Insights
By Ege Engin Özdaş (Co-Founder & CEO), Şener Özer (Co-Founder & CTO) & Gökçe Özyurt (Co-Founder & CPO) — Staterics · Published 2026-08-03 · Updated 2026-08-03 · 12 min read
The short answer: An AI receptionist covers four groups of front-desk work: booking and changing appointments, answering the questions your documents already cover, routing calls to a named person with the context attached, and the write-up afterwards. Everything outside the scope you define goes to a person. Staterics puts one live within 14 days of data handover.
An AI receptionist answers inbound calls and messages in natural speech and finishes the job on the line instead of taking a message for someone else. It works inside a scope you define, from your own documents rather than a general model’s guess, and everything outside that scope routes to a person with the call’s context attached.
The rest of this page is task by task: which ones it closes alone, which it hands over and on what trigger, and what the whole thing costs.
An IVR offers a menu; an AI receptionist changes something in your systems and says so. That second half is the whole category, and it is the only part a demo can prove.
The test is one call long. If nothing was written and nothing updated by the end of it, you have an answering service with a better voice.
The category also splits. One kind sits on the phone number and exports a summary afterwards; the other lives inside the operation, reading the same calendar everyone else writes to. Ask whether you are buying a one-way export or shared access to the operational record.
A phone layer greets a nine-year regular as a stranger.
Two lists decide it. The second is the one worth reading first.
| Likely worth it | Likely not worth it yet |
|---|---|
| The phone rings several times an hour, mostly repeat questions. | Someone answers on the second ring, doing nothing else. |
| Calls arrive when nobody is free, in evenings, at weekends, during your busiest hours. | You staff every hour you advertise. |
| A typical call ends in a booking, a status check, or a routing decision. | Almost every call is a negotiation, a diagnosis, or a bespoke quote. |
| Your answers exist in writing: prices, policies, hours, inclusions. | The answers live in one person’s head, unwritten. |
None of those disqualifications are permanent. Answers that live only in people’s heads are a writing job, and the first real task of any build.
Four task groups cover most of an inbound line, named for the task rather than the benefit. Each assumes the system can write back to the records it reads, which is what an operating system layer means in practice.
It reads real availability, offers the nearest slots, writes the booking, and confirms, carrying your rules on notice periods, deposits, and slots that need approval. A system that emails staff to enter it later has moved the typing, not booked.
It answers hours, parking, what to bring, which insurers you work with, and where an order stands, all from your own material rather than a model’s guess. It never guesses at prices, medical advice, or commitments.
It identifies the caller, retrieves the open job, and sends the call on with a summary, so the caller never repeats themselves.
The transcript, the record update, the follow-up task, and the confirmation are four jobs someone does after every call. A configured system writes all four before the next call arrives.
A voicemail box asks the caller to wait until morning. A configured system does at nine in the evening what it does at noon: books, answers, routes, flags. An after-hours booking is only as good as the calendar rules behind it. Undefined rules mean refusing everything, or filling slots you would not have.
Check your own log for whether night calls skew urgent; if they do, “someone will call you at nine” needs a named owner. Treat the sector’s statistics carefully too, because vendor figures for after-hours bookings and hang-ups are rarely traceable. Your own phone records hold the honest version.
Partly. “Complex” usually means several requests in one call, a rule that needs bending, or an upset caller. The first is largely solved. The other two are policy questions that belong to you.
| Handled end to end | Routed to a person, with context |
|---|---|
| Booking, rescheduling, cancelling, including callers who change their mind. | Anything needing a rule bent, like out-of-hours slots or waived fees. |
| Questions with a written answer: hours, location, inclusions, order status. | Questions without one: advice, diagnosis, prognosis. |
| Prices from a list you approved, read back verbatim. | Custom quotes, discounts, and terms, or anything else heard as a commitment. |
| Identifying the caller and routing on with a summary. | A caller asking for someone by name, or a flagged account. |
| Recognising an emergency phrase from your list, and triggering the route within the same turn. | Every emergency, without exception, to your on-call route. No triage. |
| Confirming what the record says about amount owed, due date, and appointment time. | Card numbers and payment details, never read to an automated system. |
| One recovery attempt on an unclear caller, confirming back and re-asking. | A caller who repeats, loops, asks for a person, or trips an escalation word. |
The right-hand column is not a list of failures. It is the configuration, and every row is a decision somebody makes once, in writing. A vendor who cannot show it has decided for you.
It says so, and hands over. The trigger list is what to interrogate, because this is where vendor copy goes soft. Intercom’s documentation for its Fin agent is unusually specific, and makes a fair template. Handover fires when:
Every guardrail you add trades a resolution for an escalation. The dial belongs to you.
A trigger is half of it. The handover has to carry the call: the caller’s identity, the request, the attempts already made, and anything promised. Ask a demo to show the receiving end, not the greeting. A caller moves Thursday’s appointment and mentions being charged twice; the AI books the slot and transfers to accounts. Judge the first sentence from that desk. “I have both charges here, and the appointment you moved” means the context travelled. “How can I help?” means a transfer with extra steps.
One tradeoff sits underneath. Every trigger you add makes the system safer and less useful at once. Published resolution rates will not settle it either, because the figure never arrives with the trigger list that produced it, and that list is the only part that would transfer to you.
Two things have to be true, and both are your decisions. The system must recognise an emergency in the words a real caller uses, vocabulary you supply during the build. What follows must be a route rather than an answer, either a live transfer to an on-call number or a scripted instruction to dial the emergency services.
What should never happen is triage. A system that decides how urgent a caller is has taken a safety judgement out of a person’s hands. The honest claim is narrow and sufficient. It reaches a human faster than a voicemail box.
Demos show the call that goes well. Here is the other one. A caller on a poor line, with an unfamiliar name, asks for an appointment on “the 12th”. The name lands wrong and the date is ambiguous. A weak system accepts both, and nobody notices until a stranger’s name appears in the diary.
A well-built one does not resolve ambiguity silently. It spells the name back, restates the date with its weekday, and if the line is too poor to confirm, becomes a messenger, sending a booking link and writing nothing. The unresolved call then follows a defined path:
Three numbers tell you whether it works: how many calls resolved alone, how many routed with context intact, and how many ended with a caller who got nothing. Only the third has to be near zero. Transcripts are business records. Ask where they live and what happens the day you leave; both answers belong in the contract.
Framed as replacement, the comparison is the wrong shape. The person at your front desk does two jobs. One is predictable traffic. The other is exceptions: the upset caller, the odd request, the regular whose situation nobody wrote down. Software takes the first well and the second badly. It replaces the queue, not the judgement.
Some front desks should stay human. Where the first ninety seconds is the relationship, as in bespoke work or with clients who expect to be recognised by voice, automating the greeting is a downgrade. The clearest case is the opposite shape: clinics, service operations, and multi-location businesses, where the same twelve questions arrive all week. The live decision is usually whether to add the next front-desk hire, not whether to remove the person already there. That is arithmetic worth doing properly.
Standalone phone products tend to publish a low monthly figure with per-minute charges on top. Checked in August 2026, IONOS lists $39 to $99 a month, and RingCentral’s AI receptionist page publishes no price at all.
Staterics publishes two lines, and both always apply.
Development is not billed. You pay for the system running, not for building it.
Payment runs monthly or annually, and development is not billed either way. On monthly, both lines begin at go-live, so the first invoice follows the first working day. Annual plans are the one exception to go-live billing: the annual platform fee is split into a deposit up front and a balance on delivery — still platform fee, not development cost.
Voice runs about a token a minute and a chat message a quarter of one. Rates step down by monthly volume, at $0.25 each for the first 500, $0.22 to 2,000, $0.20 to 5,000, $0.18 to 10,000, and $0.16 above that. So a line answered around the clock at roughly 3,000 minutes lands near $655 in tokens, about 22 cents a minute, on top of the platform fee. The full breakdown sits in the pricing piece.
The other number is time, and a timeline means nothing without a starting gun. “Live in five minutes” describes a demo. Staterics AI employees are live within 14 days of data handover, the complete operating system within 90 days. Handover is a short list: hours and locations, who takes what, calendar access, the price list it may read aloud, and your escalation triggers. The clock starts when that list is in hand, not at signature. Current deployments include Doruk İşitme and IMTEK Cryogenics.
The annual guarantee has an objective trigger. If delivery runs past day 90, you choose: 100% of the deposit back, or pay the balance only on delivery.
Before any of that, the free 45-minute Operations X-Ray walks your call flow and leaves you the friction map.
An AI receptionist is not an answering service, an IVR, or a chatbot bolted to a phone line. Those take a message, offer a menu, or type at you. It works in speech, inside a scope the business defines, and completes the request rather than logging it.
An IVR offers a menu and moves the call to a queue. An AI receptionist understands the request in the caller's own words and completes it, so something in your systems has changed by the end of the call.
Yes, and a good handover carries four things: who is calling, what they asked, what was already tried, and what was promised. The person picking up does not start over. A transfer that carries none of those is a queue with extra steps.
It says it cannot answer and routes to a person instead of improvising. Common triggers are a caller asking for a human, frustration, a conversation that loops, a repeated request, and keywords defined during the build.
It recognises the emergency vocabulary the business supplied and runs the route defined for it, either a live transfer to an on-call number or a scripted instruction to call the emergency services. It should never triage.
Yes. The usual arrangement leaves the number where it is and forwards it, whether every call, after-hours only, or overflow when the line is busy. Porting is optional, and forwarding reverses in minutes.
Yes, and without a queue. Concurrency is a seat count for people, not for software, so the thirtieth simultaneous caller is answered like the first. The real limit is the rule set behind the calls: what it may book, quote, and promise.
Voice quality is rarely the limit now; timing and interruptions are the audible tell. The weak points are poor lines, background noise, and unfamiliar names, which is why confirming risky details back matters more than how natural it sounds.
Setup time is measured from data handover, not from signature, which is the number worth asking any vendor for. Handover means hours, calendar access, approved answers, and escalation triggers. Staterics AI employees are live within 14 days of that handover, and the complete operating system within 90 days.
Standalone phone products publish low monthly figures with per-minute charges on top; IONOS listed $39 to $99 a month as of August 2026. Platform-level systems start higher. Staterics publishes two lines that always apply together: a platform fee from $750 per month depending on system complexity and number of users, plus AI usage metered in RIC Tokens. The meter is billed monthly on what the AI actually did. Development is not billed, and on monthly plans billing starts at go-live.