# AI call center: what it handles and what it costs Source: https://dialmilo.com/hub/ai-call-center Updated: 2026-10-30 An AI call center for a small team: which calls AI can own, which need a person, real per-minute costs for inbound and outbound, and a 30-day rollout plan. For a small or mid-sized business, an "AI call center" isn't a building or a platform migration. It's a phone number answered by an AI that knows your documents and can book, transfer and take messages, plus an outbound queue that makes reminders, follow-ups and check-ins. The AI should own the routine calls and the after-hours calls; your people should own complaints, negotiation and anything that needs judgement. Priced per minute, inbound runs about 12¢ a minute on a US number and outbound roughly 14–20¢ a minute all in for US and UK numbers. The phrase "AI call center" is mostly used by enterprise vendors selling to contact centres with hundreds of seats. Their articles talk about orchestration layers, workforce management and omnichannel routing. If you run a team of five, or one, most of that doesn't apply. What does apply is a simpler question: which of the calls your business makes and takes could be handled well by an AI, and what would that cost? ## What an AI call center actually is Strip away the enterprise vocabulary and there are two jobs: - **Inbound.** Someone rings your number. The AI answers in one or two rings, greets them with your script, answers questions from the documents you uploaded, books into your calendar, takes a message, sends a follow-up text, or warm-transfers to a person. - **Outbound.** You have a list of people to call: reminders, confirmations, follow-ups, renewals, win-backs. The AI calls each one in their own local time, says it's an AI, has the conversation, and records an outcome you can act on. Dial Milo does both from one account. If you only need the inbound half, [what is an AI receptionist](/hub/what-is-an-ai-receptionist) is the simpler starting point. If outbound is the bigger job, the [AI outbound calling guide](/hub/ai-outbound-calling) covers it in depth. ## The honest split: what AI should own and what it shouldn't The research on AI in customer service has a finding that's more useful than any vendor's case study. A study of 5,179 customer support agents by Brynjolfsson, Li and Raymond, published as [NBER working paper 31161](https://www.nber.org/papers/w31161), found an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced, highly skilled agents. The AI was good at spreading the knowledge your best people already have. It added little where judgement was the job. That's a fair guide to what an AI voice agent should take on: | Call type | Who should handle it | Why | | --- | --- | --- | | Opening hours, location, prices, policies | AI | Answers live in your documents | | Booking, rescheduling, cancelling appointments | AI | Structured, with a calendar to check against | | Order or job status | AI, if the data is reachable | Routine lookup; transfer if it isn't | | Appointment and payment reminders | AI | Short, predictable, high volume | | Lead follow-up and qualification | AI, then a person to close | Speed matters; nuance comes later | | After-hours calls of every kind | AI first | Capture and triage beat voicemail | | Complaints with history | Person | Needs authority and context | | Negotiation, custom quotes, refunds outside policy | Person | Terms change in the conversation | | Distressed or vulnerable callers | Person | The AI should offer to transfer or take a message | The line isn't fixed. It moves as you write more of your know-how into documents. But if you start by giving the AI only the top half of this table, you'll get the benefit without the embarrassments. Dial Milo doesn't offer human agents of its own. When a call needs a person, it warm-transfers to you or a team member, or takes a detailed message. AI billing stops at the transfer. ## What it costs, compared with the alternative Dial Milo is pay as you go, with no subscription, no per-seat licence and top-ups from $10. - **Inbound** calls to a US number are about 12¢ a minute, billed per second. - **Outbound** calls are the phone line, priced for the exact number dialled and your caller ID and billed per started minute (US numbers from 4¢, a UK mobile from a UK number 8.5¢, a Netherlands mobile 10¢), plus AI at 9.57¢ per talk minute. Roughly 14–20¢ a minute all in for US, UK and Dutch numbers. Calls that don't connect have no minute charge. Full detail is on the [pricing page](/pricing). **Worked example (illustrative figures).** A regional home-services company with a three-person office takes 1,200 inbound calls a month and wants to add 500 outbound reminder calls. - **Inbound:** 1,200 calls × 3 minutes average = 3,600 minutes × 12¢ = **$432**. Some of those are transferred to the office; AI billing stops at the transfer, so this is an upper estimate for the AI part. - **Outbound reminders:** assume 50% answer, 1.5 minutes each, billed as 2 line minutes. Line: 250 × 2 × 4¢ = $20. AI: 250 × 1.5 × 9.57¢ = $35.89. Short voicemail connections, allow $6. About **$62**. - **Total:** roughly **$494 a month**, covering every hour of the week. For context, the US Bureau of Labor Statistics puts the [median pay for customer service representatives](https://www.bls.gov/ooh/office-and-administrative-support/customer-service-representatives.htm) at $21.53 an hour (May 2025 data). One full-time person at that rate costs about $3,700 a month in wages alone, before taxes, benefits and management, and covers roughly 40 of the week's 168 hours, one call at a time. That's not an argument for replacing the three people in the office. It's an argument for not asking them to answer the 9pm calls, the Saturday calls, and the hundredth "what time do you open?" of the week. ## The rules that apply to AI on the phone As of October 2026, three rules shape how an AI call center should behave: - **Disclose that it's an AI.** In the EU, [Article 50 of the AI Act](https://artificialintelligenceact.eu/article/50/) requires people to be informed they're interacting with an AI system, from 2 August 2026, unless it's obvious from the context. Dial Milo's outbound calls say it's an AI at the start by default. - **Outbound consent in the US.** On 8 February 2024 the [FCC ruled](https://docs.fcc.gov/public/attachments/DOC-400393A1.pdf) that AI-generated voices are "artificial" under the TCPA, so telemarketing calls with an AI voice need prior express written consent. Informational calls such as reminders are treated differently from sales calls. - **Outbound consent in the UK.** The ICO's [telephone marketing guidance](https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/guide-to-pecr/electronic-and-telephone-marketing/telephone-marketing/) says automated marketing calls need consent that "must specifically cover automated calls". This isn't legal advice. It's a summary of published guidance as of October 2026. Before any outbound sales campaign, check what your sign-up forms actually ask people to agree to. ## What makes the AI good at your calls The quality of an AI call center is decided less by the model than by what you give it. Four inputs matter: 1. **A short script.** Who it is, what your business does, how to handle the common call types, when to transfer. Ninety words is often enough. Don't put prices in it. 2. **Your documents.** Prices, policies, service areas, FAQs, in [uploaded knowledge docs](/help/uploading-knowledge-docs) or connected Notion pages. When a fact changes, you change the document, not the script. This is where the NBER finding bites: the AI can only spread knowledge that's written down. 3. **A calendar.** With Google Calendar connected, Milo checks real availability and books on the call rather than taking a "please call me back" message. 4. **Clear transfer rules.** Who gets which call, and what happens when nobody picks up. For outbound, add the offer and its exact terms (Milo presents them and never invents or changes them), calling hours, retries and a contact cooldown so nobody is called twice in a few days. ## A 30-day rollout Point your number at the AI outside office hours. Every call it takes is a call that would have gone to voicemail. Read every transcript. The transcripts show what callers ask that your documents don't answer. Write those answers down and upload them. This step improves quality more than any other. Let the AI answer when the office line is busy or after a few rings. Set transfer rules for the calls your team should take. Start with something low-risk and useful, such as appointment reminders or a check-in on customers whose usage dropped. Upload a list or trigger calls from your CRM by webhook. Track how many calls end resolved, how many are transferred, and what outbound outcomes look like. Move a call type across the line in the table above only when the transcripts say it's ready. ## Where AI call centers go wrong The failures are predictable, and almost none of them are about the voice sounding robotic. **Hiding the human.** Callers forgive an AI that can't help them. They don't forgive one that won't let them reach a person. Every call type should have an exit: a transfer, or a message with a named callback time. If nobody is available, say so plainly rather than looping. **Answering from memory instead of documents.** An AI that's allowed to improvise will produce a plausible price or policy that isn't yours. Milo answers from the script and documents you give it and says when it doesn't know. Test this on day one by asking something your documents don't cover; the right answer is "I don't have that, let me take your details". **Using outbound as a megaphone.** Calling the whole contact book because calls are cheap is how a number gets labelled as spam by phone networks. Dial Milo watches the last 20 finished calls and pauses a campaign by itself if none connected or most failed, but the better protection is calling people who expect to hear from you, within their local calling hours, with a cooldown between campaigns. **Never reading the transcripts.** The transcripts are the training material. Teams that read twenty a week in the first month end up with an AI that handles their calls well. Teams that don't end up with the generic version. ## What to measure - **Resolved without transfer**, by call type. Low is fine for complex types; low on "what are your hours?" means a missing document. - **Transfer reasons.** The AI's summary says why it handed off. Patterns tell you what to write down next. - **Outbound outcomes**: accepted, declined, callback requested, opted out, wrong person, no decision. The AI campaign report labels each person with what they need next. - **Cost per resolved call**, from the cost breakdown on each call rather than an estimate. An AI call center works best when it's treated as a new member of the team who has read everything but decided nothing. Give it the routine work and the hours nobody wants, keep the judgement with your people, and widen its job as the transcripts earn it. The automation side of this is covered further in [call center automation](/hub/call-center-automation), the outbound machinery in [automated calling systems](/hub/automated-calling-system), and one of the most valuable outbound uses in [churn prevention calls](/hub/churn-prevention-calls). ## Frequently asked **What is an AI call center?** An AI call center uses AI voice agents to answer and place phone calls that human agents would otherwise handle. The AI listens, understands what the caller wants, answers from your own documents, takes actions such as booking an appointment, and hands the call to a person when it needs judgement. For a small business it's usually a phone number answered by an AI plus an outbound calling queue, not a new building or a platform migration. **Can an AI call center replace human agents?** For routine calls, largely yes: opening hours, bookings, order status, reminders, confirmations and simple follow-ups. For complaints with history, negotiation, distressed callers and anything where a mistake is costly, no. The realistic model is the AI taking the routine volume and the out-of-hours calls, with a warm transfer to your team for the rest. **How much does an AI call center cost?** With pay-as-you-go pricing you pay per minute rather than per seat. On Dial Milo, inbound calls to a US number cost about 12 cents a minute, billed per second. Outbound calls are the phone line, from 4 cents a minute for US numbers, plus 9.57 cents per AI talk minute, so roughly 14 to 20 cents a minute all in for US and UK numbers. There's no subscription. **Do callers have to be told they're talking to an AI?** In the EU, Article 50 of the AI Act requires that people are informed they're interacting with an AI system, applying from 2 August 2026, unless it's obvious from the context. In the US, the FCC treats AI-generated voices as artificial voices under the TCPA for outbound calls. Telling callers up front is the safe default everywhere and it builds trust. This isn't legal advice. **How long does it take to set up an AI call center?** A basic inbound line can be answering the same day: buy or forward a number, write a short script, upload your FAQs and test it. A sensible rollout takes about a month, starting with after-hours calls, then daytime overflow, then the first outbound campaign, reading transcripts at each step.