Hashlogics
Voice AI

How AI phone receptionists work, and what to check before you buy

Six systems in a row, and you can hear where each one breaks on a ten-minute demo call.

The short version

5 things that decide this

  1. 01An AI phone receptionist is six systems run in sequence: a phone line, speech-to-text, a language model, text-to-speech, a booking write into your calendar, and a transfer path to a person.
  2. 02Latency is the first thing a caller judges. People swap conversational turns in about 200 milliseconds (Stivers and colleagues, 2009), and ITU-T G.114 calls one-way delay under 150 milliseconds "essentially transparent". A two-second pause reads as a dropped call.
  3. 03Hallucination on a phone call looks like an invented opening time, a price you don't charge, or a booking confirmed out loud and never written. Grounding fixes it: the model reads your systems through tool calls and reads details back before it commits.
  4. 04In Upfirst's analysis of 450,702 calls (June 2026), telling callers they're talking to an AI was associated with about 20% fewer hang-ups. In Maine, and for regulated occupations in Utah, that disclosure is also the law.
  5. 05Before you buy, check five things on the vendor's own site: what you're billed per, which plan includes booking, whether the overage rate is printed, what writes natively into your calendar, and what happens when a transfer isn't picked up.
The pipeline

Six systems, one phone call

When someone dials your number, the call forwards to a phone line the vendor controls, and the audio arrives as a stream. Speech-to-text turns that stream into words a model can read. A language model reads them, decides what to say, and calls tools: check your calendar, look up your hours, hold a slot. Text-to-speech turns the reply into a voice your caller hears. They answer, and the loop runs again.

Two more systems sit outside that loop, and they decide whether the call was worth anything to you. A booking write puts the appointment into the calendar your team already uses. A transfer path hands the call to a person when your caller asks for one or the AI can't answer. A demo shows you the loop and nothing else. Your business runs on the other two.

We build these pipelines for clients, and ZhoopZhoop, a multi-branch auto repair business, is one of them. ZhoopZhoop's receptionist books repair appointments into the shop's own schedule while the technicians are under a car. Holding the conversation was the easy half of that build. Writing into the real schedule, with conflicts and branch rules, was the half that took the engineering.

One caller turn, end to endLive
  1. Phone lineYour forwarded call arrives as an audio stream.
  2. Speech-to-textWords reach the model while the caller is still talking.
  3. Language modelDecides the reply; calls your calendar or FAQ through a tool.
  4. Text-to-speechThe first words play while the rest is still being written.
  5. Booking writeSlot held, read back, confirmed, then written to your calendar.
  6. TransferA person, or a second number if the first doesn't answer.

Latency lives in stages two to four. Your money lives in five and six.

Failure mode one

Latency: the pause your caller hears

Your caller judges the pipeline in the first second. Stivers and colleagues (2009) measured turn-taking across ten languages and found the average gap between speakers near 200 milliseconds. ITU-T G.114, the phone industry's own standard for delay, says most applications get "essentially transparent interactivity" when one-way delay stays under 150 milliseconds. It sets 400 milliseconds as the ceiling for network planning. An AI receptionist adds transcription, inference and synthesis on top of your network, so it can't reach the 150 figure. It has to hide the difference from you.

Streaming is how a good pipeline hides it. Each stage starts on partial output from the one before it. Your words reach the model while you're still talking, and the voice plays its first words while the model writes the rest. A pipeline that waits for each stage to finish adds the delays together. You'll hear the difference on your second question, not your first, because the second one usually needs a lookup. We wrote up the full voice AI latency budget in an earlier post.

Latency also fails in the other direction. If the system decides too early that you've finished, it cuts you off mid-sentence. If it waits too long, you get dead air. Both show up on a demo call the moment you pause to think.

Failure mode two

Hallucination: confident, wrong, and on the record

On a phone call, a hallucination is the model saying something plausible that it never looked up. It tells your caller you open at 8 when you open at 9. It quotes a price for a service you stopped offering last year. Worst of all, it says "you're booked for Tuesday at 2" and nothing was ever written to your calendar. Your caller turns up, and your front desk has no record.

It happens because the model is answering from its prompt instead of from your systems. Grounding is the fix, and it's a build decision. Your hours, services and prices come from a source the model reads through a tool call, at the moment your caller asks. Nothing comes from a paragraph someone pasted in at setup. A booking is confirmed only after the write succeeds, and when the model isn't sure, it says so and transfers you rather than guessing.

Names and numbers are the third trap. Speech-to-text hears "Sean" as "Shawn" and drops a digit from your caller's callback number. Your model then passes the error into your CRM as fact. A receptionist that reads the name, number and time back before it commits catches most of this. One that doesn't will book people you can't call back. We covered why voice agents fail on names, numbers and noise in an earlier post.

  • 01Invented facts: hours, prices and services that came from the prompt, not from a lookup.
  • 02The phantom booking: confirmed out loud, never written down.
  • 03A transcription error passed through as truth, such as one wrong digit in a callback number.
  • 04A guess where a transfer was the right call.
The test

What to do on the demo call

Every vendor publishes a demo number you can ring today. Call it from your own phone, from a noisy place if you can, and run these in order. Ten minutes of your time is enough.

  1. 01

    Ask for something you don't offer

    Request a service the business doesn't provide, or a location it doesn't have. A grounded system tells you it can't help with that, or offers you a transfer. One that invents a yes for you will invent one for your real callers too.

  2. 02

    Give it a hard name and a phone number

    Use an unusual name and a real number, then ask it to read both back to you. Listen for a dropped digit in your number. That's the most common way a booking turns into someone you can't reach.

  3. 03

    Ask for a slot that can't exist

    Try 3am on a Sunday, or a time you know is already taken. It should offer you the nearest real option, not accept the impossible one.

  4. 04

    Pause, then interrupt

    Stop mid-sentence for two seconds and see whether it cuts you off. Then talk over it while it's speaking. Both are latency decisions, and your callers make both of them every day.

  5. 05

    Ask if you're talking to a person

    It should tell you plainly that it's an AI. Upfirst's analysis of 450,702 calls found disclosure associated with about 20% fewer hang-ups. Maine's Title 10 §1500-DD requires it wherever a reasonable caller might otherwise be misled.

  6. 06

    Ask for a human

    Say you'd like to speak to someone and note what happens: a transfer, a message taken, or a promise nobody keeps. Then ask the vendor what happens when your transfer isn't picked up, because after hours it won't be.

  7. 07

    Hang up and check the paper trail

    Within a few minutes you should have a transcript, a short summary and, if you booked, an entry in a calendar. If the booking lives only inside the vendor's dashboard, your front desk will never see it.

  8. 08

    Call twice at once

    Ring the number from two phones at the same time. A receptionist that makes your second caller wait is an answering machine with a queue.

Before you sign

Five things to check on the vendor's pricing page

A demo tells you whether the pipeline works. A pricing page tells you what owning it will feel like. Read it for five facts, and treat a fact you can't find as a fact the vendor chose not to print.

One more setting sits on your side: whether recording is on, and whether your caller is told. Upfirst's data associated a recording disclosure with about 30% fewer hang-ups. Consent rules differ by state, so treat that disclosure as a switch you turn on, not a legal question you settle from a blog post.

  • 01The billing unit. Per minute punishes long calls, per call doesn't, and per unique caller means a regular who phones daily costs you once. Match the unit to how your callers behave.
  • 02Which plan includes booking. Some vendors sell answering on the entry plan and booking on the next one up. A receptionist that can't book is an answering machine with a nicer voice.
  • 03Whether the overage rate is printed. If it isn't, ask for it in writing before you forward your number.
  • 04What writes natively. Google Calendar and Microsoft 365 are the usual direct writes; anything routed through Zapier is one more account and one more thing that can break.
  • 05What happens when a transfer fails. Ask for the fallback: a second number, a message, or a text to whoever is on call.
Six options, read today

How six AI receptionists compare, as of 25 September 2026

GreetKeeper is our own product: Hashlogics builds and runs it, so read that row as a vendor's row. Every other row comes only from that vendor's public pricing page, read on 25 September 2026. Prices move often, so we've kept to the facts that decide your fit. Smith.ai publishes two price lists; this row is its AI Receptionist plans, not its human-receptionist plans.

OptionWho answersBilled perBooking on the entry planTransfer to a personOverage rate on the pricing page
GreetKeeper (greetkeeper.com)AI, in English; transfers to your teamCall answered; very short and screened-out calls don't countYes, on every planYes, with a second fallback number if the first doesn't answerYes, printed per plan
Smith.ai (AI Receptionist plans)AI-first, with optional escalation to North America-based receptionistsCallYes: AI scheduling with Calendly and others, on every plan including the free oneYesYes, printed per call
GoodcallAI agentUnique customer per month, per agent; no fees for calls, minutes or tokensNot stated on the pricing pageNot stated on the pricing pageYes, printed per customer
RubyHuman receptionists ("100% live")MinuteYes, scheduling on every planYesNot displayed
DialzaraAI, in English and SpanishMinuteYes, on every planCall routing on every planYes, printed per minute
RosieAIMinuteNo; booking starts on the second planLive transfer starts on the second planNot displayed
Sources

Where each row comes from

Every competitor cell above was read on 25 September 2026 from the page listed here and nowhere else. If a vendor's page has changed since you read this, the vendor's page wins.

  • 01GreetKeeper: greetkeeper.com/pricing (our own product; see the disclosure below).
  • 02Smith.ai AI Receptionist plans: smith.ai/pricing/ai-receptionist
  • 03Goodcall: goodcall.com/pricing
  • 04Ruby: ruby.com/pricing
  • 05Dialzara: dialzara.com/pricing
  • 06Rosie: heyrosie.com/pricing
  • 07Upfirst hang-up analysis (450,702 calls, 503 receptionists, published 16 June 2026): upfirst.ai/blog/ai-receptionist-hang-ups. Upfirst sells an AI receptionist, and its figures are associations, not causes.
  • 08ITU-T G.114, One-way transmission time (May 2003): itu.int/rec/T-REC-G.114-200305-I/en
Our product

Where GreetKeeper fits

GreetKeeper is the AI receptionist Hashlogics builds and runs. It answers the questions above the way we'd want them answered for our own phone line. You're billed per call answered, and you get booking on every plan. It tries a second number you nominate when your first one doesn't pick up. Its overage rates are printed on the pricing page, so you see them before you sign. Every call reaches you as a transcript and a short summary. It reads the name, number and time back before it books. A webhook fires after each call, booking, message or transfer, so your own systems get the record.

If you're pricing one, start with GreetKeeper's breakdown of what an AI receptionist costs. Then read GreetKeeper's plans and the printed overage rates. And its industry pages show how GreetKeeper is set up for the trade you run, and greetkeeper.com covers the rest of what it does for you. Every plan starts with a 14-day free trial, and you don't need a card to start it.

Questions, answered

Questions this raises

01How does an AI phone receptionist work?+

An AI phone receptionist forwards your calls to a phone line it controls and turns your caller's speech into text. A language model decides the reply, calls your calendar or FAQ through a tool, and text-to-speech turns that reply back into a voice. A booking write and a transfer path sit around that loop. Each stage adds delay, so good systems stream: the model reads words while your caller is still talking.

02What is an acceptable response delay for an AI receptionist?+

Under a second from your caller finishing a sentence to the reply starting, measured on a question that needs a lookup. Human turns swap in about 200 milliseconds (Stivers and colleagues, 2009), and ITU-T G.114 treats one-way delay under 150 milliseconds as transparent. No AI pipeline reaches 150; you're checking that it hides the rest with streaming instead of stacking it.

03Can an AI receptionist make things up?+

Yes, when it answers from its prompt instead of from your systems. You'll see it most in hours, prices, and a booking that was confirmed out loud and never written. Ask the vendor where your hours and services are read from at call time. Then ask whether a booking is confirmed only after the calendar write succeeds.

04Does an AI receptionist have to say it's an AI?+

In Maine, yes, since 1 October 2025. Title 10 §1500-DD bars using an AI chatbot, by text or voice, to mislead a reasonable consumer into thinking they're talking to a person. Clear notice up front is the exception. In Utah, regulated occupations such as medical and legal practices must disclose it at the start of the call. Elsewhere it's still your better setting: Upfirst's 450,702-call analysis found disclosure associated with about 20% fewer hang-ups.

05Should I build a custom voice agent or buy an AI receptionist?+

Buy when your job is answering, booking and transferring on a standard calendar; a product like GreetKeeper or Smith.ai covers that. Build when the call has to read and write your system of record mid-conversation. A repair schedule with branch rules, a patient system or a dispatch board all sit on that side of the line. That's where ZhoopZhoop sat, and it's what we build for you.

By Abdul Basit, CEO, HashlogicsUpdated
Start

Let’s deploy working AI into your business.

We build AI agents and automation, ship them into the tools you already run, then stay on under an agreed service level. A senior engineer reads every brief, and your call gets scheduled within 24 hours.

What happens next

  1. 01

    You send a brief or book a call

    Two minutes, whichever you prefer.

  2. 02

    A senior engineer replies within 24 hours

    Not a sales rep.

  3. 03

    Honest scoping, in writing

    And if we’re not the right fit, we say so.

Abdul Basit, CEO of Hashlogics

“I started Hashlogics because too many teams ship a demo, get paid, and disappear. We build to a standard we’d run ourselves — and we stay to keep it running.”

Abdul Basit · CEO · a direct line

Not ready to talk? Take the checklist.

12 questions to ask any AI agency before you sign. They separate a demo shop from a team that ships to production.

Get the checklist

Free · no newsletter