Can an AI agent answer your inbound support calls?

Can an AI agent do phone support?

Yes, for routine inbound calls, with a person always one transfer away. An AI phone agent for customer service can answer the call, work out why the customer is calling, answer questions from your help content, take bookings, give status updates and send everything else to your team with a summary. Organizations run this in production today. PolyAI reports that its agent at PG&E, a utility receiving about 16 million calls a year, reached 67% overall containment. It also reports that a UK pub restaurant chain has 70% of calls fully handled across 140 sites, and that a retail bank resolves 30% of calls with its voice assistant. These are the vendor's own numbers, they measure different things, and they come from different call mixes.

The difference from chat is that nobody can approve an answer before the caller hears it. Your controls sit before the call (what the agent may say and do), during it (when it must transfer) and after it (reviewing recordings). Treat it as a supervised hire.

This page covers live inbound calls. For written tickets see tier-1 support and ticket triage; for the role as a whole see customer service representative.

What the agent does and what stays with you

Agent

5of 10

• Answer the call, greet the caller and identify the reason for calling
• Answer common questions from approved help content
• Send a link or text so the caller can finish online
• Route or transfer the call to the right person with context
• Record the call outcome and summary in the help desk

Agent with review

3of 10

• Take, change or cancel bookings and appointments
• Verify the caller's identity before discussing an account
• Give account-specific status, such as an outage, order or bill

Human

2of 10

• Issue refunds, credits or bill adjustments
• Handle complaints, distressed callers and safety-critical calls

The split follows one rule: the agent can do what a reviewer can check from a recording and undo afterwards, and a person does what moves money or needs judgment about a distressed customer. Answering from help content, routing and logging are checkable by sampling. Bookings, identity checks and account status depend on your systems being right, so a person reviews the failures. Refunds, bill adjustments and complaints stay human. O*NET lists "resolve customers' service or billing complaints by performing activities such as exchanging merchandise, refunding money, or adjusting bills" as a core representative task, and none of the sources here shows a voice agent doing that unsupervised.

Some calls are too risky for an agent to finish at all. PolyAI's PG&E case says the agent handles calls on sensitive topics, including potential gas leaks, "where any wrong answer or error can have disastrous consequences." Decide which of your call types belong in that group and put them on the always-transfer list.

A worked example: one call

Assume, for this example, a home services company that books repair visits.

A customer calls at 7:40 p.m. The agent answers with the company's greeting, says it is an automated assistant and that the call is recorded. The caller says the technician never showed up today. The agent asks for the booking reference, confirms the caller's postcode and name against the booking, and sees the visit was marked "missed, technician delayed." It apologizes, offers the two open slots tomorrow, and the caller takes 9:00 a.m. The agent rebooks, sends a text confirmation and logs the call: "Missed visit, rebooked for tomorrow 9:00, customer asked about compensation."

Because compensation is on the always-transfer list, the agent says a team member will call back about it and creates a callback task instead of promising anything. The next morning a support agent reads the summary, listens to the 2-minute recording, and decides on a credit. The agent did the booking; a person decided the money.

What you need to provide

The agent touches personal data (names, phone numbers, addresses, account details) and keeps voice recordings and transcripts. Give it the least access that does the job:

• Call forwarding from your support number to the agent, and a separate escalation number that rings real people. Intercom's setup for Fin Voice, for example, asks for a main support line and a separate escalation line, and forwards the call there when Fin cannot resolve it.
• Help-center content that is current and owned. Intercom lists a knowledge base with "rich, up-to-date support content" as a prerequisite. If your articles are stale, fix them first; see knowledge-base maintenance.
• Read access to booking, order or account systems through a role that cannot refund, change payment details or close accounts.
• Written identity-check rules and an always-transfer list.
• Approved disclosure and recording wording.

Recording and disclosure rules depend on where you and your callers are. Under U.S. federal law, 18 U.S.C. 2511(2)(d) says interception is not unlawful under that chapter where the person is a party to the communication or one party has given prior consent, unless it is done to commit a criminal or tortious act. State laws can add their own requirements. In the EU, Article 50 of the AI Act requires providers to design AI systems that interact directly with people so those people are informed they are talking to an AI system, unless that is obvious from the context; the information must be given at the latest at the first interaction. Ask counsel which obligations apply to your business and the places you take calls from. Our recommendation, not a legal requirement: say it in the greeting on every call.

Turn on recordings and transcripts so you can review the agent. Intercom notes that its recording setting is global for all phone calls, so check that your notice covers every line it applies to.

Job brief template

Use this as the brief for a vendor, an operator or an internal team setting up the agent. Replace the bracketed values with your own.

The always-transfer list and the 50-call weekly sample are the parts to keep. They tell you more than any containment figure a vendor quotes.

What it costs

The human figure is the 2025 national median annual wage for customer service representatives (SOC 43-4051), as published on O*NET OnLine from BLS data: $44,770 a year, or $21.53 an hour. It is wage only. Dividing the hourly wage by 60 gives about $0.36 per paid minute, but a representative is not on a call for every paid minute, so this is not a per-call-minute comparison.

The agent figures are published per-minute prices from two vendors. Amazon Connect Customer lists voice at $0.038 per minute, with standard telephony rates charged on top, and says it has no minimums. Retell AI's pay-as-you-go plan lists $0.07 to $0.31 per minute for voice agents, depending on the language model and voice, with no commitment and $10 in free credits. Intercom does not publish a Fin Voice price; its documentation says to ask your account manager, and that Intercom Phone usage charges apply to forwarded calls.

Two costs sit outside both numbers: the time your team spends reviewing calls and handling transfers, and the work of getting your help content ready. Neither has a reliable public figure, so measure both in your trial.

Risks and human checkpoints

The most damaging failure is a caller who cannot reach a person. Set up the transfer before anything else, and test it every day. Intercom's Fin Voice documentation says to forward the call to your escalation line when Fin cannot resolve it, but does not document when a phone call escalates. Its separate conversational Fin guidance, whose examples cover chat, Messenger and email, says Fin escalates directly when a customer clearly asks for a person and offers escalation when the customer is angry or repeats themselves for 3 rounds; that guidance does not verify how Fin Voice behaves on calls. Whatever tool you use, check its phone escalation defaults with test calls and change them if they add friction you do not want.

The quieter failure is a call marked resolved that was not. Intercom's conversational Fin guidance, again not specific to voice, warns that a customer saying they will wait for an update can be read as acceptance and close the conversation. Count repeat calls from the same number as a signal that the first call did not end the problem.

How to run a trial

Run the agent on a narrow slice for 4 weeks before you widen it.

1. Pick a slice: after-hours calls, or one queue such as bookings or order status. Keep your current team on everything else.
2. Before launch, place 10 scripted test calls covering each always-transfer type, an angry caller, a request for a person and a caller who fails identity checks. All must route correctly.
3. Check the greeting on the first day: disclosure, recording notice and the transfer option.
4. Each week, review 50 calls against the brief's acceptance criteria. Log every wrong answer, missed transfer and false "resolved".
5. Track repeat calls within 48 hours and transfers that dropped. A missed transfer matters more than ten unnecessary ones.
6. Time your reviewers and the team taking transfers. Those minutes decide whether the agent saves money.

1. 1

   Pick a slice

   After-hours calls, or one queue such as bookings or order status

2. 2

   Place 10 scripted test calls before launch

   Each always-transfer type, an angry caller, a request for a person and a caller who fails identity checks

3. 3

   Check the greeting on the first day

   Disclosure, recording notice and the transfer option

4. 4

   Each week, review 50 calls

   Against the brief's acceptance criteria; log wrong answers, missed transfers and false "resolved"

5. 5

   Track repeat calls within 48 hours

   And transfers that dropped; a missed transfer matters more than ten unnecessary ones

6. 6

   Time your reviewers and the team taking transfers

   Those minutes decide whether the agent saves money

If the criteria hold for two weeks in a row, add the next queue or extend the hours. For structuring the test, see how to run a paid trial with an AI agent; for writing the brief, how to write a job description for an AI agent; and for ongoing review, how to supervise an AI agent. More support jobs are on the customer support overview.

Sources

1. 1
   43-4051.00 - Customer Service Representatives

   O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026

2. 2
   18 U.S.C. 2511 - Interception and disclosure of wire, oral, or electronic communications prohibited

   U.S. Government Publishing Office (govinfo)PrimaryAccessed Nov 14, 2026

3. 3
   Article 50: Transparency obligations for providers and deployers of certain AI systems

   European Commission AI Act Service DeskPrimaryAccessed Nov 14, 2026

4. 4
   Integrating Fin Voice with Zoom Phone using call forwarding

   IntercomPrimaryAccessed Nov 14, 2026

5. 5
   Conversational Fin experience

   IntercomPrimaryAccessed Nov 14, 2026

6. 6
   How Pacific Gas and Electric saved 35,000 labor hours with PolyAI

   PolyAIVendorAccessed Nov 14, 2026

7. 7
   How a UK pub restaurant chain automates 70% of guest phone calls

   PolyAIVendorAccessed Nov 14, 2026

8. 8
   How a popular retail bank resolves 30% of calls with a customer-led voice assistant

   PolyAIVendorAccessed Nov 14, 2026

9. 9
   AI Phone Agent Pricing

   Retell AIVendorAccessed Nov 14, 2026

10. 10
    Amazon Connect Customer Pricing

    Amazon Web ServicesVendorAccessed Nov 14, 2026