What does an AI phone or chat agent actually do for a small business?
It talks, sorts, and hands over. Everything past that point is a choice somebody made, and most of the bad ones are made there.
By Eugene Romanov·Published
The short answer
An AI phone or chat agent holds the conversation, then hands it to a person. The webhook behind one of these systems refuses any call report without its shared secret, reads the duration to separate a real conversation from a voicemail, matches the number to a lead record, and alerts a human only for the real ones. Short calls stay silent. It is built inside a custom CRM and never sold alone. It does not quote or take money.
SHT-02What drives the number
The whole design is the handoff
A Vapi assistant carries the voice conversation, and the code worth reading runs after it ends. Vapi posts an end-of-call report to app/api/vapi/webhook/route.ts, which refuses anything not carrying the shared secret, reads the duration to tell a real conversation from a voicemail or a wrong number, normalises the phone number and matches it to a lead record, and for a real conversation pushes the summary, the end reason, a preview of the transcript and a link to the recording into one chat through lib/telegram.ts, which in that system is the studio's own founder DM. That is the entire job, and nothing on that path moves a deal forward by itself. Two things this page will not tell you, because the repository does not: which direction those calls run in, which is a setting in the provider's dashboard, and how many are flowing today. What is described here is what the handler does with a call report.
Duration is the cheapest honest classifier there is
The webhook sorts every finished call into three buckets by how long it lasted: a real conversation, something short, or effectively no answer. Only the first raises an alert. The other two are written down and stay quiet. That rule looks crude on paper, and it is the reason the alerts are still worth reading, because an alert that fires for every wrong number is an alert somebody mutes in week two. The interesting part of building agents is almost never the model. It is deciding what the system is allowed to interrupt a human for, and every vendor demo gets this backwards by showing you the conversation instead of the interruption.
Why there is no accuracy figure anywhere on this page
Because there is not one this studio would be willing to defend, and inventing a plausible one is the easiest thing on this page to do. What can be shown instead is the code that decides what counts as a real conversation, and the specific failure each rule was written after. That is the same standard the AI service page holds itself to. If a vendor quotes you a containment or accuracy figure, ask what it was measured on, over how many calls, and in which month, then ask to see the same measurement on calls from your own business. A figure that cannot survive those three questions was marketing.
The telephony underneath belongs to the CRM, not to the agent
A phone call is not a webhook, and handling one is most of the engineering. Six files under app/api/twilio/ do it in that system, and they are the CRM's own dialler for a human caller rather than anything the agent touches: a token endpoint minting browser credentials, a route that starts an outbound call over the provider's API, a voice route returning the dialling instructions, a status callback recording how a call ended, whether no answer, busy, failed or completed, a hook that fires when a recording is ready and sends the audio for transcription from inside that same handler, and a hardcoded debug endpoint for checking the line at all. Nothing in those files mentions the AI assistant. Ask any vendor which parts of their telephony diagram the model actually touches. It is usually fewer than the diagram suggests.
It is never priced on its own, and the running cost is yours
There are two numbers and only one of them is a project fee. The build is part of a custom CRM starting at $2,000 over 30 to 45 days, because an agent answering into nothing is a toy and this studio does not sell it separately. The running cost is per-minute for voice, per-message for chat and per-token for the model, and it is billed to your own accounts by the providers directly. That is deliberate: the invoices arrive in your name and nothing is rented back through the studio. A quote that mentions only the build number is hiding the half you pay every month.
When an answering service or nothing at all is the better buy
Buy an answering service if what you need is a human voice and a message taken, and you have no system for the message to land in. Buy an off-the-shelf AI receptionist if you need it running this week and you can live with the lead sitting inside somebody else's product. Buy nothing if your average ticket is small and your call volume is high, because at that shape a good booking form and a follow-up sequence cost less and do more. An agent built into your own system earns its place when the ticket is large enough that one recovered conversation pays for the month, and when there is a system on the other end for the record to land in. Say which direction you need covered when you ask anyone for a quote: answering calls that come in and running conversations out are different products.
SHT-03In this hub
What does an AI agent cost, and what should it be allowed to do?
Two separate numbers: what it costs to build, and what it costs to keep answering. Most quotes only mention the first.
AI voice agent, answering service, or voicemail?
All three answer the phone. They differ in what exists five minutes after the caller hangs up, and that is the only part worth choosing on.
AI receptionist for contractors: what it costs and what it should never do
The interesting question is not whether it can answer the phone. It is what happens to the call afterwards, and what the thing is not allowed to decide on its own.
Leads arrive in WhatsApp. How to stop losing them
The messenger is not the problem. The problem is that a conversation is not a record: it has no status, no owner, and no yesterday you can go back to.
A2P 10DLC: why your business texts do not arrive without an EIN
It is not a setting and not a checkbox. It is registering your company with the US carriers, and either it is done or the messages are filtered.
What is missed-call text-back, and is it worth building?
The smallest useful agent there is: one trigger, one sentence, and a list of things it must refuse to do. Most of the work is the refusals.
SHT-04General notes
The questions people actually ask.
- N01What does an AI agent do for a small business?
- It carries a voice or chat conversation, decides whether that conversation was real, attaches it to the right record, and alerts a person while the lead is still warm. It is a triage and routing layer, not a salesperson, and everything past the handoff is done by a human on purpose.
- N02Can an AI agent book jobs and take deposits?
- No, and it is not wired to. Sending an estimate, accepting money and marking a job complete are deliberate human actions in every system built here. That is a design decision rather than a limitation waiting to be lifted: those are the actions that generate a refund, a dispute or an angry phone call when a machine gets them wrong.
- N03How much does an AI phone agent cost?
- It is delivered inside a custom CRM build starting at $2,000 over 30 to 45 days and is not sold on its own. Voice minutes, message volume and model usage are ongoing and are billed to your own provider accounts rather than through the studio, so the running cost is yours to see and yours to cap.
- N04Will it replace my receptionist?
- No. The design is the handoff, which only makes sense if there is a person to hand to. What it removes is the gap between a call ending and somebody knowing it happened: the report lands in front of a human with the summary, the transcript and the recording within seconds, and what happens next is a person's decision.
- N05How accurate is an AI phone agent?
- No figure is published here because there is not one worth defending. What is published is the rule the system uses to decide whether a call was a real conversation and the file it lives in, so you can judge the mechanism rather than a number. It gets tested against your own real calls in the first days rather than promised in advance.
- N06Does it work in Russian and Spanish?
- Yes. The studio works directly in English and Russian and multilingual intake is a normal part of a build. The language your customers hear and the language your staff works in are separate choices and do not have to match.
- N07What happens when the agent cannot match the caller to a record?
- The alert still goes out, carrying the number and the summary, and a human picks up the thread from there. The lead lookup is best effort in the code and wrapped so that a miss cannot swallow the notification, which is the right priority: a call you hear about with no record attached is recoverable, and a call nobody hears about is not.
SHT-05Evidence
The systems this argument comes from.
Everything above was learned building these. Each sheet names the business, what the trade's arithmetic actually is, and the file in the repository that proves the capability. No client revenue and no client's customers appear on any of them.