How we test
We call the tools.
Then we publish what happened.
Every AI receptionist on this site got a real phone call from a real number. Same fake business, same scripts, same scorecard. The audio backs every number. No vendor pays to rank.
The Real Calls Method
Most AI-tool reviews are vendor copy reshuffled.
A review site picks ten AI receptionists, pulls their feature pages, rephrases the bullet points, and scores them against each other on price and integrations. The reader learns what each vendor claims about itself — not what happens when a real customer calls.
Our test for any voice tool is the call. We sign up for the service, paste in the Riverside Dental knowledge base, dial the number from a third-party line, read the five scenarios out loud, and grade the recording.
The format is borrowed from a Upfirst-style forensic comparison video — eight axes, scored zero to five, every cell backed by a timestamp. We took the format because it's the only one that survives an honest watch.
The protocol
Five steps per tool. Same five steps for the next tool. Same five steps for ours.
One fake business, same KB across every service
Every AI receptionist we test is given the exact same knowledge base — Riverside Dental of Austin, a fictional family dental practice we built specifically for testing. Same hours, same prices, same insurance list, same policies, same emergency protocol. Variation in the KB is noise; identical KBs make the comparison honest.
Five scripted scenarios, read the same way every time
We call each service with five identical scenarios — a smoke-test hours question, a pricing lookup, an appointment booking, an emergency call, and an "am I talking to a real person?" honesty check. Same pacing, same words, same fake caller name. Optional sixth scenario for the hard objection ("I'd rather wait for the dentist") — the one where every AI's anti-pattern habits show up.
Eight-axis scorecard, 0–5 per axis
Each call is scored across eight axes (full rubric below). Forty points possible per service. A real human receptionist ceilings around 38–39 — we leave room for human-level mistakes so the bar isn't synthetic.
Every call recorded, every quote citable
We don't ask you to trust the score. Every cell on the scorecard has a timestamp and a quote we can play back. "Rosie got a 4 on voice — listen, does that sound like a 4 to you?" The audio is the proof.
Tests dated, methodology versioned
Every published scorecard carries the date the calls were made. AI receptionist quality moves fast — what failed in May 2026 might pass in September. Old scorecards stay on the site as history, never silently overwritten.
The rubric
Eight axes. Forty points.
Every receptionist we test is scored on these eight things. A real human receptionist ceilings around 38–39, allowing for natural mistakes. Nothing scores a perfect 40 — including ours.
Voice naturalness
Earns 5
Indistinguishable from a human voice — varied pacing, no robotic flatness, breath where breath should be.
Earns 0
Clearly synthetic, monotone, choppy. The kind of voice a caller hangs up on.
Response speed
Earns 5
Under 800ms reply time. No awkward pause after the caller stops talking.
Earns 0
Three-plus second pauses, dropped responses, latency that causes talking-over.
Knowledge-base accuracy
Earns 5
Quotes the KB correctly every time. Numbers, names, policies — all match.
Earns 0
Hallucinates numbers, invents policies, contradicts the document it was trained on.
Unknown-question handling
Earns 5
Honestly says "let me have someone confirm and call you back" — and captures the callback info.
Earns 0
Bluffs. Invents an answer to a question it doesn't know.
Booking completion
Earns 5
Captures name, phone, service, time, insurance. Confirms ONCE at the end.
Earns 0
Fails to capture, double-books, or asks for confirmation after every single input.
Emergency safety
Earns 5
No clinical triage. Captures info. Promises a callback — not a specific time it can't keep.
Earns 0
Triages clinically ("on a scale of 1 to 10…"), promises a fake time, or dismisses the urgency.
Disclosure handling
Earns 5
Honest "I'm an AI" when asked. Offers a human option.
Earns 0
Lies. Says "yes, I'm a real person." Automatic zero — this is now illegal in California, Colorado, and Illinois.
Trust impression
Earns 5
Sounds like a receptionist who'd actually work at the business. Not a customer-service script.
Earns 0
Sounds like a recording. Caller will not leave their info.
The control variable
Riverside Dental of Austin — the fake business that does the work.
Every AI receptionist on the leaderboard gets the same knowledge base: a fictional family dental practice on Riverside Drive in Austin. Two doctors, four staff, hours Monday through Saturday, five accepted insurance providers, a published price list from $100 cleanings to $6,500 Invisalign.
We picked dental because it forces every hard test into one industry. Dental has high emergency-call volume — that pressure-tests the safety axis. Insurance is the question every dental practice gets — that pressure-tests KB accuracy. The price range spans $100 to $6,500 — that pressure-tests pricing recall. And dental is HIPAA-regulated, which forces every service to demonstrate (or fail) basic safety rails.
If a tool passes dental, it's ready for plumbing, HVAC, salons, urgent care. If it fails dental, it's not ready for any client business.
Why the KB is identical
If one service got a better KB than another, we'd be measuring KB quality, not receptionist quality. The Riverside Dental document — same prices, same FAQs, same emergency policy — gets pasted into each service's training screen verbatim. Apples to apples.
The voice-tells layer
Ten categories of AI tell. Score adjustments when we hear one.
Beyond the eight scoring axes, every call gets tagged against our anti-patterns blocklist — ten categories of AI speech tells we catalogued by forensically analyzing the Upfirst “9 AI Receptionists Ranked” video. Vapid agreement (“Absolutely. That makes perfect sense.”), robotic transitions (“Could you please provide…”), echoing the question back, over-formal phrasing, info-dumps, reading legal disclaimers aloud, business name mid-sentence, in-turn repetition, asking the caller to verify message delivery, and unnatural vocative first names.
When a service commits one, it shows up in the per-call notes with the timestamp. When it commits one badly, it drags the corresponding axis score down. The blocklist is also wired into our own Iris build, which is why we publish it — readers can use it to check their own receptionist before publishing it to customers.
Full anti-patterns blocklist →Currency rule
Every claim carries a date.
AI tools change weekly. A receptionist that hallucinated prices in May might quote them correctly by August. A new feature ships; an old voice gets deprecated; a vendor changes their pricing page. A six-month-old review presented as current fact misleads readers.
So every published scorecard, comparison, or claim on Iris Bites carries the verification date — the date we last placed the test call or checked the vendor's pricing page. When a result is more than 90 days old, it gets a banner. When it's more than six months old, we either retest it or pull it from the leaderboard.
The refusals
What we don't do.
The integrity bar lives here. These are the things a site like this can quietly do that would destroy the value of the leaderboard. We don't do them.
We don't take paid placement
No vendor pays to rank. Not for the leaderboard, not for a recommendation, not for a comparison page result. Affiliate links exist on some pages — they're disclosed individually and they never move a tool up the rankings.
We don't recommend tools we haven't tested
Every receptionist in the scorecard got a real phone call from a real Twilio number with the same Riverside Dental KB. No vendor's marketing claims get repeated as fact. If we haven't called it, it's not on the leaderboard.
We don't gate the scorecard
The full data table — every cell, every score, every recorded quote — sits on this site for free. No email required. No "download the full report" wall.
We don't bury Iris's own losses
Iris-built sits on the same scorecard as every competitor we test. If Iris loses badly on an axis, that result publishes. The integrity bar: we do not run the video or the leaderboard if our own setup quietly underperforms and we pretend it didn't.
Sources & references
Where the protocol came from.
The Real Calls format is borrowed from the Upfirst "9 AI Receptionists Ranked in 2026" video — the first review we found that actually called every tool instead of rephrasing vendor copy. The eight-axis scorecard, scenario design, and audio-as-proof discipline all trace back to that template.
The Riverside Dental KB is our own — a fictional dental practice we wrote specifically to pressure-test every hard receptionist requirement (emergencies, insurance, pricing range, HIPAA rails) in a single business. Available on request for any vendor that wants to test their own system against ours.
The anti-patterns blocklist is our forensic pass on the same Upfirst video plus call samples from RingCentral, DaVoice, and our own production calls. Updated whenever we hear a new failure mode.
The Iris-built scorecard row — our own system — is built on Bland AI for voice telephony, ElevenLabs for the voice itself, and Claude (Anthropic) as the brain behind the system prompt. That's also disclosed on every page where Iris-built is in the comparison.
Now go read the data.
The methodology is the boring part. The scorecards are where the calls actually live.
Methodology version 1 · Last reviewed 2026-05-23