Call center automation is software that handles parts of customer contacts without a person: routing calls, answering the common questions by voice or chat, drafting replies for agents, writing the after-call notes and sorting tickets. In 2026 the stack runs from 1990s touch-tone IVR to AI voice agents that talk back in under a second, and the design question is which contacts to automate, and where the hand-off to a human sits.
The forecasts pull in opposite directions. Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues, cutting operational costs by 30%. The same firm predicts none of the Fortune 500 will have fully eliminated human customer service by 2028, and that half of the companies planning big service headcount cuts will drop those plans by 2027. Both are probably right. Routine contacts go to machines; exceptions stay human.
Here’s the stack layer by layer, where people stay in the loop, the rules that apply, and the numbers to watch.
The 5 layers of call center automation
| Layer | What it handles well | Where it breaks |
|---|---|---|
| IVR and routing | Menu choices, identification, skills-based routing | Callers mashing 0 to escape |
| Chat and email agents | Order status, password resets, returns, FAQs | Multi-issue tickets, angry customers |
| AI voice agents | Bookings, status checks, simple changes, after-hours | Accents plus noise, policy exceptions |
| Agent assist | Suggested replies, knowledge lookups, summaries, after-call work | Agents ignoring bad suggestions |
| Ticket triage and back office | Classification, priority, routing, drafting | Edge cases outside the rules |
1. IVR and call routing
The oldest layer and still the most used. Touch-tone and speech menus identify the caller, capture the reason and route to a queue by skill, language and priority. Modern platforms fold this into the same console as everything else: Amazon Connect, for example, prices voice at $0.038 per minute with AI capabilities included in the channel price, with no seat licences.
The failure mode hasn’t changed in 30 years: a menu with 7 options and a caller who wants the eighth. Keep menus to 3 or 4 choices, always offer a path to a person, and feed the “zero-out” rate back into the design.
2. Chat and email agents
This is where resolution rates are highest and the pricing has shifted to outcomes. Intercom’s Fin claims an average 76% resolution rate across 12,000+ customers and charges only when it resolves. Zendesk’s AI agents are included in its plans with usage billed on successful resolutions, through resolution credits, in 80 languages. Salesforce’s Agentforce service agent runs on no resolution, no charge pricing, and Salesforce says it handles 85% of its own support queries. Sierra sells outcome-based pricing across chat, SMS, WhatsApp, email and voice.
Outcome pricing changes incentives. The vendor is motivated to count a contact as resolved. Define “resolved” in the contract (no repeat contact within 7 days, customer didn’t ask for a human) and audit a sample monthly.
3. AI voice agents
The newest layer, and the one most people mean by “AI call center” in 2026. Speech comes in, a language model decides what to say, speech goes out, fast enough to feel like a conversation.
The plumbing is now off the shelf. OpenAI’s Realtime API is built for barge-in, low first-audio latency, natural turn taking and realtime tool use, over WebRTC or WebSocket. Twilio’s ConversationRelay turns caller speech to text, sends it to your application over a WebSocket, and plays back your text as speech, with configurable interruption handling so the caller can cut the bot off. ElevenLabs’ agents support 90+ languages with sub-second responsiveness and connect to Twilio, Genesys, Vonage or any SIP system.
Purpose-built vendors publish bigger numbers. PolyAI says its largest deployments resolve 90% of calls end to end with no human agent, with voice-to-voice response under 300 ms. Those are its best cases, and the vendor says so.
Where voice agents break: overlapping speech, background noise, strong accents, callers who change topic mid-sentence, and anything the policy doesn’t cover. Say a caller asks to cancel, then asks about a refund for last month, then mentions a bereavement. That call belongs with a person, and the agent needs to know it within 2 turns.
4. Agent assist
The layer with the least hype and, in most centers we’ve seen, the fastest payback. The AI listens to the live call or reads the chat and surfaces the knowledge article, drafts the reply, fills the CRM fields and writes the wrap-up summary.
Salesforce’s Service Rep Assistant builds step-by-step plans from customer data and conversation history, and its Conversation Summaries write the issue and resolution notes for the rep to review. Amazon Connect’s AI teammates suggest answers and take on after-call tasks. Cresta’s Agentic Assist offers live guidance, knowledge and automation during conversations.
After-call work is the quiet win. Every minute an agent spends typing notes is a minute the queue waits. Summaries written by the model and approved by the agent cut that, and the agent still owns the record.
5. Ticket triage and back-office automation
The unglamorous layer: read every incoming email and form, classify it (order status, return, billing, complaint, spam), set priority, route it, and draft a first reply from the customer’s order data. Most of the volume in an ecommerce support inbox is 5 intents. A model that handles those 5 well, with a person approving anything that refunds or changes an order, removes most of the queue.
This is also the safest place to start, because email tolerates a 2-minute delay that voice doesn’t. We cover the broader office version in our guide to AI assistants for work, and the difference between rules and models in AI vs automation.
Where humans stay in the loop
Klarna is the case study everyone cites, and both halves of it matter.
In February 2024, a month after launch, Klarna said its AI assistant had handled 2.3 million conversations, two-thirds of its customer service chats, doing the work of 700 full-time agents, cutting resolution time from 11 minutes to under 2, and reducing repeat inquiries by 25%.
In May 2025, after quality complaints, the company started hiring human agents again. CEO Sebastian Siemiatkowski told Bloomberg the priority was now the quality of human support, and that there will always be a human if you want. Klarna called it a dual-track approach, scalable AI plus human support. The same month he told CMSWire the assistant was still handling about 1.3 million errands per month, the equivalent of about 800 people, and was now eating into level-two support.
So the AI kept the volume and the humans came back for the hard cases. That’s the design, and Gartner’s Kathy Ross puts the reason plainly: AI excels at routine and well-defined problems and often struggles with exceptions and high-risk scenarios.
Customers agree, loudly. In Gartner’s December 2023 survey of 5,728 consumers, 64% would prefer that companies didn’t use AI for customer service, and 53% would consider switching if a company used it. Their top worry: it gets harder to reach a person.
What “human in the loop” means in practice:
- Give a visible exit. “Agent” or pressing 0 works at every step and carries the transcript with it. The caller never repeats their order number.
- Put approval gates on money and account changes. Refunds, cancellations, address and payment changes are drafted by the agent, approved by a person, at least until you have 3 months of clean data.
- Set confidence thresholds. Below a set score, the bot stops guessing and transfers.
- Add sentiment and risk triggers. Bereavement, legal threats, medical issues, fraud claims: straight to a human.
- Sample the output. A person reads 2% of resolved contacts every week and scores them.
Gartner also warns that over 40% of agentic AI projects will be canceled by the end of 2027, citing cost, unclear value and weak risk controls, and names “agent washing”, the relabelling of old chatbots and RPA as agents. Ask any vendor what the agent decides on its own and what it only recommends.
The rules that apply
Two regulations reach into call center automation directly.
United States. On 8 February 2024 the FCC ruled that calls made with AI-generated voices are artificial under the Telephone Consumer Protection Act, effective immediately. For outbound calls that means the same prior-consent rules as any prerecorded message. It covers outbound calls, so your reminder and collections flows are in scope.
European Union. Article 50 of the AI Act, which applies from 2 August 2026, requires that AI systems interacting with people are designed so the person is informed they are interacting with an AI system, unless it’s obvious from context. Say it at the top of the call. Most customers prefer knowing anyway.
Metrics that matter
Pick 6 and report them before and after every change.
| Metric | Definition | Why it’s easy to game |
|---|---|---|
| Automated resolution rate | Contacts closed by the bot with no repeat contact in 7 days | Vendors may count “closed” as “resolved” |
| Containment rate | Share of calls that never reach a human | Includes callers who gave up |
| Transfer rate and context carry-over | Share handed to humans, and whether the transcript arrived | Rarely measured at all |
| Average handle time | Talk plus hold plus wrap-up per contact | Drops if agents rush the hard calls |
| First-contact resolution | Issues solved in one contact | Depends on how you count a contact |
| CSAT and repeat-contact rate | Survey score and share returning within 7 days | Survey only reaches people who stayed |
Two of these deserve a second look. Containment counts a caller who hung up in frustration as a success; pair it with repeat-contact rate. And when a vendor charges per resolution (Fin, Zendesk, Agentforce, Sierra all do some version of it), the vendor’s definition of resolved becomes your bill. Write your own definition into the contract.
How to start without a transformation programme
- Pull 3 months of contacts and tag the top 10 intents. In most centers, 5 of them are 60% of volume.
- Automate the email and chat versions of those 5 first. Voice comes after you’ve seen the failure modes in text.
- Run the agent in shadow mode for 2 weeks: it drafts, humans send, you compare.
- Switch on auto-send for the lowest-risk intent only (order status, opening hours). Keep refunds and account changes behind approval.
- Review 2% of resolved contacts weekly. Expand one intent at a time.
- Before any outbound AI voice, check consent against the FCC ruling and add the disclosure the EU AI Act requires.
Our list of AI implementation questions to ask is a decent checklist for the vendor conversation.
Where AutomationHero fits
We’re the email layer, mostly. Our Customer inbox agent reads an ecommerce or SaaS support inbox, pulls the order or account data, and drafts order-status, shipping and return replies that your team approves before they send. That takes the 5 routine intents out of the queue your agents (and your phone line) would otherwise absorb. We don’t do voice. We install in one session and run it for a monthly fee; the catalog is here.
FAQ
What is call center automation?
Software that handles parts of customer contacts without a person: routing calls, answering common questions by voice or chat, suggesting replies and writing notes for human agents, and classifying and routing tickets. Humans handle exceptions, complaints and anything that moves money.
Can AI replace call center agents?
For routine, well-defined contacts, largely yes. Gartner predicts 80% of common issues resolved autonomously by 2029, and also that no Fortune 500 company will have removed human service by 2028. Klarna’s AI handles about 1.3 million contacts a month and the company still hired humans back in 2025 for the hard cases.
How much does an AI voice agent cost?
Platform pricing is usage-based: Amazon Connect lists voice at $0.038 per minute with AI included; chat and email agents from Intercom, Zendesk and Salesforce charge per resolution. Specialist voice vendors quote per deal. Model and telephony costs come on top if you build your own on the OpenAI Realtime API or Twilio.
Do I have to tell customers they’re talking to an AI?
In the EU, yes, from 2 August 2026 under Article 50 of the AI Act, unless it’s obvious. In the US, the FCC’s February 2024 ruling treats AI-generated voices as artificial under the TCPA, so outbound AI calls need the same consent as prerecorded ones. Disclosing up front is good practice everywhere.
What should I automate first?
The 5 highest-volume, lowest-risk intents, on email and chat before voice: order status, shipping updates, password resets, opening hours, return instructions. Keep refunds, cancellations and account changes behind a human approval until you have months of clean data.