Building
AI call centers in 2026: what they handle, what they cost, and the rules for AI outbound calls
What an AI call center handles, what AI call center software costs per minute and seat, and the US rules for AI outbound calls: consent, hours, disclosure.
In short
An AI call center, also called contact center AI, uses AI voice agents and other software to answer calls, help human agents, score calls and make outbound calls. AI call center software publishes AI voice usage at about $0.04–$0.75 a minute, with Amazon Connect including AI in $0.038 a minute, and seats at $75–$249 per agent a month; an inbound call handled by people averages $7.20. The savings depend on how many calls the AI finishes: on the IRS individual tax line, about 62% of callers sent to its voice bot still transferred to a person. Answering calls with AI isn’t covered by the US robocall consent rules, but calling people with an AI voice needs their prior express consent, signed written consent for marketing calls, and calling-hour, do-not-call and opt-out rules.
Key takeaways
- An AI call center is four jobs: AI voice agents that answer, help for human agents, call scoring and AI outbound calls. Start with one.
- Published AI voice prices run about $0.04–$0.75 a minute; the share of calls the AI finishes decides the savings.
- Real deployments transfer most calls: 62% of callers sent to the IRS voice bot went to a person. Design the handoff first.
- Calling people with an AI voice needs their prior consent, and signed written consent for marketing calls.
- Keep people on the phones: only 20% of service leaders have cut staff for AI, and some that did have reversed it.
What is an AI call center?
An AI call center is a contact center where software does part of the phone work. You’ll also see it called call center AI, contact center AI or contact center automation. In practice it is four separate jobs, and most businesses start with one:
- AI voice agents that answer. An AI voice agent, or AI phone agent, takes the call, answers from your knowledge base, looks things up in your systems and transfers to a person when it can’t finish. For a single business line, that is an AI receptionist or AI answering service.
- Help for human agents. While a person handles the call, AI suggests answers, fills in the record and writes the summary afterwards.
- Call scoring. Every call transcribed and checked, instead of a supervisor sampling a few a week.
- AI outbound calls. Reminders, callbacks and follow-ups made with an AI voice. This is the part with the most legal rules.
It is not the older automated call center built on phone menus (“press 1 for billing”), which only routes calls. And it rarely replaces the team: several companies that tried have reversed course.
What does an AI call center handle well, and badly?
It does well with calls that repeat and have a written answer:
- Hours, prices, policies and how-to questions from your help articles
- Status lookups: an order, an appointment, a claim, a delivery
- Collecting the caller’s details and reason before a transfer
- After-hours and overflow calls, when everyone is busy
- Summaries and notes after every call
It struggles with exceptions, disputes, collections, upset callers and anything your documents don’t cover. The largest independent data on voice bots comes from the IRS. In the National Taxpayer Advocate’s 2025 report to Congress, about 62% of the 23.4 million callers sent to the voice bot on the individual tax line still transferred to a person, and the collections bot finished just 2.98% of calls on its own. The bot couldn’t pass what it had learned to the person who picked up, so callers had to start again.
The wider numbers agree. In a 2024 Gartner survey, only 14% of service issues were fully resolved in self-service, and only 36% even of issues customers called very simple. Gartner predicts that by 2029 AI agents will resolve 80% of common issues, but that is a forecast. In 2026, only 24% of service leaders could show positive financial returns across their AI projects.
Companies that cut staff have walked it back:
- In August 2025, Commonwealth Bank of Australia reversed the cut of 45 roles it had replaced with a voice bot, saying its assessment “did not adequately consider all relevant business considerations.”
- In May 2025, Klarna’s chief executive said it was hiring human agents again: “it’s so critical that you are clear to your customer that there will always be a human if you want.”
- Gartner found that only 20% of customer service leaders had actually reduced staff because of AI, and predicts half of the companies that did will rehire by 2027.
Helping human agents has better evidence. In a study of 5,172 support agents published in the Quarterly Journal of Economics, an AI assistant raised issues resolved per hour by 15% on average, with the biggest gains for newer agents. That was text chat at one software company, not phone calls, but it is the most rigorous result there is.
Test with your own callers. Speech recognition is less accurate for some voices: in a 2020 study of five commercial systems, the average word error rate was 0.35 for Black speakers against 0.19 for white speakers.
What does AI call center software cost?
Prices from each vendor’s own pricing page or documentation on October 7, 2026. Contact center platforms charge per seat, per minute or both, and several don’t publish what their AI voice agents cost.
| Platform | Seats or usage | AI voice agents |
|---|---|---|
| Amazon Connect | No seats; $0.038 a voice minute, plus phone charges | Included in the minute price; outbound campaigns $0.045 a minute |
| Google Conversational Agents | Usage only, billed on seconds of speech | $0.06 a minute of speech for scripted flows, $0.12 for generative |
| Genesys Cloud CX | $75–$240 per user a month, billed annually | Voice bots use one $1 token per 17 minutes, about $0.06 a minute (token price) |
| Microsoft Dynamics 365 Contact Center | $110 per user a month, paid yearly; calls billed separately | 10–75 credits a minute at $0.01 a credit: $0.10–$0.75 a minute |
| Twilio Flex | $150 per named user, or $1 per active user hour | ConversationRelay $0.07 a minute, plus call minutes |
| Dialpad | $80–$150 per user a month, billed annually | Sold as credits; price not published |
| Talkdesk | $85–$225 per user a month | Not published |
| NICE CXone | $110–$249 per agent a month | Not published |
| Five9 | $119–$159 per seat a month, 50-seat minimum | Not published |
People are the larger cost. The median US customer service representative earned $21.53 an hour in May 2025, about $0.36 a paid minute in wages alone, before benefits, idle time and supervision. ContactBabel’s survey of 207 US organizations puts the average inbound call at $7.20.
So the price per AI minute matters less than how many calls the AI actually finishes. Here is an illustration, not a quote: 10,000 calls a month, every call answered by AI first at $0.10 a minute, two minutes on the calls it finishes, one minute before it transfers the rest, and $7.20 for each call a person handles.
| AI finishes | AI usage | People | Total | Saved |
|---|---|---|---|---|
| No AI | — | $72,000 | $72,000 | — |
| 10% of calls | $1,100 | $64,800 | $65,900 | 8% |
| 30% of calls | $1,300 | $50,400 | $51,700 | 28% |
| 60% of calls | $1,600 | $28,800 | $30,400 | 58% |
Seats, the build and running it come on top. Measure what share of calls the AI finishes, and whether those callers call back within a week, before you plan around savings.
Are AI outbound calls legal?
Yes, to people who agreed to them. In the US the rules for AI outbound calls are much stricter than for answering. This is a summary, not legal advice.
- AI voices count as artificial voices. In February 2024 the FCC ruled that the robocall law’s limits on “artificial or prerecorded voice” calls cover AI-generated voices.
- Consent before the call. Calling a mobile phone with an AI voice needs the person’s prior express consent, unless it is an emergency. A marketing call to a mobile or home phone needs prior express written consent: a signed agreement that clearly authorizes calls from you. In practice, that rules out cold sales calls with an AI voice to consumers and mobile phones.
- What every call must say. Name the business at the start and give a number to call back. Marketing calls must also offer a way to opt out during the call.
- When and whom. Telemarketing calls only between 8 a.m. and 9 p.m. in the person’s time zone, and not to numbers on the National Do Not Call Registry.
- Opt-outs. People can revoke consent in any reasonable way, and you must honor it within ten business days. In October 2026 the FCC adopted changes that will let callers name one way to opt out; they take effect 30 days after publication and weren’t yet in force on October 7. An opt-out from marketing still stops all your marketing calls.
- Penalties. Anyone you call without consent can sue for $500 a call, up to $1,500 if willful. Since McLaughlin Chiropractic v. McKesson (2025), courts aren’t bound by the FCC’s reading of the law, so don’t rely on edge cases.
- States go further. Florida and Maryland require written consent for marketing calls that use automated dialing or recorded messages, and Oklahoma allows them only between 8 a.m. and 8 p.m. California requires prerecorded calls from automatic dialers to say if they use an artificial voice, and Utah requires a business to disclose AI when someone asks.
- Saying it’s AI. The FCC proposed in 2024 that AI calls disclose it at the start, but there was no final rule as of October 2026. Disclose anyway: it costs one sentence.
Enforcement has started with the worst cases. After AI-voice robocalls imitating President Biden in 2024, the carrier that transmitted them agreed to pay a $1 million penalty.
AI outbound calls work for calls people asked for: appointment reminders, callbacks they requested, delivery and order updates, renewals they opted into. Keep a record of each person’s consent, and treat a callback your AI makes as an outbound call.
Answering is simpler, which is why an AI answering service is the usual first step. The FCC has said “the TCPA’s requirements do not extend to technologies used to answer inbound calls,” though state recording laws still apply: in 2025 a federal court let a California wiretapping claim proceed against the company whose AI answered phone orders for Domino’s restaurants. The AI receptionist guide covers the inbound rules in full.
Should you build, buy, or switch it on?
Switch on your contact center AI if
- You already run Amazon Connect, Genesys or another contact center platform.
- Your routing, queues and reporting live there, and its AI covers your call reasons.
Buy an AI phone answering service if
- You have one business line and standard calls: hours, prices, messages, simple bookings. See the AI receptionist comparison.
Build your own AI voice agent if
- It must answer from your own documents and act in your own systems: orders, accounts, bookings, your CRM.
- Your platform’s AI agents are a “contact sales” add-on, or you don’t have a platform at all.
- You want custom transfer rules, warm handoffs with a summary, and your own call logs.
We build the third kind, connected to the phone line or queue you already have. An agent that answers from your FAQs and help articles goes live in three days, and one that also books or looks things up takes about a week. See AI voice agents and AI customer service agents.
How do you roll out an AI call center safely?
- Count your calls. List last month’s top ten call reasons and what a call costs you today.
- Pick one or two. Start with reasons that have a written answer. Leave disputes, collections and complaints with people.
- Design the handoff first. Transfer with a summary so nobody repeats themselves; that was the IRS bot’s biggest failure.
- Write the greeting. Name the business, say it’s an AI assistant, mention recording if you record, and tell callers they can ask for a person.
- Test by phone. Real questions, different accents, bad lines and background noise.
- Start on overflow and after hours. Watch the share of calls finished, transfers, repeat calls within a week and complaints.
- Add outbound last. Only to people with consent on record, inside calling hours, with an opt-out on every call.
- Cap and review. Set a monthly cost ceiling and read a sample of transcripts every week.
What next?
Pull last month’s call log, count the calls that asked the same few questions, and price them at what a call costs you now. That number tells you whether an AI call center is worth building. Voice AI for business pays off on volume: the same questions, many times a day. When you’re ready, send us your requirements or see how a three-day voice agent build works.
Sources
- National Taxpayer Advocate — 2025 Annual Report to Congress (voice bots)
- Gartner — Only 14% of issues resolved in self-service (2024)
- Gartner — Agentic AI to resolve 80% of common issues by 2029 (2025)
- Gartner — 24% of service leaders show positive AI returns (2026)
- Information Age — CBA reverses AI-driven job cuts (2025)
- Entrepreneur — Klarna hires human agents again (2025)
- Gartner — Half of companies that cut staff for AI will rehire (2026)
- Brynjolfsson, Li and Raymond — Generative AI at Work, QJE (2025)
- Koenecke et al. — Racial disparities in speech recognition, PNAS (2020)
- Amazon Connect — Pricing
- Amazon Connect — Pricing details
- Google Cloud — Conversational Agents pricing
- Genesys Cloud CX — Pricing
- Genesys — AI token usage by feature
- Genesys — Pricing and tokens
- Microsoft Dynamics 365 Contact Center — Pricing
- Microsoft — Copilot Studio credits per voice minute
- Microsoft Azure — Copilot Studio pricing
- Twilio Flex — Pricing
- Twilio — US voice pricing
- Dialpad — Pricing
- Talkdesk — Pricing
- NICE CXone — Pricing
- Five9 — Pricing
- US Bureau of Labor Statistics — Customer service representatives
- ContactBabel — US contact center decision-maker’s guide
- FCC 24-17 — AI-generated voices under the TCPA
- 47 CFR 64.1200 — Delivery restrictions
- FCC 26-67 — Consent revocation changes (2026)
- 47 U.S.C. 227 — Telephone Consumer Protection Act
- McLaughlin Chiropractic v. McKesson (2025)
- Florida Statutes 501.059 — Telephone solicitation
- Maryland Commercial Law 14-4502 — Stop the Spam Calls Act
- Oklahoma HB 3168 — Telephone Solicitation Act
- California Public Utilities Code 2874 — Artificial voice disclosure
- Perkins Coie via JD Supra — Utah AI disclosure laws (2025)
- FCC 24-84 — Proposed rules on AI-generated calls
- FCC — Lingo Telecom $1 million settlement (2024)
- Taylor v. ConverseNow — Order on motion to dismiss (2025)
- California AB 1609 — Customer service chatbots