Industries · Appliance repair
Everything you need to know is on the plate. None of it is anywhere you can act on it.
Corveonic builds the back office for independent appliance shops, weighted toward built-in and premium-brand work: claims reconciled from submitted to banked, an appliance registry that can turn a serial into a date, and a co-pilot beside the person who answers your phone. Nothing we build speaks to your customers.
The serial tag
One plate, read once, on every visit
A technician kneels at a toe kick and reads a plate. Model and serial go onto the work order, because the claim requires them. That is the whole transaction, and it has happened on every visit your shop has ever run.
Consider what those two strings already tell you. The model. The build date, derived from the serial. The warranty tier the machine sits in today, and the date it crosses into the next one: the day the labor stops being the manufacturer's problem and becomes the homeowner's. What failed on this model before, across every job in your history. Which part it took, and whether that part still orders under that number.
Then the turn: none of that is anywhere your shop can act on it. Four structural reasons, and none of them is anyone's fault. Dispatch systems are organised around the job, not the appliance; the machine is free text on a work order, spelled differently every time. Nothing in the stack has a horizon measured in years, so knowledge with a five-year fuse decays into history. The event that matters is a non-event: nothing happens on the day coverage lapses, no ticket, no alert, no call. And the certification is treated as a cost of doing business, never as a list.
The clock argument, stated as an observation: premium built-in brands publish tiered warranties, and the tiers change who pays as the appliance ages. Early on, the manufacturer covers sealed-system parts and labor. Later, the parts stay covered and the labor does not; the homeowner pays, and the work must go to a factory-certified shop. Every appliance a certified shop touches under warranty today is a scheduled, high-ticket, cash-paying repair some years out, at a date computable from the serial tag.
As far as we can tell, almost nobody tracks it. That is our read of the trade, not a study, and you can check it against your own dispatch history this afternoon.
The leaks
Three leaks, and where each one hides
The national average appliance repair is $179, on a survey of 30,000+ customers. If you run built-ins, that number describes nothing you do, and the distance between it and your sealed-system ticket is the whole reason a tool priced and designed against a $179 job is the wrong tool for your shop.
Source: HomeAdvisor appliance repair cost guide — survey of 30,000+ customers · verified 2026-08-01
The lead market prices the trade the same way: appliance-repair leads sell for $22 to $85 on a published rate card, against $65 to $325 for HVAC and $85 to $550 for roofing. Everything sold to this trade assumes low tickets and high volume. Your economics are the opposite, which is why your leaks are different too.
Source: Service Direct published pay-per-lead rate card · verified 2026-08-01
01
The truck roll
The intake call is where the truck roll is decided, and it is the least-instrumented moment in the business. A homeowner describes a symptom in homeowner language; from that one sentence someone infers a model, a fault class, a part, a window and a technician, from memory, in ninety seconds. The customer reports the loudest symptom, so the door gasket and the dead icemaker surface on site, and now the window is over-run or the job is under-scoped. The knowledge that would prevent it lives in one or two heads, and it leaves the building at five o'clock.
02
The claim
Submitted is not paid. A claim has a life after it leaves the building: acknowledged or not, approved in full or short, rejected on paperwork or on the merits, pending for ninety days. Settlements arrive as batch deposits that bundle many claims, so a short-pay is quieter than a rejection: it is just a smaller number inside a batch. And the person who would notice is measured on getting claims out, not on auditing what came back. The work has no owner, so it does not happen.
03
The clock
The customer whose labor coverage lapsed fourteen months ago never got a call, because nothing rang. The transition from covered to out-of-pocket is silent by construction; only a system that computes it will ever see it. When she called somebody, it was whoever answered. The date was in your records the whole time.
The refusals
What we won't tell you, and what we won't build
This page has fewer numbers than its siblings, by design: most of the figures circulating in this trade do not survive being traced. Here is what we refuse to say, and refuse to sell.
We won't tell you what a sealed-system job is worth. The figures in circulation come from individual shops' price pages in other states. You know your book better than any of them.
We won't tell you your industry's first-time-fix rate. The 75% average everyone quotes traces to marketing blogs citing an analyst report nobody can produce. We'll measure yours.
We won't tell you what your missed calls cost. No credible source exists for that number, and the ones being quoted at you were modelled by companies selling phone products.
We won't promise a booking rate, because we're not putting a bot on your phone.
We won't tell you we'll automate your warranty claim submissions. Your system already sends them. We'll tell you which ones came back short.
We won't make earnings claims. In March 2026 the FTC permanently banned the owners of an AI sales-automation company from marketing business opportunities over exactly those claims.
We'll run a 20% holdout for 60 days on anything we claim credit for. It makes our numbers look worse, and it's the only way either of us learns the truth.
We have no case studies to show you, and we'd rather say that than invent one.
And the standing refusals, product by product:
A warranty claim submitter. Your system already sends them.
Portal automation using your credentials. Your authorization is the business.
Review gating, in any of its shapes.
A blast to a list you have not messaged in years, including the list the warranty clock just produced.
An AI voice on the phone with a homeowner.
A ranked technician leaderboard in month one.
An OEM manual ingestion. Access is not ingestion, a retrieval index is a copy, and a contract can forbid what fair use might allow.
A diagnosis of a sealed system from a photograph. That is physics, not a roadmap.
A dollar figure for a sealed-system job, a first-time-fix benchmark, a missed-call cost, or an earnings claim of any kind.
The enforcement posture: FTC press release, 24 March 2026 — Air AI owners permanently banned from marketing business opportunities · verified 2026-08-01
The work
What we actually build
Ten items, in dependency order rather than sales order, because some of these gate the others and pretending otherwise is how projects fail in week six. Each carries its caveat.
Claim reconciliation: billed against banked
Join claim to job to invoice to settlement, on authorization number where it exists and serial where it doesn't. Classify every claim: paid in full, paid short, rejected on paperwork, rejected on the merits, pending and aged, or unmatched, and unmatched is never quietly dropped. Billed against banked, displayed as an arithmetic identity you can check against a statement by hand.
The honest caveat: it starts with a paid discovery that can end in no, because neither warranty portal in this category publishes a self-serve API; expect exports, not silent integration. We do not log into manufacturer portals as you. And it will not recover a claim rejected on the merits; nothing will. This is the strongest first project because the output is verified against your own bank deposits.
Pre-submission claim QA
A validation pass at job close, before the claim transmits: required fields, authorization format, part on the PO reconciled against the invoice, superseded numbers, date inside the window. On failure the claim is held, not blocked, with the specific field named.
The honest caveat: the rules come from your own rejection history, which means this is genuinely second, not first. A vendor quoting it in month one is quoting from a template. It will also make visible how many claims were being rejected for paperwork, which is a conversation about a person's work; have it deliberately.
The warranty clock: registry and worklist
A customer-and-appliance registry, the entity your dispatch system does not have; a tier-clock function from brand, model family and date to a crossing date; and a worklist of appliances crossing into the labor-not-covered window, sorted on a contactable field computed from your consent records and the 18-month relationship window.
The honest caveat: the backfill is most of the first project, and the registry is only as good as the dates in your history, so a data-quality view ships on day one, not on request. The clock tells you who to contact. It does not tell you that you may: the customers it surfaces are usually outside the relationship window by definition, and some of them a competitor already took. We will not multiply the count by a conversion rate or a job value, because neither number exists.
Plate capture
The job cannot close without the plate photo, and the technician can see why: his claims get approved and his callbacks get defended. Vision reads model and serial, validates against the known serial format, and anything low-confidence goes to a correction queue worked by a person.
The honest caveat: OCR on a corroded plate behind a toe kick fails often; the queue is permanent, and it is the product, not the fallback. A wrong serial is worse than a missing one, so nothing low-confidence writes straight to the registry. This is the only item here that needs new technician behavior, which is why it is never the first project.
The CSR intake co-pilot
While your CSR talks, it reads the plate, pulls the closest past calls on that model and complaint from your own history, generates the right questions, including the cheapest fix in this whole list: is there anything else wrong with it?, and produces a ranked parts pick-list with a confidence band and an explicit abstention when the fault smells sealed-system. It routes: longer window, certified technician, load the gear.
The honest caveat: it cannot diagnose a sealed system from a phone; that is physics, not a roadmap item. It is weakest exactly where it is most valuable, the first-time caller with no history. And we do not ingest OEM service manuals: they are contract- and copyright-gated, and a vendor who says they have the manuals in their product is telling you something about how they will treat your data too. Without the confirm-or-edit step at job close, this build should not start.
Parts staging, lead time, and NLA
Usage normalised through a supersession map; lead times measured from your own ordering history, order to in-hand, never quoted by a supplier; two buckets, stock on the truck and pre-order on diagnosis; and an NLA watch that flags a part as unorderable before a technician discovers it in a kitchen.
The honest caveat: no stocking level fixes a part you cannot buy; for long-lead and NLA parts the deliverable is a sourcing process and a customer conversation. Any vendor who models your truck stock without asking about lead times has not worked in this trade.
First-time-fix and callback measurement
The definitions first, in writing, because a callback is a return trip on your own work and a recall is the manufacturer's word for a defect campaign, and your office and your techs almost certainly disagree about the edge cases today. Then auto-detection on same serial within a set number of days, which is cleaner in this trade than in any other, with a caused-versus-ran split.
The honest caveat: this is a scoreboard, not an intervention; sold alone, it lands as blame. And we will not publish an industry benchmark beside your number, because the benchmarks in circulation do not survive being traced.
Review requests, on paid invoice
A suppression gate first: return trips, warranty jobs, open complaints, recent asks, opt-outs, and anything the technician flagged. Then one message from your own number, naming the technician, linking straight to the review dialog. One reminder. Then stop, permanently. Responses drafted for a person to approve, because half of consumers are put off by templated replies.
The honest caveat: the one real benchmark here is academic and general, not appliance-specific: responding to reviews raised ratings by 0.12 stars and volume by 12% in a peer-reviewed study of businesses generally. On Google we ask everyone, unconditionally; on Yelp we don't ask at all, because Yelp's rule is the opposite. You will get more one-star reviews too. We say that before go-live, in writing.
Technician reporting, with counterweights
Per technician: return-trip rate on the caused split, first-time fix, plate-capture rate, claim rejections on their jobs, reviews, on-time arrival, with job mix displayed above the metrics, because your sealed-system specialists carry the highest tickets and a naive ranking just rediscovers job mix and calls it performance.
The honest caveat: this is the worst first project in the trade, and we decline to start here even when asked. No ranked leaderboard in month one, not behind a toggle. Every metric carries a counterweight or it is not published.
Missed-call follow-up, honestly
The honest recommendation first: this is a commodity. One platform vendor publishes an AI receptionist add-on at $29 a month; if that solves it on the software you already pay for, buy that, and we will say so before we quote you. What earns a custom build is the layer underneath: one customer lookup, one de-duplication rule, one shared opt-out ledger serving every automation on this page.
The honest caveat: a text into an inbox nobody watches at six on a Friday is worse than a missed call, and a homeowner standing in front of a leaking refrigerator wants a voice. If nobody will answer the replies, this should not be built, and we will not build it.
The review benchmark: Proserpio & Zervas, Marketing Science (2017) — peer-reviewed, causal identification · verified 2026-08-01
The $29 add-on: Jobber published pricing (AI Receptionist add-on; free on the Plus plan) · verified 2026-08-05
Worksheet
The arithmetic, on your numbers
Every field starts empty, because defaults are claims. Every number in the arithmetic is on this page and editable. And it can return an unimpressive answer: if the gap comes out near zero, your office is ahead of most, and we would rather tell you that than sell you a build.
Nothing you type leaves this page. There is no request, no storage and no analytics on these fields; you can verify that in your browser's network tab.
Schedule A · the return leg
From your own board and your own books. We will not supply any of these, and no industry figure for them survives tracing.
- 1
- 2
- 3
- 4
- 5What return legs cost you in the period(Line 1 × Line 2 × Line 3 × Line 4)
Line 1 × Line 2 × Line 3 × Line 4 = —
A cost you already paid, not a number we promise to remove. The fraction a pre-screen removes is measured against a 20% holdout, never assumed, and for long-lead and NLA parts it is structurally capped.
Schedule B · billed against banked
Take one closed quarter. Line 5 comes from your dispatch system's invoice extract; line 6 from your bank feed or distributor statement.
- 6
- 7
- 8The gap(Line 6 − Line 7)
Line 6 − Line 7 = —
We do not know what this is for your shop. Nobody does, including any vendor quoting you a warranty-leakage percentage; no defensible industry estimate exists. Decompose it, short-paid, rejected, pending, unmatched, before you conclude anything.
For scale, beside your result rather than instead of it: Corveonic retainers are published at $597, $1,197 and $1,997 a month plus a one-time setup fee.
Client-side only. Your numbers never leave your browser; we never see them.
The design position
Why there's no robot on your phone
Your call is the first time a stranger asks to come into someone's kitchen. For this kind of work we build the thing that sits next to whoever answers: the tool prompts the right diagnostic questions for that model and symptom while she talks, pulls what the shop has seen before on that model and complaint, and helps get the right part on the truck the first time. The human talks. The tool remembers.
The reasons are public, not sentimental. 75% of consumers say they prefer a human for customer service, and Klarna publicly walked back an AI-only support strategy and re-hired people. AI-disclosure rules are tightening in step with adoption, and a design that never needs a disclosure exemption ages better than one that does. And the category's own honest benchmark is about a 70% booking rate, published by the one vendor that discloses both its adjusted and unadjusted numbers. Selling a voice product into a trade where a missed nuance means a wrong part and a second trip is selling the wrong 30%.
To be precise about the claim: voice AI demonstrably works for some shops in adjacent trades. This is the design we chose for this work, with the reasons stated. Not a claim that it can't be done.
The 75%: Five9 study, via Business Wire, October 2024 · verified 2026-08-01
The 70%: ServiceTitan's own published figures for its AI voice agent — 90% “capacity adjusted,” 70% overall · verified 2026-08-01
The incumbent stack
Submission was solved. Reconciliation wasn't.
Your dispatch platform already generates warranty claims on job close and transmits them to the manufacturer portals in batch, and it has done so for years. It is good at it. Any vendor offering to automate your claim submission is telling you they have never stood in a shop.
What nothing in the stack tells the owner is which claims came back rejected, which were paid short, which have been sitting in pending for ninety days, and what the difference between billed and banked came to last quarter. The claim leaves the building, and the money either arrives or quietly doesn't. That gap, between submitted and paid-what-was-owed, is where we work.
We could not find a claim-lifecycle product serving shops in this band; if your platform has one, turn it on, and skip us for this. The dashboards above are what we build when it doesn't.
Compliance
The rules we build to, and the homeowner's front door
Your customer file is a physical-access list
An appliance shop's database is a list of affluent households, their addresses, when they are home, which door to use, and increasingly the code to get in. That is a materially different security object from a marketing list, and it deserves to be treated as one: gate codes, alarm codes and access notes are restricted fields in anything we build, never in a model prompt, never in an exported spreadsheet, never in a message body.
Vetting us is your legal obligation, so we make it easy
New York's SHIELD Act applies to any business holding a New York resident's private information, with no size threshold, and it requires reasonable safeguards including vetting service providers by contract. You get the data-processing agreement, the named sub-processor list, the no-training commitment and the breach-notification path in writing, because that paperwork is how you document your own diligence.
The technician is federally licensed, and the software is not
Sealed-system work means refrigerant, and refrigerant means EPA Section 608 certification under 40 CFR Part 82, Subpart F. A diagnostic suggestion is an assist, not an authority: anything touching the sealed system defers to a certified technician with gauges on the machine. That is a federal rule, not our policy.
The standard message rules, unchanged
Quiet hours 8am to 9pm at the customer's local time, enforced as a hard send window. Opt-outs honored across every automation from one shared ledger, records kept. Even the photo request, the most useful intake step in this trade, is a consented, windowed outbound message, and we build it as one. The 18-month relationship window bites harder here than anywhere, because your repeat cycle runs in years: the customer the warranty clock surfaces is usually outside it, which is exactly why the clock produces a worklist and never a blast. Review gating is prohibited and we won't build it; on Yelp we don't ask at all. Where any automation writes to a customer, it is disclosed. Always.
The certification: EPA — Section 608 technician certification (federal regulation) · verified 2026-08-01
The message rules: 47 U.S.C. § 227; 47 CFR 64.1200 (TCPA and FCC rules) · verified 2026-08-01
What happens to client data is stated in plain language in our privacy policy.
Read the privacy policyFAQ
Asked by owners, answered without a script
Are you going to put an AI on my phone?
No, and we won't sell you one. Your call is the first time a stranger asks to come into someone's kitchen. We build the thing that sits next to whoever answers.
My software already submits my warranty claims.
It does, and we won't pretend otherwise. We're not building a submitter. We're building the part nobody built: which claims came back rejected, which came back short, which have been pending since March, and what the gap between billed and banked actually was.
Can you get the manuals into it?
No. OEM service documentation is contract- and copyright-gated, and we won't bulk-ingest it. We ground the co-pilot in your own job history, which is the more valuable corpus anyway, because it knows what actually fails in the field on the models you see.
Can it tell what's wrong from a photo?
Sometimes, and we'll be specific about when. On the inside of a built-in, treat it as a hint that helps you ask better questions, not a diagnosis. A federally certified technician with gauges makes the call on anything sealed.
How much of my customer data do you hold, and where?
As little as the workflow needs. Access notes, gate codes and alarm codes are restricted fields: never in a model prompt, never in an export, never in a message body. You'll get the DPA, the sub-processor list and the no-training commitment in writing, because under the NY SHIELD Act vetting us by contract is your obligation, and we should make it easy to discharge.
Read the privacy policyHow long until something works?
First automation live in two to four weeks, and it will need no new behavior from your techs. Your data sets the timeline, not our code: on the warranty registry specifically, the backfill is usually the project. We'll tell you which you are in week one.
How will I know it worked?
Baseline from your own system, in writing, before we build. One metric per automation, measured at the invoice, or for claims, at the deposit. And a 20% holdout for 60 days.
What if we part ways?
Everything runs in accounts registered to you, and your phone numbers and 10DLC brand registration stay in your name. Your customer and appliance registry is yours, exportable, on the way out as well as the way in.
What this costs: retainers run $597 to $1,997 a month plus a one-time setup fee, published in full. Which tier fits depends on what the discovery finds, and we'll say so plainly.
See the tiersOr run your own numbers firstBring one closed quarter's claims.
We'll reconcile billed against banked, and you can check the result against your own deposits. If the gap is small, we'll say so, shake your hand, and save you the retainer. That answer is free too.