $420 |
| Total before LLM usage | $2,220 |
Calls on numbers imported from Twilio or connected over SIP bill telephony at your provider’s rate instead.
Compare that with your loaded cost per handled minute for the same call type, and count only calls the agent actually resolved. A call the agent fumbles and then transfers costs you both the automated minutes and the human ones. The honest number is cost per resolved call: total automation spend plus the human cost of transferred calls, divided by verified resolutions.
Measure resolution, not containment
Containment counts calls that never reached a person. It’s easy to report and easy to fool: an agent that hangs up on everyone has 100% containment. Track these instead:
- Verified resolution rate: the share of automated calls where your backend confirms the request was completed.
- Repeat contact within seven days for the same reason.
- Transfer rate and reason: the reasons tell you what to fix next.
- Abandonment during the automated part of the call.
- Incorrect actions, from wrong appointment times to updates on the wrong account.
- Response latency: how long callers wait after they stop talking.
- Satisfaction on automated calls compared with human-handled calls of the same type.
Managed Agents score every completed call with built-in metrics, including call success and time to first audio, and let you add LLM-judged metrics such as “was the caller’s issue resolved?” Spot-check transcripts against your backend anyway. A judge model can be wrong too.
What it looks like in production
Teams already run these workloads on Cartesia’s voice models:
- Assort Health runs specialty-specific voice agents for medical practices that handle scheduling calls against each practice’s rules. Assort reports 89% shorter wait times and 63% lower operational costs.
- Forethought uses Sonic in its customer-service voice agents and reports a 30% increase in deflection rate over its first nine months.
- Replicant automates contact center conversations for enterprises such as Fanatics. Its product team singled out spelling out emails, phone numbers, and order numbers as the place other voice providers struggled.
- SuperDial automates outbound calls to insurance payers, including navigating phone trees and waiting on hold, and reports a 3x reduction in operational costs.
These are each company’s own results on their own workloads. Use them as a sense of what’s possible, and measure your own queue before you promise anyone a number.
Roll out in stages
- Test before live traffic. Run the agent against scripted test calls built from real call reasons, including the awkward ones.
- Start after hours or on overflow, where the alternative is voicemail or a long hold.
- Take one queue during business hours with a fast path to a person.
- Add one call type at a time, and only after the last one holds its resolution rate for a few weeks.
Keep the old route reachable throughout. Callers forgive an agent that hands them to a person quickly. They don’t forgive one that won’t let them leave.
To try the pieces, build an agent in the Cartesia playground and point a test number at it. Our AI phone agent page covers numbers, transfers, and outbound calling.