One topic, 5 expert lenses — a solo operator, the caller, a stat auditor, a field empiricist, and compliance counsel — discovered for this topic (STORM method), forced to contradict each other, then every claim checked against its primary source.
Nobody neutral has ever measured the problem AI receptionists claim to solve — the "62% of calls missed / $126K lost" numbers all trace back to the companies selling the cure. What IS solidly measured cuts both ways: the only randomized trial on bot disclosure found that telling people they're talking to a machine cut conversions by roughly 80%, and state wiretap and disclosure laws increasingly force you to tell them. That collision would kill the whole idea — except the honest comparison for a small business isn't AI versus a good human, it's AI versus the voicemail most callers hang up on. The defensible move: deploy it after hours as a disclosed voicemail replacement with a fast path to a human — and pull your own 30-day call log before you believe anyone's statistics, including these. Replacing a competent human at the desk during business hours has no independent evidence behind it at all.
The only randomized controlled trial on voice-bot disclosure (Luo et al., Marketing Science 2019 — 6,255 randomized sales calls) found undisclosed AI voice bots sold as well as proficient human agents and roughly four times better than inexperienced ones — but revealing "this is a bot" before the conversation cut purchase rates by 79.7% versus the undisclosed bot, with 56% of disclosed callers hanging up. The overlooked nuance: disclosing after the call largely erased the penalty — it's the pre-conversation reveal that kills. The caveat that caps the score: the setting was outbound fintech sales calls, and the authors themselves flag that "more dynamic inbound calls" are untested. The direction is solid; the magnitude may not transfer.
The FCC's 2024 AI-voice ruling (FCC 24-17) governs calls you make, not calls you answer. The live risk for inbound AI answering is state law: a California wiretap (CIPA) class action against AI phone-answering vendor ConverseNow was allowed to proceed in 2025 — exposure runs $5,000 per call — and under Ambriz v. Google, the AI vendor itself can be the unauthorized eavesdropper on every call in an all-party-consent state. Utah fines undisclosed AI up to $2,500 per violation (disclosure on request; proactive only for high-risk interactions since its 2025 amendment). Both cases are pleading-stage rulings, not liability findings — but one pre-answer sentence ("you're talking to an AI assistant; this call may be recorded") collapses most of this exposure either way.
"62% of small-business calls go unanswered" traces to a single January 2016 study by 411 Locals — an SEO agency that sells businesses more phone calls — sampling 85 businesses with no published sampling method, yet routinely relabeled "a 2024 study." Its own breakdown: 37.8% answered live, 37.8% voicemail, 24.3% no response. The 62% counts voicemail as "missed." The "85% of missed callers never call back" first appears in a 2016 Aircall blog post with zero attribution and has no locatable methodology anywhere; the "$126K/year lost" is answering-service arithmetic stacking those two numbers on an assumed customer value. No independent (non-vendor) measurement of SMB missed-call rates exists — we looked.
Every lens converged here from a different direction. The peer-reviewed field evidence shows voice AI persistently reduced complaints but drove demand to escalate to humans (Wang et al. 2023), and the strongest published productivity result is AI assisting humans (+15% resolutions/hour — Brynjolfsson et al., QJE 2025, text chat), not replacing them. Even the "20–40% typical containment" range comes from chatbot vendors' own blogs; the one independent anchor is Gartner: just 14% of issues get fully resolved in self-service. Meanwhile roughly 1 in 3 callers say they'd hang up on a bot — but most callers hang up on voicemail too. After hours, the bot competes with silence.
Metrigy (n=503): 84.7% prefer a human; 80.1% still do even when told the AI would resolve the issue. Callvu 2024: 81% would rather wait on hold for a person. UJET: 80% say bots increase frustration — but that data is from 2021-22, before modern voice AI. AnswerConnect (a human answering service running a "People, Not Bots" campaign): would-hang-up-on-AI rose from 29% to 31% — a delta within polling noise. All but Metrigy are vendor-run or vendor-commissioned; Metrigy, the cleanest of them, is an analyst survey, not peer review. The direction is consistent enough to respect, the magnitudes are soft. And the Luo RCT suggests stated preference and actual behavior diverge: people bought from bots just fine until told.
These circulate in every AI-receptionist pitch. The competitor stat and the dollar figure have no locatable primary methodology; the ROI case studies ("$48K recovered for a 5-truck plumber," "37% of unconverted calls recaptured") name no verifiable businesses and are published by the vendors themselves. Treat all of them as marketing, not measurement — do not repeat them in your own planning.
Findings 1 and 2 collide head-on: the law increasingly requires the exact disclosure that the only randomized trial says cuts conversion by ~80%. Read together, they look like a verdict against AI answering entirely.
They aren't — because of Finding 3. The vendors' own inflated statistics hide the real shape of the opportunity. Once you correct the 62% down to the ~24% of calls that get no response at all, the business case stops being "replace your receptionist" (where disclosure-depressed conversion competes with a persuasive human, and loses) and becomes "answer the calls that currently die in voicemail" (where a disclosed bot competes with silence, and silence books nothing). The legally compliant deployment and the economically defensible deployment turn out to be the same deployment — and it's the one the vendor pitch undersells.
The disclosure law and the disclosure penalty cancel out exactly where the vendors aren't pointing: after hours, against voicemail, with an honest "I'm an AI" up front. (This leans on Finding 4's convergence — medium confidence, no controlled SMB trial exists.)
Discovery surfaced a sixth perspective that was cut for overlap: The Front-Desk Veteran — the human who answers now, cleans up botched bookings, and hears the relieved sigh when a panicked caller reaches a person. Her question never got answered: "Which callers quietly hung up on the bot and never called back — and does anyone even log those?" No vendor dashboard measures silent churn; every dashboard measures captured calls.
That's the omission that could invert these findings: if business-hours callers who bail on the bot outnumber the after-hours callers it captures, the net effect is negative and invisible in the tooling. Build with this caveat in mind — it's why the attribution test below is not optional.
For the owner of a small service business being pitched an AI receptionist — specific moves, not principles.
Your carrier or VoIP portal has the real number. Count calls with no response at all — not voicemail, which you often return and book anyway (Finding 3). Your number, not 411 Locals' 85-business sample, is the business case.
After hours the bot competes with voicemail and wins by existing (Finding 4). Business-hours replacement is where the evidence is thinnest and the disclosure penalty (Finding 1) bites hardest.
"You're reaching our AI assistant — I can book you now, or a person will call you back first thing." That one sentence is your CIPA/Utah control (Finding 2) and the caller's own stated tolerance condition (Finding 5). If the vendor can't do a pre-answer disclosure plus fast escalation, walk.
Tag every AI-booked job: "would I have called this person back and booked them anyway?" Subtract those before computing what the tool earned. The dashboard claims credit for your own callback habit — and never logs the callers who hung up on it (the missing lens).
In all-party-consent states (CA, FL, IL, MD, MA, MT, NH, NV, PA, WA…), the vendor listening in is the exposure — and the lawsuit names you alongside them (Finding 2). Ask directly: who indemnifies whom for CIPA claims?
It's the one number that captures goodwill damage no dashboard shows. If AI-booked customers don't come back at your usual rate, the bot is buying you one-time jobs with repeat relationships.
Nobody has published this. The vendors hold exactly this data across thousands of deployments and release none of it; the academics have only studied adjacent settings. It's the question that decides whether business-hours AI answering is quietly negative — and until someone answers it independently, every small business running the 4-week attribution test above knows more about the real answer than the published literature does.