What an AI call assistant can do, what it costs, and when not to use one

The short version

• An AI call assistant answers or places phone calls using speech recognition, a language model and a synthetic voice.
• AI call assistants suit calls where the request is predictable, the answer sits in a reachable system and a mistake is cheap to fix.
• Without connections to a calendar, CRM or ticketing system, an AI call assistant can only talk and take messages.
• A caller who asks for a person should always reach one, or be told plainly when someone will call back.
• Compare AI call assistants on cost per resolved call, because a low usage rate means little if most calls end in a transfer.

How does an AI call assistant work?

An AI call assistant runs a loop on every turn of the conversation. Speech recognition turns the caller's words into text, a language model decides what to say or do, and a synthetic voice speaks the reply. Telephony software connects that loop to an ordinary phone number.

The useful version has a fourth part: tools. These are connections that let it check a calendar, look up an order or open a ticket. Without them it can only talk. It can recite a knowledge base and take a message, which is not far from what an FAQ page and voicemail already do.

The phone is harder than a chat window for mechanical reasons. A pause nobody notices in chat sounds like a dropped line on a call. Noise, accents and people talking over the assistant all damage the transcript, and every later step depends on that transcript. A misheard name or postcode does not show up as a typo. It shows up as a wrong booking.

Which calls can an AI call assistant handle well?

An AI call assistant handles calls well when the request is predictable, the answer sits in a system it can reach, and a mistake is cheap to fix. Booking, routing and after-hours messages usually meet that test. Complaints and advice usually do not.

Most businesses have a mix, so the decision is rarely all or nothing. Pull a sample of recent calls and sort them by type before you look at any product. The sort tells you how many calls could be finished without a person, which is the number the whole business case rests on.

Which calls suit an AI call assistant?

When is an AI call assistant not worth using?

An AI call assistant is not worth using when call volume is low, when most calls are exceptions, or when the personal call is part of what customers are paying for. In those cases it adds a step and a bill without removing much work.

A few varied calls a day, answered by someone already in the room, do not justify the set-up and upkeep. Neither does a line where the assistant would transfer nearly every call, since the caller just waits longer. Cheaper fixes are often enough: an online booking link, a clear voicemail with a firm call-back time, or a human answering service for overflow.

What does an AI call assistant cost?

An AI call assistant usually costs a usage or plan fee, plus items the headline price leaves out, plus the work of connecting and maintaining it. No honest single figure exists, because the price depends on call length, volume and how much integration you need.

Pricing pages show why. Retell AI publishes a pay-as-you-go range, but its page does not say what the range assumes, some single components are priced above the top end, and items such as phone numbers are billed separately. Smith.ai publishes a list price for its Human-first Starter plan, with optional per-call and monthly add-ons on top. Both are vendors' own list prices, so read the live pages before you budget.

The usual comparison is a receptionist's wage. The US Bureau of Labor Statistics lists 2025 median pay for receptionists as $38,010 per year. That is US pay before employer costs, and a receptionist does far more than answer the phone. A straight comparison flatters the software, because the software takes over only part of the job.

The number to calculate is cost per resolved call. Add usage, phone numbers, integration upkeep and the staff time spent reviewing transcripts, then divide by the calls finished without a person. A low rate means little if most calls still end in a transfer.

Is an AI calling assistant different from an AI call assistant?

Not in any consistent way: vendors and searchers use both names for the same kind of software. The distinction that matters is direction. Inbound assistants answer calls people chose to make. Outbound assistants ring people who may not want to hear from them.

Inbound is the easier case, because the caller wants something and the assistant only has to provide it. Outbound starts with a harder problem: getting anyone to pick up. In Hiya's survey of 12K+ consumers across 6 countries, 14% said they immediately answer numbers they do not recognise. That is self-reported behaviour, not measured call data, and Hiya sells branded caller ID, so it has an interest in the finding.

Outbound also carries more legal risk. Rules on consent, disclosure and automated calling differ by country and sometimes by state, so take advice before the first call goes out. The outbound calls that make sense are ones people expect: a reminder for a booking they made, or a call-back they asked for. Cold calling with a synthetic voice is the use we would leave alone.

Will callers accept talking to an AI?

Some callers will accept it and some will not, and no survey can give you the split for your own callers. Only testing on your own phone line can. Acceptance also depends on the outcome: a caller who gets a booking done quickly judges it differently from one stuck in a loop.

Survey evidence is a starting point at best. Moneypenny's survey of 5,001 UK consumers, with fieldwork by Censuswide, records stated comfort, not how callers behaved on a real call. It is a UK sample and Moneypenny sells call answering, so read it with that in mind.

Tell callers at the start that they are speaking to an AI, and tell them how to reach a person. Callers who work it out halfway through have a reason to doubt everything said so far. Then watch behaviour: how many ask for a person straight away, how many hang up in the first few seconds, and how many ring back about the same thing.

How should an AI call assistant hand over to a person?

An AI call assistant should transfer on request, without argument, and pass on what it has already learned so the caller does not repeat it. A caller who says 'get me a person' should reach one, or be told plainly that nobody is available and when someone will call back.

Set the triggers before launch. An explicit request always transfers. So does the same misunderstanding happening twice, any topic on a restricted list, and any sign of distress. Write these as rules the system enforces. A language model can be talked out of an instruction, and a rule in code cannot.

Check who is on the other end of every transfer path, at every hour the assistant answers. A transfer to an empty line at lunchtime is worse than no assistant, because the caller has now spent time twice and still has no answer.

How do you know whether an AI call assistant is working?

You know by comparing it with a baseline from your own phone line, taken before launch. Record missed calls, call types and how calls ended. Without a baseline, any improvement is a guess.

Do not rely on the share of calls finished without a person, usually called containment. A caller who gives up and hangs up can count as contained. Pair it with repeat calls on the same issue, early hang-ups, requests for a person, and errors found by reading a sample of transcripts every week.

Vendor benchmarks are context, not targets. Invoca's lead conversion benchmarks draw on 70M+ phone calls and 600M+ conversation minutes across 10 industries. They are averages across Invoca's own customers, not a sample of small businesses, and Invoca sells call tracking software. Your own numbers will tell you more.

Should you buy an AI call assistant or build one?

Buy one first, unless your calls depend on systems or rules a packaged product cannot reach. A standard booking or message-taking line does not need custom engineering. We sell custom AI engineering, and that is still our advice for a standard phone line.

Custom work starts to make sense when the assistant has to act inside your own systems with real permission checks, when the data cannot go to a general-purpose vendor, or when the call is one step in a longer process. At that point the voice is a small part of the job. The hard part is the integration, the evaluation and what happens after the call ends.

What should you ask a vendor before buying an AI call assistant?

• •Which of our systems can it read from and write to, and what does each connection cost?
• •How does a caller reach a person at any point in the call, and what happens when nobody is free?
• •What happens to a live call if the speech or model service fails?
• •Where are recordings and transcripts stored, who can read them, how long are they kept, and are they used for training?
• •Can we keep our phone number and export our call data if we leave?
• •What would the all-in cost be at our call volume, including numbers, transfers and add-ons?