Voice AI agents: when AI answers your business phone

A garage's phone rings on a Tuesday at 2 p.m. The owner is under a car, both employees are busy. The call goes to voicemail, and the person who wanted a tire appointment calls the next garage on Google. This scene repeats thousands of times a day in Quebec, and it has a measurable cost.

Until recently, the only real fix was human: hire a receptionist or pay an answering service. Since 2025, there is a third option: a custom AI agent that picks up, understands the request, checks your calendar and books the appointment, at any hour. The technology just crossed the threshold where it becomes usable for a small business. It has not crossed the one where it replaces a human in every situation.

This article sorts it out: what voice agents do well as of July 2026, what they really cost, what still breaks, and the cases where one is the wrong purchase.

The short answer, for the busy

A voice AI agent is software that answers a business phone line: it understands speech, queries tools like the calendar or the CRM, and replies out loud in real time. Since 2025, speech-to-speech models, such as those behind OpenAI's Realtime API, connect directly to the phone network over SIP and handle interruptions naturally, with sub-second responses. On running costs, current platforms (ElevenLabs, Retell, Vapi) charge between USD $0.08 and $0.31 per minute all-in, before custom design and integration fees. The investment case rests on missed-call economics: a 411 Locals study measured that only 37.8% of calls to small businesses reach a live person, the rest going to voicemail or ringing out. But the technology remains imperfect: strong accents, complex negotiations and frustrated customers still require a transfer to a human, and a low call volume simply does not justify the purchase.

The economics of the missed call

Start with the problem, not the technology. The best public data on missed calls comes from a study by 411 Locals, which monitored the phone lines of 85 small businesses across 58 industries for 30 days: 37.8% of calls reached a live person, 37.8% went to voicemail, and 24.3% got no response at all. Seven businesses out of ten answered fewer than half their calls. The study dates from 2016, a detail voice-agent vendors tend to omit when they recycle it. But the cause it identifies has not changed: in a service business, the person answering the phone is the same person doing the work.

Rather than quoting unverifiable annual-loss statistics, run your own numbers. Take your average customer value: a tire change, a leak repair, a dental exam. Estimate how many calls per week hit your voicemail, counting evenings and weekends. Then ask the honest question: what share of those people call back, and what share calls the next competitor? For a plumber whose average emergency job is worth several hundred dollars, two or three lost calls a week are enough to exceed the full annual cost of a voice agent.

What changed: the machine speaks in real time

If your last memory of a "robot on the phone" is a menu asking you to press 2, the 2026 technology is a different species. Two technical advances explain the jump.

First, speech-to-speech models. Older systems chained three steps (transcribe the voice, generate a text reply, read it out loud), and every step added delay. Recent models process audio directly. OpenAI's Realtime API, out of beta since 2025, is the best-known example: it connects to the phone network over SIP, meaning an ordinary phone number can terminate directly on the model. ElevenLabs offers the equivalent with its Agents platform, and orchestrators like Retell and Vapi assemble the pieces for integrators.

Second, interruption handling. A real call is not a polite exchange where everyone waits their turn. People cut in, change their mind, rephrase. Current platforms detect the interruption and stop talking. Response latency is now measured in fractions of a second rather than seconds. That point, more than voice quality, is what moved voice agents from frustrating gadget to usable tool.

What it costs, per minute

An AI-handled call adds up four line items: speech understanding, the language model deciding what to say, voice synthesis, and the phone line. Platforms bundle them differently, but the July 2026 orders of magnitude are public.

Retell advertises $0.07 to $0.31 USD per minute depending on the model and voice; its sample configuration comes to about $0.11 per minute. ElevenLabs sells its agents as minute bundles (from $6 USD per month for 75 minutes up to $990 for 12,375 minutes), with extra minutes at $0.08, language model and telephony billed on top. Vapi charges $0.05 USD per minute for its platform and passes model costs through at cost. And for teams building directly on OpenAI, the gpt-realtime-2.1 model costs $32 USD per million audio input tokens and $64 per million output, with a mini version roughly three times cheaper.

In concrete terms: a three-minute call costs between $0.25 and $1 USD. The marginal cost is a rounding error. The real expense, as we explain in our article on what an AI agent costs for a Quebec SMB, is elsewhere: designing the scenarios, wiring in the calendar and CRM, testing, and monitoring the agent in production.

What still breaks

An honest vendor will tell you where the technology fails in 2026, because that is where the design work happens.

Accents and noise first. The models understand a clear voice in a quiet room very well. A customer calling from a job site, engine running, with an accent the model rarely heard in training, still produces comprehension errors. A well-designed agent confirms critical information (name, number, address) instead of assuming it.

High-stakes conversations next. A voice agent excels at structured exchanges: booking an appointment, answering questions about opening hours, qualifying an emergency. It is bad at anything requiring commercial judgment: negotiating a price, assessing a complaint, retaining an angry customer. A frustrated caller who realizes they are talking to a machine gets more frustrated. The design rule is simple: the agent handles the standard cases and transfers the rest to a human, with the context already collected. An agent that cannot transfer is a badly designed agent.

Discipline, finally. A language model wants to be helpful, and on the phone that is a defect: without guardrails, it can promise a discount that does not exist or a slot already taken. The fix is engineering work: strictly limit what the agent may state, and connect it to your real data rather than its memory.

The Quebec French test

For a business here, one question eliminates half the market: does it work in Quebec French?

Separate two things. Comprehension: large models have made real progress on spoken Quebec French, but they are still trained mostly on European French. Local phrasing and informal vocabulary produce more errors than broadcast diction. Synthesis: most default French voices sound Parisian. That is not a technical defect, but to your customers, a voice from France answering for a garage in Trois-Rivières immediately sounds like an offshore call centre. Solutions exist, including cloning the voice of a Quebec speaker, available at ElevenLabs among others.

The practical consequence: never buy a voice agent off an English-language demo. Demand a pilot tested with real calls from real Quebec customers, and measure the comprehension error rate before going live. It is a one-week test that avoids months of frustration.

When a voice agent is the wrong purchase

Three situations where we advise against buying, even though we sell agents.

If you get a handful of calls a day, the economics do not hold: the design cost pays for itself through recovered call volume. Call forwarding to your cell and online booking solve the problem for a fraction of the price. If your calls are the sales moment, as in high-fee professional services, trust is built voice to voice, and delegating that moment to a machine costs more than it returns. And if your schedule is already full, capturing more calls only lengthens your waiting list: your constraint is production capacity, not the phone.

The voice agent is one use case among several, and not always the first to prioritize: our overview of AI agent use cases for Quebec SMBs helps place the phone next to email, booking and back-office work. And if the very notion of an agent is still fuzzy, start with what an AI agent is, explained for an SMB.

Where to start

Before you watch a single demo, open your call log. How many missed calls per week, at what hours, for what kind of request, at what average value? With those four numbers, the decision takes one conversation: either the voice agent pays for itself or it does not, and either way you will know before spending a dollar. That is exactly the scoping we do in a first meeting. 30 minutes, no commitment.

→ First conversation, no commitment

Prices cited are those publicly posted by vendors in July 2026, in US dollars; they move fast in this market. Check the official pricing pages before any decision. Law 25 reminder: call recordings and transcripts contain personal information; inform your callers and set retention rules.