Voice agents · Competitor teardown · 16 Nov 2026
What competing AI phone receptionists have that our voice agents lack, how they were built, and what goes into the master agent as a result.
Decided 16 Nov 2026 (Rachel)
patient turn-taking setting while a caller gives an address. The July tests recorded in LEARNINGS found 4 to 15 second gaps with it, so it only goes in after a handset test.Summary
Competing AI receptionists have moved past message-taking: calendar booking appears in 11 of the 14 UK products surveyed and some form of human hand-off in 10 of 14, and in the US booking, warm transfer, mid-call text links and a website-built knowledge base are standard at the $49 to $149 tier (UK product pages, e.g. Callout AI; Rosie pricing). RomanLogic's voice agent lacks all of those, plus returning-caller memory, an owner correction loop and job-software logging, but it leads on the things almost nobody else documents: postcode read-back with lookup and fixed-wording emergency safety lines, matched only by Answrd (Answrd trades).
Almost every rival bought its voice layer (Retell, Bland, Vapi, LiveKit) and spent its engineering on onboarding, transfers and integrations, so our ElevenLabs Agents + Telnyx + Claude Haiku 4.5 stack sits at the same level as theirs and does not need replacing.
The research supports the plan to route the model step through our own service with a per-call record: this is the published mainstream pattern, and "having to repeat information already given" is named by 41% of consumers who have dealt with an AI tool (AnswerConnect). It changes the plan in five places: the custom model endpoint needs an explicitly configured hosted fallback or a failure ends the call, the service should choose the next question in code instead of blocking a bad reply, re-asks need a two-ask cap with a change of method, the test rig must run each scenario many times on real speech-to-text output, and three ElevenLabs behaviours must be tested before building.
The evidence is uneven: vendor documentation is solid, independent review evidence is thin, no competitor demo line was called, and no one has published a before-and-after re-ask figure for any of these patterns.
The gap list in the brief is accurate, and the research ranks it. Two missing features are now expected by buyers. Calendar booking is offered by every UK product read except Answrd, Down To Earth AI and AI Answerphone, usually against Google, Outlook or Cal.com with a confirmation text to the caller (Sayora; TradeVoice; Fasthosts). Live transfer is common but paywalled, which shows where vendors think the value sits: it is on Callout AI's £149 Pro tier, Arcscribe's £129 Pro tier and a £29 a month add-on at TradeVoice (Callout AI; Arcscribe; TradeVoice). In the US Rosie gates direct booking, warm transfers and mid-call texts at $149, while Upfirst includes them on every plan and gates only on volume (Rosie pricing; Upfirst pricing).
The remaining gaps are differentiators, held by a few. Returning-caller memory is claimed on every plan by Callout AI and listed by Arcscribe in the UK; Upfirst shipped it in March 2026 as per-contact notes the owner can edit, and Jobber greets matched clients by name from caller ID (Callout AI; Upfirst changelog; Jobber help centre). A mid-call photo link is live only at Arcscribe; Answrd's is still marked as in testing (Arcscribe; Answrd). About six UK products answer price questions, and all of them quote only figures the owner entered (TradeVoice; Answrd trades). The table summarises where each gap stands.
| What we lack | Who has it | Status in the market |
|---|---|---|
| Calendar booking | 11 of 14 UK products; Rosie, Goodcall, Upfirst, Jobber, Housecall Pro, ServiceTitan, Workiz, FieldPulse | Table stakes |
| Live transfer | 10 of 14 UK products; Rosie (cold, warm, waterfall), Smith.ai, Upfirst, Jobber, ServiceTitan | Table stakes, often the paid upgrade |
| Website or Google listing knowledge base | Sayora, Whoza, BT, Fasthosts, Rosie, Goodcall, Upfirst, Dialzara | Table stakes, and the onboarding method |
| Mid-call text link | Rosie, Upfirst, ServiceM8 (booking form), ServiceTitan (payment link) | Common in the US |
| Photo request by text | Arcscribe (live), Answrd (in test) | Rare |
| Returning-caller memory | Callout AI, Arcscribe, Upfirst, Jobber, Workiz, Avoca | Differentiator |
| Owner correction loop | Goodcall, Smith.ai, Sayora, Frontdesk | Differentiator, and a retention feature |
| Job-software logging | Rosie (18 connectors); native only at ServiceM8, Jobber and similar | Open lane in the UK |
Rosie's documentation is the most useful design reference because it states the rules. Its warm transfer puts the caller on hold, rings the team member, gives a spoken briefing, and connects only on a spoken yes; a decline or a voicemail pickup sends the agent back to take a message, there is one attempt per call, and recording stops at hand-off (Rosie call transfers). Its intake is a short prose brief per scenario, explicitly "not a question list", and the agent skips anything the caller already volunteered (Rosie message scenarios). Its text-a-link feature is a table of trigger phrase, message and URL, and the agent always asks permission before sending (Rosie text a link). A photo request can be built on that same pattern.
Two findings cut the other way. First, the job-software vendors are not as far ahead in the UK as their press suggests. Powered Now's Receptionist is a message-taker like ours: its current product pages label the link to its own customer database and diary as "coming soon", even though its July blog and September press release said it books into the diary (Powered Now landing page; Powered Now blog). It sells at £29 or £39 a month with unlimited calls, which makes it the UK price anchor (Powered Now Receptionist). ServiceM8's agent submits a booking request for staff review and does not write to the schedule (ServiceM8 help centre). No standalone UK product lists Powered Now, Tradify, Commusoft, ServiceM8 or Xero as an integration, and Jobber is "coming soon" at Callout AI (Callout AI). A standalone agent that books into any calendar is at parity with or ahead of the UK's native trade-software receptionist today. Second, our emergency handling is ahead of much larger products: ServiceTitan's own release note admits that until July 2026 its agent "treated all calls the same regardless of urgency" and booked gas leaks as routine jobs (ServiceTitan release hub). Only Answrd describes the same combination we run: fixed safety text spoken word for word, postcode read-back, and Ideal Postcodes lookup (Answrd privacy).
Seven UK stacks and two North American ones are disclosed, mostly through privacy policies. Answrd runs Retell AI on Telnyx with Anthropic and Cartesia, the nearest to ours; Whoza runs Retell with ElevenLabs voices on Twilio; Callout AI runs its own LiveKit pipeline with Deepgram, OpenAI, ElevenLabs and Twilio; AiRingDesk uses Claude, Deepgram and Twilio; AI Answerphone uses Google Gemini (Answrd privacy; Whoza privacy; Callout AI privacy; AiRingDesk; AI Answerphone). Jobber's sub-processor list names Vapi, LiveKit, Anthropic, OpenAI and Azure (Jobber sub-processors). Powered Now, Housecall Pro, ServiceTitan, Workiz, Goodcall, Smith.ai, Sayora and TradeVoice name nothing, so their builds are unknown.
Rosie's history is the clearest account of how a successful product was built. It started with its own orchestration of separate transcription, voice and model services, moved to Vapi, then to Bland as a single vendor; founder Jordan Gal said "we were coordinating and orchestrating all these services, and it didn't feel like a good long-term solution" (Bland). He had earlier said plainly that the company does not do the transcribing, the voice or the model, and that the value is the application layer; a team of about six went from start to paying customers in roughly four months (Startups for the Rest of Us). The application layer is where the disclosed engineering effort went: onboarding from a Google Business Profile (about 80% of Rosie's sign-ups), a sample agent that speaks the prospect's own business name before an account exists, 18 CRM connectors that log a summary after the call without booking anything, and a rule that owner-entered knowledge is treated as facts with instructional text stripped on publish (Startups for the Rest of Us; Rosie CRM integrations; Rosie business information). That facts-versus-behaviour split is the same one we already keep between the profile and the base prompt.
Where rivals book, they do not build a scheduler. Jobber books through its existing Online Booking form, FieldPulse through its Booking Portal availability, ServiceM8 through its online booking form (Jobber help centre; FieldPulse help). Rosie offers two levels: text a scheduling link on every plan, or create the event directly in Google Calendar, Calendly, Acuity or Appointlet on higher plans (Rosie scheduling). The cheap first step for us is the link; the second is a direct write to one calendar.
Quality tooling has become a product surface. Goodcall's builder writes practice callers, lets the owner flag a bad call and say what to change, holds edits as a draft until publish, and can roll back (Goodcall how it works). Smith.ai sells simulated scenarios by plan and drafts an FAQ entry for human approval whenever the agent could not confidently answer (Smith.ai). FieldPulse has a Test Mode that creates no records and a conflict check on owner-written rules (FieldPulse help). The only detailed in-call engineering account comes from a single former Avoca engineer's personal blog and is unverified: a state object re-injected every turn, confirmation when transcription confidence is low, hand-off after three failed attempts at one field, a pronunciation dictionary of about 200 trade terms, and an 850 ms average to first word (personal blog). No vendor publishes measured turn latency.
Independent review evidence is thin: Rosie and Upfirst have no Trustpilot reviews, no review of any kind was found for Whoza, Callout AI, TradeVoice, Answrd or BT, and Reddit, G2 and Capterra were not readable. No independent head-to-head test with call counts exists. What follows rests on Jobber's community forum (read through search snippets), two vendor-sponsored OnePoll surveys, Upfirst's own call data and one consultant-run trial at a Kent heating firm.
The sharpest trades complaint is about content, not speech. A plumbing owner on Jobber's forum reported that the receptionist took a name and contact details but never asked what was wrong, so bookings arrived with no job description; a tree-service owner turned it off because it "would mismatch clients, or not link to an existing client" and produced "more dropped balls" (Jobber community). In the same forum an owner said it made callers "hang up or get angry due to the issues with taking information", and Jobber's reply was that there is no way to train the receptionist (Jobber community). Dialzara's five-star reviews praise the opposite: "teaching it, correcting it, and refining how it responds didn't take much time at all" (Trustpilot, Dialzara). Ease of correction keeps customers; its absence sits beside the hang-up complaints.
On the caller side, the OnePoll survey of 6,000 adults for AnswerConnect, a seller of human answering, found that 41% of people who had spoken to an AI tool named repeating information already given, 62% distrust AI to relay information accurately when contacting a tradesperson in an emergency, and 31% say they would hang up on AI (AnswerConnect). Observed behaviour is milder: at the Kent firm 6 of 61 callers hung up on realising it was not a person, and 43 of 61 otherwise-missed calls became bookings (Lilach Bullock). That trial's failures were steering complaints back to scheduling, cutting in on slow-speaking elderly callers, and following the calendar literally on urgent calls; an instant transfer on "leak", "no heat" or "emergency" to the on-call mobile "fixed most bad experiences" (Lilach Bullock). UK press on the EMMA GP receptionist documents the accent risk, with patients hanging up because it could not understand a broad Yorkshire accent (Silicon UK).
Upfirst's analysis of 450,702 calls across 503 receptionists is correlational vendor data, but it is the only large dataset. Agents with booking, transfer, texting and business-specific training had roughly 30 to 40% lower odds of a hang-up; disclosing that the agent is an AI went with about 20% fewer hang-ups; voice choice made no measurable difference; and agent settings explained only about 13% of the variation, with many hang-ups coming in the first seconds (Devdiscourse). Three sources with different interests agree on disclosure: the Upfirst data, the Kent test, and 86% of survey respondents who want AI clearly indicated (ServiceForge). On review platforms the one-star reviews are mostly about money and support, such as charges after cancellation at Goodcall and My AI Front Desk, not about the call (Trustpilot, Goodcall; Trustpilot, My AI Front Desk).
Holding captured details in a structured record outside the chat history, and letting code decide what is still missing, is the established fix for the re-ask bug. Evalgent, a testing vendor, defines the bug as an agent whose "state never advances" and prescribes storing slot values outside chat history, re-injecting them every turn and enforcing per-slot ask counts in code (Evalgent). Our specific failure, where the caller answers and asks something in the same turn and the agent answers the question then re-asks, matches Evalgent's "context" cause: the answer is in the transcript but nothing marks it captured. FormBharo, an August 2026 paper on a phone form-filling agent, is the closest published system to the plan: a language model paired with rule-based validation and flow control, in which "the rule-based controls recover many turn-level extraction errors" and let smaller, cheaper models match or beat frontier ones (arXiv). That supports keeping Haiku 4.5 and adding deterministic control. Pipecat's August 2026 benchmark agrees on the model: Haiku 4.5 passed 98.0% of graded turns at a 637 ms median first answer token, among the strongest inside the roughly 700 ms budget, though its 95th percentile is 1,615 ms (Pipecat benchmark).
Putting our own service in the model path, instead of using ElevenLabs' native Update state tool and dynamic variables, also has a stated rationale. The native route depends on the model choosing to call the tool; Zep's ElevenLabs guide, which documents the same proxy architecture, says the model "may skip retrieval when it shouldn't" and that a proxy runs "on every request, without LLM involvement" (Zep; ElevenLabs Update state). The native route is still the lighter option with no service to keep up, and it is the right fallback if the pre-build tests below fail. There is no public account of anyone running a receptionist on ElevenLabs with this proxy in production, so uptime and latency for the exact design have no precedent.
Fallback. With a custom model as primary, ElevenLabs bypasses its hosted fallback by default: the same endpoint is retried at least three times, the per-attempt wait defaults to 4 seconds, and a sustained failure ends the conversation (ElevenLabs LLM cascading). The backup preference must be set to override with an ordered list of hosted models, and because the backup runs the raw system prompt with no record, the base prompt has to remain able to carry a whole call unaided.
Choose the next question in code; do not block and regenerate. No source describes blocking a reply and regenerating it inside a live voice turn, and doing so doubles model latency for that turn. FormBharo's design extracts every field answered in the turn into a form state that is the single source of truth, then picks the next question by rule and tells the model what to ask (arXiv PDF). The extraction belongs in a separate or parallel call. If the design includes a hidden note written before the spoken reply in the same completion, it delays first audio by the note's length, and it must be stripped or sent in the reasoning_content field or it will be spoken (ElevenLabs custom LLM). No source measured that delay; the rig should.
Cap asks at two, then change method. Blocking a re-ask is half the remedy. Evalgent's rule is "two asks, then change strategy": re-ask with structure, then spell-back, then move on or hand off, and never loop on a value that has failed twice (Evalgent). For us the third step is to move on and flag the missing detail on the owner's card. This rests on one vendor blog whose author calls the thresholds starting points, plus the unverified Avoca three-strikes rule.
Test on real transcripts, many times. FormBharo found that form completion fell by up to about 41 points when the model received real-speech transcripts instead of clean reference text, and that component scores do not predict end-to-end results (arXiv). Replayed calls must use the actual speech-to-text output, errors included. Each scenario should run at least ten times with a pass count, and the re-ask check should be a deterministic count of asks per detail taken from the transcript, not a judge model's opinion (Evalgent). A text rig that calls our service directly cannot see turn-taking, barge-in or tool execution, and Hamming's guidance notes that voice can behave differently from chat (Hamming). ElevenLabs already ships simulation, next-reply and tool-call tests, creation of a test from a real conversation, 2 to 20 repeat runs with a pass rate, and a command-line runner for CI (ElevenLabs agent testing). A small native suite plus the existing handset call covers the layer the rig misses. The scenarios should include the cases owners actually complained about: job reason captured, a business caller, a complaint, a slow speaker, an urgent call, and a caller who answers two questions at once.
Verify three things before building. ElevenLabs' custom model page says system tools are passed in the standard tools parameter but does not mention webhook tools; a third-party gateway guide says custom tools work and warns that a plain-text connection test proves nothing about tools (ElevenLabs custom LLM; API7). So test that both of our webhook tools arrive and execute through the custom endpoint, that a per-call identifier reaches the endpoint (Zep reports the extra-body setting must be enabled under Security overrides), and what the caller hears when the endpoint times out (Zep). The route also rules out ElevenLabs' realtime speech-to-speech models, which costs little: independent evidence still shows them weaker on tool use and rule-following (Hamming).
The service opens two further gains. Ideal Postcodes reports a 20 ms median response, so the service can run the postcode lookup itself whenever a transcript contains something postcode-shaped and place the result in the record, removing the second model pass a tool call costs (Ideal Postcodes). The safe form is validate and confirm by place name, never silent replacement, which is consistent with why check_details was removed; PolyAI's published nearest-valid-postcode correction is for a service with bounded geography (PolyAI). The same service is the natural home for returning-caller memory, keyed on caller number, as Zep's guide shows (Zep).
Evidence strength is graded as follows. Strong: first-party documentation or a published study, backed by at least one other independent source type. Moderate: documentation from several vendors, or one dataset or trial with a stated method. Weak: a single unverified source, a search snippet, or inference. Booking and pricing recommendations apply to client agents; the rule that demos never book or price is unaffected.
| No. | Adopt | Why | Evidence | Decision |
|---|---|---|---|---|
| 1 | Per-call record outside chat history; separate extraction of every detail answered each turn; code picks the next question | Published fix for the re-ask bug; rules let a small model match larger ones (Evalgent; arXiv) | Strong for the pattern; no measured before-and-after figure exists | First build |
| 2 | Hosted fallback set to override; base prompt able to run a call unaided |
Default is three retries then the call ends (ElevenLabs) | Strong | First build |
| 3 | Pre-build tests: webhook tools through the custom endpoint, per-call identifier, timeout behaviour | Docs are silent on webhook tools (ElevenLabs) | Moderate (a documented gap) | First build |
| 4 | Test rig: at least 10 runs per scenario, deterministic ask count, replay on real speech-to-text output, small native ElevenLabs suite, handset call | 41-point drop on real transcripts (arXiv; ElevenLabs testing) | Strong | First build |
| 5 | Two asks per detail, then change method, then move on and flag to the owner | (Evalgent) | Weak to moderate | First build |
| 6 | Emergency live transfer after the safety line: owner-set trigger words, spoken briefing, connect on a spoken yes, voicemail detection, fall back to message and alert text | Fixed most bad experiences in the Kent trial; 62% distrust AI in a trades emergency; gated as the paid upgrade by three UK rivals (Lilach Bullock; Rosie) | Strong | Next, 1st |
| 7 | State up front that the caller is speaking to an AI; test the opening line hardest | About 20% fewer hang-ups; most hang-ups come in the first seconds (Devdiscourse) | Moderate | No change |
| 8 | Patient turn-taking while collecting addresses and numbers; a fallback for callers the agent cannot understand | Cut-ins on slow speakers; accent failures; ElevenLabs recommends patient eagerness for addresses (ElevenLabs; Silicon UK) |
Moderate | Later |
| 9 | Calendar booking in two steps: text a booking link first, then direct write to Google or Outlook | 11 of 14 UK products; ahead of Powered Now as shipped (Rosie; Powered Now) | Strong on prevalence; moderate on effect (correlational) | Next, 3rd |
| 10 | Owner correction loop: flag a call, edit facts and FAQs as a draft, publish, roll back; behaviour stays in the base prompt | Named reason for staying; its absence sits beside hang-up complaints (Trustpilot, Dialzara; Goodcall) | Moderate | Not now |
| 11 | Knowledge base built from the website or Google listing, treated as facts only | The norm for onboarding and the instant personalised demo (Sayora; Rosie) | Moderate (vendor documentation; times are vendor claims) | Next, 2nd |
| 12 | Mid-call text link with spoken consent, including a photo-upload link | Rosie's trigger-and-URL pattern; photo link live at one UK rival (Rosie; Arcscribe) | Moderate for links; weak for photos | Later |
| 13 | Postcode lookup run by the service in parallel; validate and confirm by place name; owner-set included and excluded service areas | (Ideal Postcodes; FieldPulse) | Moderate | Later |
| 14 | Returning-caller memory as a short, owner-editable brief per number, with a stated retention period and a mismatch safeguard | Retell and Upfirst store briefs, not transcripts; Jobber owners report mismatches; ServiceTitan escalates on number mismatch (Retell; PolyAI docs) | Moderate on design; weak on value | Later |
| 15 | Job-software logging after the call (summary to the matching record, no booking), starting with one UK tool | Rosie's 18 connectors work this way; no standalone UK rival offers it (Rosie CRM) | Moderate | Later |
| 16 | Non-intake paths for complaints and business callers | Kent trial and one forum thread (Jobber community) | Weak to moderate | Later |
| 17 | Small copies: known-number bypass list, a test mode that sends nothing, an in-product divert test, a per-trade pronunciation list, a monthly value report | (Powered Now; FieldPulse; Sayora; Answrd) | Weak (vendor pages only) | Later |
Four things the research says to leave alone. Voice selection does not move hang-ups in Upfirst's data, so voice testing is low value (Devdiscourse). Open web search during a call has only search-vendor evidence behind it. A larger model is not indicated when rule-based control recovers the errors. Tone-based escalation is contested: Rosie refuses to transfer because a caller sounds upset, while Moneypenny and ServiceTitan do (Rosie; Moneypenny). Invented pricing is offered by nobody; if price answers are added later, the market model is owner-approved figures only.
The competitive picture is narrower than the feature lists suggest. Rivals are ahead on what the agent can do after it understands the caller, and nearly silent on whether it understands the caller correctly; only Answrd documents address checking, and no vendor publishes a re-ask rate, a postcode accuracy figure or measured turn latency. Owners, meanwhile, judge by listening to their own customers on recordings and switch off agents that capture a name without the job. That makes the state service and the test rig a product in their own right, not just a bug fix: a method-stated run of fixed scenarios with pass counts would be the only such evidence in the UK market, and nobody has published a mystery-shop of the five callable UK demo lines.
Timing matters on one item. Powered Now's job-software sync, the one thing a standalone agent cannot match, has not shipped, and no standalone UK rival writes to trade software at all. Calendar booking and a single after-call logging connector, added once the state service is stable, would put the master agent level with or ahead of every UK product read here while its address and emergency handling remain the part others have not copied. The main unknowns to close first are the three ElevenLabs behaviours that the documentation leaves unstated, since they decide whether the proxy route or the native Update state route carries the plan.