# Call center automation: what to automate first Source: https://dialmilo.com/hub/call-center-automation Updated: 2026-10-23 Call center automation without the hype: which inbound and outbound calls to hand to AI first, what customers object to, the costs, and a 30-day rollout plan. Call center automation works when you automate calls, not people: pick the high-volume, predictable, low-risk call types, hand those to an AI voice agent, and keep humans one step away for everything else. Most centres start with inbound FAQs and forget the outbound half (reminders, follow-ups, callbacks, renewals), which is often easier to automate because you control the script and the timing. Say it is an AI, make reaching a person easy, and judge it by resolved outcomes, not deflection. Ask ten contact centre leaders what "automation" means and you will get ten answers: IVR menus, chatbots, workforce scheduling, after-call summaries, dialers, AI voice agents. All of them are automation. Only some of them change what a customer experiences on the phone. This guide is about that part: which calls a modern AI voice agent should take, which it should not, and how to roll it out without making your customers feel managed. ## What customers actually object to Start with the uncomfortable data, because it shapes everything else. In a Gartner survey of 5,728 customers carried out in December 2023, [64% said they would prefer companies did not use AI in customer service](https://insurance-canada.ca/2024/07/15/gartner-survey-ai-for-customer-service/), and 53% said they would consider switching to a competitor if they found out a company was going to. Their top concern was not robotic voices. It was that it would become harder to reach a person. Read that carefully. Customers are not objecting to automation; they are objecting to being trapped by it. Every bad IVR menu taught them that "automated" means "blocked from a human". The design rule that follows is simple: automation should remove waiting and repetition, never remove the exit. Gartner's own forecast points the same way. It predicts that by 2029 [agentic AI will resolve 80% of common customer service issues](https://www.cmswire.com/the-wire/gartner-predicts-agentic-ai-will-autonomously-resolve-80-of-common-customer-service-issues-without-human-intervention-by-2029) without human intervention, with a 30% cut in operating costs. Note the word "common". The other 20% are the calls that decide whether customers stay. ## The three tests for any call type Before automating a call type, put it through three questions. 1. **Volume.** Does it happen often enough that saving two minutes on each one matters? A call type you handle twelve times a month is not worth the setup. 2. **Structure.** Can you write down the three or four outcomes it usually ends in? "Confirmed / rescheduled / cancelled" is structured. "Customer is upset about something" is not. 3. **Risk.** What happens if the AI gets it slightly wrong? A reminder that needs a second try is cheap. A misunderstood hardship request is not. High volume, clear structure, low risk: automate it. Fail any one test: keep a person on it, or automate only the admin around it. ## Which calls to automate, and which to keep | Call type | Direction | Automate? | Why | | --- | --- | --- | --- | | Opening hours, location, pricing questions | Inbound | Yes | High volume, answers come from your own documents | | New bookings and rescheduling | Inbound | Yes, with calendar access | Structured, and the outcome is a calendar entry | | After-hours overflow | Inbound | Yes | The alternative is voicemail | | Appointment reminders and confirmations | Outbound | Yes | Clear outcomes, strong evidence it reduces no-shows | | New lead follow-up | Outbound | Yes | Speed is the whole game; humans take the qualified ones | | Requested callbacks | Outbound | Yes | Called at the time the customer asked for | | Renewals, win-back, reactivation | Outbound | Yes, with consent | One offer, exact terms, recorded outcome | | Payment reminders | Outbound | Partly | Reminders yes; hardship conversations to a person | | Complaints and escalations | Inbound | No | Judgement and empathy are the job | | Collections hardship, bereavement, vulnerability | Either | No | The cost of a mistake is too high | | Complex sales and negotiation | Either | No | Discovery is the work | ## The outbound half most centres skip Most call center automation projects start and stop with inbound, because that is where the queue is visible. The outbound side is often the easier win. On outbound you choose the script, the list, the timing and the single outcome you want. Nobody calls in with an unexpected problem. And the work is exactly the kind agents find draining: four hundred near-identical reminder calls, most of which reach voicemail. A typical outbound flow on Dial Milo looks like this. You upload a CSV, add a filtered group from your contacts, or let your CRM push people in through the campaign's "start a call" webhook. Milo calls each person in their own local time within the hours you set, retries the ones who do not answer, keeps anyone from being called twice within your cooldown across all campaigns, and stops at your daily or total budget. When a voicemail answers, Milo hangs up rather than leaving your brand on it and tries again later. When someone rings back, Milo recognises them and carries on the same conversation. Every call ends with a recorded outcome (accepted, declined, callback requested with a time, opted out, wrong person, no decision), plus a transcript, summary and recording. The callback list is what your agents work from. That is the hybrid pattern that tends to last: AI makes the first touch to everyone, people take the conversations that need people. We compare it with the older approach in [predictive dialer vs AI voice agent](/hub/predictive-dialer-vs-ai-voice-agent). ## What it costs AI voice automation is billed per minute, not per seat, so cost follows volume. - **Inbound:** about 12¢ a minute on a US number, billed per second. - **Outbound:** the phone line, priced for the exact number dialled and your caller ID country and billed per started minute, plus 9.57¢ per minute of AI talk time. Roughly 14–20¢ a minute all-in for US, UK and Dutch numbers. Calls that never connect have no minute charge. **Worked example (illustrative figures).** Say a 12-seat centre makes 6,000 reminder calls a month to US numbers. Assume 2,400 are answered and last 90 seconds, and 2,000 reach voicemail. - Answered: the line bills 2 started minutes each, 2,400 × 2 × 4¢ = $192; the AI bills 1.5 minutes each, 2,400 × 1.5 × 9.57¢ = $344.52. - Voicemail: budget about 5¢ per attempt, 2,000 × 5¢ = $100. - Total: about **$637 a month**. Done by hand, the same 6,000 dials at, say, 90 seconds of agent time each (dialling, ringing, talking, logging) is 150 agent hours. Put your own loaded hourly cost against that. The more useful question is what those 150 hours could be spent on instead. Check any destination on the [rates page](/rates), and plan an outbound list with the [spend calculator](/help/using-the-spend-calculator). ## A 30-day call center automation rollout plan Choose the call type that passes all three tests. Record today's numbers: volume, handle time, outcomes, and how often customers ask for a person. Turn your best agent's talk track into a short script. Put facts that change (prices, policies, opening hours) in uploaded documents or a connected Notion database, not the script. See [writing a script](/help/writing-your-script). Have three agents call the AI as difficult customers. Hear Milo in the browser demo and fix every answer that would make you wince. A slice of real traffic, with a low daily budget and Slack alerts on. Read twenty transcripts a day. Note every moment a customer wanted a person, and why. Same metrics as your baseline. If resolved outcomes held up and complaints did not rise, widen it. If not, fix the script or narrow the scope before you scale. ## Where the AI gets its answers An AI agent is only as good as what it is allowed to say. Keep three things apart: - **The script:** stable behaviour. Who the agent is, the reason for the call, how to handle the five most common situations, when to offer a callback. - **The knowledge:** facts that change. Prices, policies, opening hours, service areas. Upload documents or connect Notion databases so Milo answers from your own pages, and update them there rather than rewriting the script. - **The offer:** on outbound, the exact offer and terms. Milo presents it once, never invents or changes terms, and only explains conditions when asked. Most of the bad AI calls you will hear in testing come from mixing these up: a price buried in the script that went out of date, or an instruction so vague the agent fills the gap itself. ## The failure modes to test for Before going live, try to break it. These are the problems that turn up most often: - **Talking over people.** A "yes" in the middle of the caller's sentence should not restart the pitch. - **Silence on pickup.** On outbound, a pause of more than a second after the person answers loses them. - **Voicemail treated as a person.** Carrier detection misses some greetings; the agent needs to listen for them too. - **The wrong language.** If your customers speak Dutch or German, pass the language per person and check the agent switches when they do. - **No way out.** Ask for a person in five different ways and check every one works. ## How to measure it honestly Deflection rate (calls that never reached a human) is the metric vendors like and customers hate, because a customer who gives up counts as deflected. Measure these instead: - **Resolved outcomes:** bookings made, reminders confirmed, callbacks scheduled, offers accepted. - **Human-request rate:** how often callers ask for a person, and whether they get one. - **Repeat contact:** did the same customer call again within a week about the same thing? - **Opt-outs and complaints:** on outbound especially, a rising opt-out rate means the wrong list or the wrong message. - **Agent time freed:** hours moved from repetitive calls to the conversations that need people. ## Disclosure and consent Tell people they are talking to an AI. Dial Milo says so at the start of every call by default, and for good reason beyond honesty: customers who find out later feel tricked, which is precisely what the Gartner data warns about. Outbound calls also carry legal weight. In the US the FCC ruled on 8 February 2024 that [AI-generated voices are "artificial"](https://docs.fcc.gov/public/attachments/DOC-400393A1.pdf) under the TCPA, so automated outbound calls need the consent that applies to prerecorded calls. In the UK, the ICO requires specific consent for automated marketing calls. Our guide to [AI outbound calling](/hub/ai-outbound-calling) covers the rules in more detail. This isn't legal advice. The summary reflects US and UK rules as of October 2026. Take advice on your own call types and the consent you hold. ## Start small, on purpose The centres that get the most out of automation are rarely the ones with the biggest launch. They pick one dull, high-volume call, prove it in a month, and then move to the next. For the inbound side, start with [what an AI receptionist is](/hub/what-is-an-ai-receptionist). For outbound, pick your reminder list and [create an account](/signup). There is no subscription, and you can hear Milo call you before you pay anything. ## Frequently asked **What is call center automation?** It is using software, increasingly AI voice agents, to handle calls and call-related work that human agents used to do by hand. That covers inbound calls such as routine questions and bookings, outbound calls such as reminders and follow-ups, and the admin around every call: notes, dispositions, CRM updates and callbacks. **Which calls should a call center automate first?** The ones with high volume, a predictable structure and low risk if something goes slightly wrong. Appointment reminders, order and booking updates, callback scheduling, simple qualification and frequently asked questions are typical first candidates. Leave complaints, hardship, negotiation and anything emotionally charged with people. **Do customers mind talking to an AI on the phone?** Many say they would rather companies did not use AI in customer service. In a Gartner survey of 5,728 customers, 64% said so, and their top concern was that it would become harder to reach a person. The practical answer is to automate calls customers find tedious anyway, say it is an AI, and always make it easy to reach a human. **How much does call center automation cost?** AI voice agents are generally billed per minute rather than per seat. Dial Milo charges about 12 cents a minute for inbound calls on a US number, and outbound calls are the phone line price plus 9.57 cents per AI talk minute, roughly 14 to 20 cents a minute for US, UK and Dutch numbers. There is no subscription. **Will call center automation replace agents?** It replaces tasks rather than whole jobs in most centres. Gartner predicts that by 2029 agentic AI will resolve 80% of common customer service issues without human help, but the remaining conversations are the hard ones. The realistic goal is agents spending their hours on calls that need judgement instead of repeating reminders.