Insights

Data and articles from the phone lines we build and run: call numbers, Canadian phone compliance, voice-AI field notes. Every statistic carries a named, dated source.

36 articles

Featured

A repeating pattern of identical shapesBuilt, not configuredWhy templated AI receptionists failA templated line does not fail loudly. It answers, books, and logs a successful call, while mishandling the calls that did not fit the template, which are the ones that mattered.Argument · from mechanismHands working at a craftsman's benchBuilt, not configuredWhat a custom AI receptionist actually meansThe word 'custom' gets used for voice and greeting. The things that actually differ between two businesses' lines are all invisible on a demo call.Argument · from mechanismArchitectural plans and drawings on a deskBuilt, not configuredThe decisions behind an AI receptionistThe full list of decisions that sit behind AI answering. A template makes all of them for you, silently and identically for every business.Checklist · 17 decisionsA raised hand signalling stopVoice AIWhat an AI receptionist should refuse to answerRefusing well is a design decision, not a missing feature. The categories that must be declined are predictable, and so is the damage when they are not.Design · refusal categories

Why we build phone lines per business instead of configuring a template.

A headset resting on a desk in a call centreVoice AIWarm transfer, cold transfer, or take a messageThree ways to get a caller to a human, each failing differently. The worst outcome is not choosing: a line that transfers by default drops people into voicemail it cannot see.Comparison · three handoff modesA warning light glowing in a dark roomVoice AIHow escalation rules actually get builtEscalation should never be a judgement call the system makes at runtime. It should be a rule someone wrote down, deliberately biased toward false alarms.Design · deterministic rulesA medical clinic reception desk with staffComplianceAlberta's Health Information Act and your clinic's phoneAlberta's HIA has a specific name for a third party that handles health information on a custodian's behalf, and a specific requirement that follows: a written information management agreement.Alberta HIA s. 66 · RSA 2000 c H-5An analytics dashboard on a screenCall dataWhy a 100% answer rate is the least interesting number we publishA 100% answer rate is what makes the other numbers comparable, not what makes them interesting. On its own it improves when the phone gets quieter.420 calls · 2 businesses · first-partyEmpty chairs in a waiting roomIndustry researchWhat the research says about appointment no-showsA systematic review put the average no-show rate at 23%, with reported rates ranging from single digits to over 50%. The variation is the finding.Systematic review · Health Policy 2018A calendar and pen used for scheduling appointmentsCall dataHow many callers are actually trying to book?Booking-intent share is what converts call volume into something you can plan against. At one med-spa it ran just over half, in a single month at a single location.340 calls · one month · first-partyA desk telephone beside paper files in an officeComplianceIs it legal to record business calls in Canada?One-party consent makes the recording legal. PIPEDA decides whether your business may keep it, and those are not the same test.Criminal Code s. 184 · OPC guidanceA hand holding a smartphone showing a messageComplianceCASL and texting a caller backAn inquiry buys you six months of implied consent under CASL. A purchase buys twenty-four. Missed-call text-backs sit squarely in the first bucket.CASL · CRTC guidanceThe Calgary skyline seen through treesIndustry researchWhat a receptionist actually costs in AlbertaGovernment of Canada wage data puts Alberta receptionists between $16.75 and $30.00 an hour. Everything interesting about the comparison sits outside that range.Job Bank · updated Nov 2025A person taking a phone call at an office windowComplianceDo you have to tell callers they are talking to an AI?There is no Canadian rule that names AI disclosure. But meaningful consent under PIPEDA is hard to argue for when the caller has been misled about what they are talking to.PIPEDA consent standard · OPC guidanceAn old answering machine on a shelfIndustry researchDoes anyone leave a voicemail anymore?An empty voicemail box is usually read as 'nobody needed us'. It is at least as consistent with 'nobody who needed us was willing to wait'.Reasoning · first-party contextTwo people shaking hands at a meetingCall dataFirst-time callers versus repeat callersCaller mix decides what a missed call costs. A returning customer calls back; a stranger comparing options calls somebody else.80 calls · 15 weeks · first-partyA clock on a waiting room wallIndustry researchHow long callers wait on hold, and why the published numbers are weakThe 80/20 service level is a real, durable industry convention. Most of the abandonment percentages quoted around it are vendor aggregates citing each other.Convention · 80/20 service levelRows of servers in a data centreComplianceHow long should you keep call recordings?Canadian privacy law gives you a test, not a number. Absent a decision, the default becomes whatever your vendor set, which is the one answer that is never defensible.PIPEDA retention principleA magnifying glass resting on printed detailBuilt, not configuredHow to evaluate an AI receptionistA four-minute phone test and eleven vendor questions, including the ones that are uncomfortable for us. Feature lists tell you almost nothing.Checklist · buyer's testDifferent tools arranged on a workshop wallBuilt, not configuredWhy every industry needs a different phone scriptEvery industry's phone call has a different hardest decision, and that decision, not the greeting, is what a receptionist is actually for.Comparison · 12 industriesA measuring tape and hand toolsOperationsFour numbers that tell you whether your phone is workingFour numbers, all sitting in your phone provider's reporting already. Answer rate is the one everyone quotes and the one that conceals the most.Method · four metricsA telephone handset resting on a desk in low lightCall dataThe Missed Call Report, Edition 1Two Calgary businesses, both answering every call. One took 340 in a month at a single location, the other 80 across a whole season -- and the quiet line was the expensive one.420 calls itemised · 2 businesses · first-partyA reception desk in a medical officeCompliancePHIPA and answering services in OntarioPHIPA defines an electronic service provider as anyone whose service lets a custodian handle personal health information. A phone system that records and stores patient calls qualifies.PHIPA · SO 2004 c 3 Sch AA worker on a construction site holding a phoneVoice AIJob sites, salons and restaurants: noise on business callsThe phone channel throws away most of the frequency range before noise is even considered. For trades, salons and restaurants, difficult audio is the normal case, not the exception.Mechanism · audio channelHands typing on a laptop keyboardIndustry researchPhone calls and web forms are not the same leadA form is a request to be contacted later. A call is contact happening now. The research on response speed says the second is worth more, and it usually gets less process.Reasoning · lead-response researchA notebook and pen on a deskOperationsWhat to capture on a first call, and in what orderMost intake scripts ask for what the business wants in the order it wants it. Order by what survives a dropped call instead, and the callback number goes first.Method · field orderingDocuments and a pen on a deskCompliancePIPEDA and an AI phone system: what actually appliesA voice agent collects personal information on every call. Three of PIPEDA's ten principles are where an AI receptionist actually gets into trouble.PIPEDA · 10 fair information principlesAn office building exterior in a Canadian cityComplianceQuebec Law 25 and recorded callsLaw 25 requires the consent request to stand on its own, separate from other information. That single structural rule is what most call-recording notices fail.Law 25 · P-39.1People talking on a busy city streetVoice AIAccents, and where speech recognition still failsKoenecke et al. measured a word error rate of 0.35 for Black speakers against 0.19 for white speakers across five commercial ASR systems. A voice agent answering a public number has to be designed knowing that.PNAS 2020 · 5 ASR systems · 115 speakersA runner in motion against a wallIndustry researchSpeed to lead: what the research actually saysThe 100x figure comes from a 2007 MIT/InsideSales study, not from HBR. The 42-hour average comes from HBR. Almost every page quoting them attributes at least one wrongly.2 studies · 2007 & 2011Two people in conversation across a tableVoice AIBarge-in: what happens when a caller interruptsLetting a caller cut in mid-sentence is the single feature that separates a conversation from a phone menu, and it is an acoustics problem before it is a product decision.Mechanism · turn-takingA stopwatch on a plain surfaceVoice AIHow much silence before a caller thinks the line is deadStivers et al. found response gaps peaking within 200ms of a question ending, across ten unrelated languages. That is the budget a voice agent is measured against, whether or not it can meet it.PNAS 2009 · 10 languagesAn hourglass with sand running through itCall dataWhat a 76-second phone call actually capturesAcross a full peak season, the average conversation on one landscaping line ran 76 seconds. That is what capturing a lead properly actually costs in time.80 calls · 15 weeks · first-partyCity lights at nightIndustry researchWhen do customers actually call? Most businesses cannot sayWe will not publish a chart of when customers call, because the honest version is per-business. What we will do is explain how to produce yours in about an hour.Method · no borrowed dataA small business owner working in their shopOperationsWho should answer your phone?Four options, and the right one depends on when your calls arrive rather than on which sounds most professional.Comparison · four optionsAn audio waveform on a dark screenVoice AIWhy voice AI mishears names and street addressesOrdinary sentences are a closed problem with a language model behind them. A surname is an open one with nothing to fall back on. That difference is the whole failure mode.Mechanism · open-set vocabulary

Want your own version of these numbers?

Every figure in these reports started as a business that could not answer its own phone. We build a working line on your calls and your booking rules, you run it for about 14 days, and at the end you have your own data instead of somebody else's.

Start free pilot