How to Choose an AI Receptionist: A Buyer's Checklist

Run the demo yourself, don't watch theirs

A scripted demo where a salesperson plays the caller tells you almost nothing. The salesperson knows the happy path. Ask for a number you can call yourself, from your own phone, unannounced, and run the same five tests on every vendor on your list.

• Interrupt it mid-sentence. Start talking over the greeting. A well-built agent stops within a beat and listens. A weak one talks over you for another five seconds, or stops, loses the thread, and restarts its whole paragraph.
• Change your mind mid-sentence. "Tuesday morning, actually no, make it Thursday afternoon." You are testing whether it holds the correction or books Tuesday. Systems that pass this are noticeably better at real calls, because real callers do this constantly.
• Give it something hard to hear. Spell an unusual last name letter by letter, say a phone number as "double four," include a unit number with a letter in it. Then ask it to read the details back. The real test is not what it says on the call, it is what arrives in the confirmation text and in whatever record it creates.
• Ask something it cannot know. A price it has not been given, a clinical or legal question, something about another customer's account. The only acceptable behavior is saying it does not know and offering a person. An agent that invents a plausible answer on a demo call will invent one on a real call, and that is disqualifying, not a rough edge to be tuned later.
• Ask for a human. Plainly the first time, then again a minute later sounding irritated. Count the turns before it routes you, and check whether the person on the other end receives what you already said or makes you start over.

Call twice, at different times of day, and once from a car or a speakerphone with noise behind you. Then ask the vendor a question they do not expect: is the demo running the same stack, the same model, and the same telephony path as production. It is sometimes not, and that answer changes what your test was worth.

Latency: ask for the number, then ask which number

The pause between a caller finishing a sentence and the agent starting to reply is the single most reliable signal of build quality, because it is the sum of everything happening underneath: speech recognition, endpointing, the model, the text-to-speech, and the network hops between them. The mechanics are covered in how AI voice agents work.

There is a reference point for what feels normal. A 2009 cross-linguistic study of conversational turn-taking, published in PNAS, found that across ten languages the gap between one speaker finishing and the next beginning peaks within roughly 200 milliseconds, with cultural variation of well under a second. That is the standard a caller's ear is unconsciously applying.

Voice-engineering write-ups converge on a practical band: under roughly 800 milliseconds of end-to-end response latency, most callers experience the conversation as smooth; between 800 and 1,200 it is workable for business calls; past about a second and a half people start talking over the agent or assume the line has dropped. Treat those as rules of thumb from the people who build these systems rather than as a published standard.

Two follow-up questions matter more than the headline figure. First, ask for the 95th percentile, not the average, because an agent that sits reliably at 900 milliseconds beats one that averages 700 and spikes to 2,500 twice a call. Second, ask what happens to latency when the agent has to look something up mid-call, such as real calendar availability. Many systems go silent for two or three seconds during a tool call. The honest fix is a spoken filler ("let me check that for you") rather than dead air, and you can hear whether they have bothered to build it.

The edge cases are the product

The middle of the call is easy and every vendor is fine at it. What separates deployments that work from deployments that get switched off after a month is what happens at the boundaries, and those boundaries are configuration, not intelligence.

Ask to see the actual escalation rules in writing, not a description of them. Specifically: what triggers a handoff to a person, what the agent says while handing off, what the person receives with the call, what happens when nobody picks up the escalation, and what happens after hours when there is no person to escalate to. "It transfers to your team" is not an answer. A transfer at 11 p.m. to a phone nobody is holding is a hang-up with extra steps.

Ask which direction uncertainty resolves. A correctly built agent escalates up when it is unsure, never down, and never resolves ambiguity by guessing. This should be a stated rule you can read, not a property people assume the model has.

Then ask the question almost nobody asks: what does a caller hear when your system is down. The right answer is an automatic failover configured at the carrier or forwarding layer, so calls ring through to your existing line or voicemail if the agent does not pick up within a set number of seconds. The wrong answers are a busy signal, endless ringing, or an apology page on a status site. Vendors who have thought about this will tell you the failover timer in seconds. The failure modes underneath all of this are catalogued in the common mistakes guide.

Read access, write access, or neither

"Integrates with your software" hides three very different situations, and the difference decides whether the agent saves you work or just moves it.

Read-only means the agent can see availability or customer records but a human still enters the result. Write access means it creates the booking, the work order, or the CRM record itself. Neither means there is no live connection at all and the vendor is describing a roadmap, an export, or a generic automation platform sitting between two systems that neither vendor supports properly.

So the useful question is not "do you integrate with our system." It is: which of those three are we, are you connected today to our exact product on our exact plan tier, and what does the system do when a write fails. That last one is where deployments quietly rot. An integration that drops every third booking without telling anyone is worse than no integration, because the failure is invisible until a patient shows up for an appointment nobody has.

Where write access genuinely is not available, the honest design is a structured handoff rather than a fake integration: a complete, categorized record delivered by email, SMS, or webhook alongside the transcript, so someone spends fifteen seconds on data entry instead of six minutes on a call. A vendor willing to say that out loud is telling you something useful about the rest of the build. Our own position on which platforms connect directly and which go through certified integration partners is on the integrations page, and the booking mechanics are in how appointment automation works.

Your phone number is the part that is hard to undo

There are two ways to put an AI agent on your line, and they carry very different exit costs. Forwarding leaves your number where it is and points it at the agent, which takes minutes and is reversible in minutes. Porting moves the number onto the vendor's telephony account, which is cleaner operationally and much harder to unwind if the relationship ends badly.

Under FCC local number portability rules, a simple port, broadly a single-line account with no complex switch translations, no unbundled network elements, and no reseller involved, is meant to complete within one business day, a standard set in the Commission's 2010 porting-interval order. That is the mechanical timeline. The commercial timeline is the one to pin down: if you cancel, who is the account holder of record on that number, what exactly do you have to do to get it back, and how long does the vendor take to release it.

Start with forwarding unless there is a specific reason not to. If you do port, get the release process in writing before the port, not after. A vendor who is relaxed about that question is usually fine; a vendor who is vague about it is telling you what leaving will feel like.

Compliance: which questions apply depends on what you do

None of this is legal advice, and the details vary by state and by industry. These are the four areas where the answer changes what you should buy.

Healthcare and anything touching patient information. If the agent creates, receives, maintains, or transmits protected health information, the vendor is a business associate and you need a signed business associate agreement before it touches a single call, under 45 C.F.R. § 164.502(e). Two dodges to watch for. The conduit exception is narrow: it covers services that transport information without storing it beyond what is incidental to transmission, in the mould of a courier, not a platform that keeps transcripts. And encryption does not remove the requirement; HHS guidance on cloud service providers is explicit that a provider storing encrypted PHI is still a business associate even when it holds no decryption key. Also note that there is no such thing as a HIPAA certification, for us or anyone else, which is the subject of the HIPAA guide.

Call recording. Most states require only one party to consent, but a group commonly counted as eleven, including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania, and Washington, requires all parties to consent, and a few others sit in between depending on whether the case is civil or criminal and whether the conversation is by phone or in person. Ask whether recording can be disabled per state, whether the disclosure is spoken before anything is captured, and whether transcripts are treated differently from audio. Confirm the specifics for your states with counsel, because the classifications do shift.

Outbound is a different legal regime from inbound. Answering calls people place to you is not regulated the way calling them is. On February 8, 2024 the FCC issued a declaratory ruling confirming that an AI-generated voice counts as an "artificial" voice under the Telephone Consumer Protection Act, so AI-voiced outbound calls sit inside the existing consent rules rather than outside them. Separately, since April 11, 2025 callers have had to honor a revocation of consent made by any reasonable method and process it within ten business days; the broader "revoke-all" provision at 47 C.F.R. § 64.1200(a)(10) has been repeatedly delayed and now carries an effective date of January 31, 2027. If a vendor pitches automated outbound campaigns, ask how consent and revocation are tracked per number, across voice and SMS together.

Disclosure that it is AI. There is no single federal rule, and the state picture is genuinely in motion. Utah's Artificial Intelligence Policy Act, as narrowed by SB 226 in 2025, requires a business to disclose generative AI use when a consumer clearly asks, and requires proactive disclosure at the start of high-risk interactions in state-regulated occupations. California's bot-disclosure law applies to online communications rather than phone calls. Colorado's original AI Act was repealed and replaced in May 2026 by a narrower disclosure-focused statute that takes effect on January 1, 2027. Our own position does not depend on which way that settles: the agent identifies itself as AI when asked, every time.

Data: what is kept, what trains on it, what you get on the way out

Call recordings and transcripts are the most sensitive thing your business will hand a vendor, and the terms are usually in a document nobody on the sales call has read.

• Is our call data used to train models, by you or by any provider underneath you, and is the opt-out contractual or a setting someone can flip?
• What is the retention period for audio, transcripts, and structured records, and can we set it ourselves?
• Who at your company can listen to our calls, under what process, and is it logged?
• Which subprocessors touch the audio, and where is it stored?
• On termination, what do we get back and in what format, and how long until it is deleted?

"We do not train on your data" belongs in the agreement, not in the sales call. If the answer is confident, having it written down costs the vendor nothing.

Pricing: read the meter, not the sticker

Published prices for AI receptionists are not comparable across vendors, because the unit is not the same. Some charge per minute, some per call, some a flat monthly rate with an included allowance, and the allowance is where the surprise lives. The cost guide breaks down the four models, the pass-through fees that never reach the plan page, and the break-even arithmetic between them.

Ask what counts as billable. Does the meter start when the call connects or when the agent starts speaking? Is silence billed? Is time rounded up to the minute per call, which is a real cost difference when the average call is 90 seconds? Are spam and robocalls billed to you? Is the time after a transfer, once a human has taken over, still on the meter? Are outbound texts, including confirmations and reminders, inside the plan or extra?

Then the commercial terms: the overage rate once the allowance is gone, the setup fee and what it covers, contract length and notice period, the price at renewal rather than the introductory price, and who pays for script changes after go-live. That last one separates a product from a service. If every wording change is a billable ticket, the system will drift out of date, because nobody wants to file a ticket to fix a sentence.

When the answer is not to buy one

There are businesses where an AI receptionist is the wrong first purchase, and it is worth knowing whether yours is one before you sit through four demos.

• Low call volume. If the phone rings a handful of times a day and someone always answers, the recoverable revenue may not cover the monthly cost. Measure your actual missed-call rate first, from your carrier's call log, before buying anything to fix it.
• Calls that are overwhelmingly existing customers asking account-specific questions, with no integration available to answer them. Without read access to the record, the agent can only take a message, which your voicemail already does.
• Businesses where the phone call is the sales conversation, not the route to it. If the call is a consultative negotiation, the useful automation is capturing and routing it quickly, not conducting it.
• No internal owner. Someone has to watch the escalation queue and the transcripts for the first few weeks and correct what the agent gets wrong. Deployments without that person are the ones that get switched off.

In several of those cases the cheaper move is missed-call text-back, which recovers the calls you are already losing without changing how the phone is answered at all.

The checklist, in one place

• Can I call the demo number myself, unscripted, from my own phone?
• Does the demo run the same stack, model, and telephony path as production?
• Does it handle interruption, self-correction, spelled names, and background noise?
• What does it do when it does not know something?
• How many turns to reach a human, and does the human get context?
• What is the 95th-percentile response latency, not the average?
• What happens to latency during a mid-call lookup, and is there a spoken filler?
• Can I read the escalation rules, including after-hours and no-answer paths?
• What does a caller hear when the system is down, and what is the failover timer?
• Is our integration read, write, or neither, and is it live today on our plan tier?
• What happens when a write fails, and how would we find out?
• Are we forwarding or porting, and if porting, what is the release process on exit?
• Will you sign a business associate agreement, if we handle health information?
• Can recording and disclosure be configured per state?
• How is consent and revocation tracked if we ever run outbound?
• Does the agent identify itself as AI when asked?
• Is our call data used to train models, and is the opt-out in the contract?
• What is the retention period, and what do we get back on termination?
• What exactly is a billable minute, and what is the overage rate?
• Contract length, notice period, renewal price, and who pays for script changes?

Take the same twenty questions to every vendor. The differences show up immediately, and they show up in which questions get a number and which get a paragraph.

Sources

• T. Stivers et al., "Universals and cultural variation in turn-taking in conversation," PNAS (2009) — supports the roughly 200 millisecond peak in response offsets across ten languages, and the finding that cultural variation in that gap is small.
• Practitioner latency guidance published by voice-infrastructure and voice-testing vendors, including Telnyx and Hamming AI — supports the under-800ms, 800 to 1,200ms, and past-1,500ms bands, and the p95-over-average point. These are engineering rules of thumb, not a published standard.
• FCC, Local Number Portability Porting Interval and Validation Requirements order (FCC 10-85, 2010), and 47 C.F.R. § 52.35 — support the one-business-day simple-port interval and the definition of a simple port.
• 45 C.F.R. § 164.502(e) and HHS guidance on business associates and cloud service providers — support the business associate agreement requirement, the narrowness of the conduit exception, and the point that encrypted storage without a decryption key still makes a provider a business associate.
• State call-recording surveys, including the Digital Media Law Project's guide and the Reporters Committee for Freedom of the Press recording guide — support the one-party and all-party consent split and the existence of mixed-rule states. Counts vary between compilations and the law changes; confirm your states.
• FCC Declaratory Ruling of February 8, 2024 — supports the holding that AI-generated voices are "artificial" voices under the TCPA.
• FCC consent-revocation rules effective April 11, 2025 (any reasonable method, ten business days), and the Consumer and Governmental Affairs Bureau order of January 6, 2026 extending the effective date of the revoke-all rule at 47 C.F.R. § 64.1200(a)(10) to January 31, 2027 — support the revocation description.
• Utah Artificial Intelligence Policy Act as amended by SB 226 (2025); California Business and Professions Code § 17941 (online bots); Colorado SB 26-189 (signed May 14, 2026, effective January 1, 2027, replacing SB 24-205) — support the disclosure-law summary.
• None of the above is legal advice. Recording, disclosure, consent, and health-information rules vary by state and industry and are actively changing; confirm the specifics with your own counsel before a script goes live.