Every trades shop that has deployed AI for more than 60 days has a story. The Avoca booking that promised free dispatch the shop does not offer. The AI-drafted estimate that listed a SEER2 rating two points higher than the actual equipment. The customer-recovery email that committed to a refund the owner never approved. The hallucinated NEC reference in a panel-upgrade proposal. The fabricated warranty year on a Goodman coil. The Section 25C credit number that was right last year and wrong this year. Each of those stories cost somebody money — sometimes hundreds of dollars, sometimes tens of thousands, sometimes a license investigation and a Reg Z lawsuit. This lesson is the failure-mode atlas. Not theoretical "AI might be wrong." Specific 2026 ways AI gets trades shops sued, fined, audited, and humiliated. Read it once. Tape the list to the inside of every CSR's monitor. Build the 30-second verify habit before you build anything else.

Failure One: Hallucinated Part Numbers and Equipment Specs

The CSR or tech asks the AI: "What's the part number for the inducer motor on a Goodman GMVC96 furnace?" The AI returns a specific-looking part number — let's say 0131M00135S — confidently formatted, no hedging. The tech orders it through Carrier Enterprise or Johnstone. The part arrives. The part is wrong. The job is a wasted trip. The customer is on the phone wondering why their no-heat call is now in day three.

This is not a hypothetical. Part numbers are exactly the kind of data where hallucination is structural. The AI's training corpus contains millions of references to Goodman, GMVC96, inducer motors, and 0131M00135-style part-number formats. When asked, the model produces the statistically most likely part number — which means a number that looks like a real Goodman part number, contains roughly the right number of characters, follows the alphanumeric convention. It is almost certainly not the actual current part. Goodman's actual part numbers change with revisions; the AI's training data is months or years stale; cross-reference lookups against the actual current supplier catalog never happened.

The fix: AI never orders parts. The AI can suggest, draft, and summarize. The actual part number comes from the manufacturer site, the supplier catalog (Carrier Enterprise, Johnstone, RE Michel, Munch), or the part-lookup database the FSM platform integrates with. The 5-second verify on any AI-suggested part is to cross-reference against the supplier catalog before the order goes out. The same rule applies to refrigerant matches (the R-454B vs. R-410A confusion in 2026 is catastrophic when mismatched), filter sizes (an off-by-one inch dimension wastes the trip), and breaker amperage (an AI suggesting a 30A breaker for a 40A circuit fails inspection and creates a fire-risk citation).

Failure Two: Fabricated Warranty and Rebate Terms

The Comfort Advisor at a Texas HVAC shop runs an AI prompt: "Draft the warranty language for a 16-SEER2 Trane XR16 condenser install with the homeowner's 10-year parts and 2-year labor coverage and ConsumerCare for the parts." The AI produces fluent paragraphs that sound exactly like a warranty disclosure. The paragraphs say "12-year compressor coverage" instead of 10. The paragraphs say "lifetime heat exchanger" on a 2026 model that carries a 20-year manufacturer term. The advisor pastes the language into the proposal. The customer signs. Six months later the compressor fails in year 11 and the homeowner is on the phone with a copy of the proposal saying year 12. The shop eats $4,200 in compressor + labor.

Same failure mode applies to rebates. AI confidently quotes utility rebate amounts that were correct in 2023 and are different in 2026. Confidently quotes federal Section 25C credit caps that were right under one IRS guidance memo and superseded by a later one. Confidently quotes state heat-pump rebate qualification thresholds that vary by ZIP code and customer income bracket — variables the AI does not have access to.

The fix: any AI-drafted warranty language gets verified against the manufacturer's actual term sheet (Goodman, Trane, Carrier, Lennox, Bryant, Rheem, York — each has a published current term sheet) before it reaches the customer. Rebate amounts come from the utility's current 2026 rebate page or from the state energy office's current heat-pump program page — not from the AI's training memory. The 30-second verify is not optional on regulated artifacts.

Failure Three: AI-Drafted Financing Language and Reg Z Exposure

This is the failure mode most likely to produce an actual lawsuit in 2026 trades AI. A Comfort Advisor asks the AI to "draft a financing summary for the homeowner with the Wisetack approval at $18K, 84 months, the APR they showed me, and the monthly payment." The AI produces a beautifully formatted summary. The APR in the summary is 7.99%. The actual Wisetack portal showed 8.99%. The advisor doesn't catch it. The customer signs the proposal. The actual loan documents from Wisetack arrive showing 8.99%. The customer claims bait-and-switch. The shop is in a Reg Z compliance event — federal Truth in Lending Act exposure, $500-$5,000 per-violation statutory damages plus attorney fees plus state-AG complaint exposure.

The fix is non-negotiable: AI does not draft financing language with specific numbers. Numbers come from the Wisetack / GreenSky / Synchrony portal output, copy-pasted verbatim. AI can draft surrounding context ("here's why financing fits this replacement," "here's how the monthly payment compares to the energy savings") but never the regulated number itself. The shop's financing-language template has a literal "[paste portal output here]" placeholder; the advisor pastes; AI never touches the regulated terms.

Adverse-action notices on soft-pull declines are an even tighter rail. FCRA requires specific language, score-range disclosure, credit-bureau identification, and notice of dispute rights. AI does not draft these. The lender's template, signed by the lender, is the only compliant artifact. Any shop letting AI generate the adverse-action notice has FCRA exposure — typically $100-$1,000 statutory per violation plus actual damages plus attorney fees.

Failure Four: Electrical and Plumbing Code Hallucinations

The Comfort Advisor on a panel-upgrade pitch asks the AI to "summarize the NEC requirements for a 200A service panel upgrade in a 1968 home in [Bel Air, CA]." The AI confidently produces a paragraph citing NEC 230.71, NEC 408.40, and NEC 250.122 — the right-sounding code sections. The advisor includes the citations in the proposal. The advisor is wrong on two of three citations. The shop's electrician, who actually pulls the permit, has to fix the proposal in the kitchen before the homeowner notices. If the homeowner already showed the proposal to a competing electrical contractor or the local building inspector, the shop's credibility takes the hit.

Code citations are structurally the worst-case AI failure mode. The NEC has been updated multiple times since most training data cutoffs. Local jurisdictions amend the NEC (and the IPC, the IMC, the IRC) with custom requirements. California's Title 24 layers on top of the NEC. Florida's building code amendments differ from Pennsylvania's. New York City and Chicago have their own code amendments. The AI's training corpus is national, generic, and stale. Citations confident in the training data may be wrong in your jurisdiction. Citations confident in the AI's output may be inverted, deprecated, or hallucinated entirely.

The fix: AI never cites code in customer-facing artifacts. The shop's electrician, plumber, or HVAC tech with the relevant license citation cites the code. AI can draft "this work meets current code requirements" as a general statement; specific section citations come from the licensed professional and the current jurisdiction's adopted code. The Level 4 program covers the AI-content QA pass that catches code-citation drift in AI-generated service pages and blog posts — same discipline at scale.

Failure Five: Customer Data and the Data Leak

The CSR pastes a customer's full call history — name, address, payment method on file, partial credit card, equipment serial, prior repair notes — into a public ChatGPT session to ask "what should we do next." The data is now in OpenAI's training pipeline (depending on plan tier and settings) and the customer's PII is outside the shop's controlled environment. Six months later the customer's email shows up in a data breach unrelated to your shop and the customer is asking why their payment data is exposed. You cannot prove it was not your CSR's ChatGPT paste.

This is the failure mode that is most often invisible until it isn't. ChatGPT, Claude, and Gemini consumer tiers have different data-handling policies than their enterprise tiers; CallRail, ServiceTitan, Avoca, and Rilla have signed BAAs and data-handling agreements; ad-hoc model use by unsupervised CSRs does not. Customer data — SSN for financing soft-pull, full credit card numbers, jobsite photos with people or kids in them — should never enter a consumer-tier AI tool. The Level 1 Chapter 5 lessons on data policy build the 5-line shop policy that prevents this; the failure mode is named here so the operator sees the exposure before it materializes.

Failure Six: The AI Voice Agent Promising What the Shop Can't Deliver

An Avoca / Jobber AI Receptionist / Housecall Pro AI Agent system prompt is not perfect. In the first 30-60 days of deployment, the AI will occasionally generate promises that contradict the shop's actual operations. "Free estimate" when the shop charges $79 for a service-call diagnosis. "Same-day service guaranteed" when the territory is fully booked. "Lifetime warranty on labor" when the shop's actual policy is one-year parts and labor. "We service that area" when the ZIP is outside the territory.

Every one of these is a customer-experience event waiting to happen. The customer arrives at the appointment expecting a free estimate and is told there's a dispatch fee; they cancel, leave a one-star review, and the shop has lost both the call and the Google reputation. The customer expecting same-day service is the customer who calls back at 4 p.m. asking why the tech isn't there.

The fix: every AI voice agent has a forbidden-phrase list maintained in the system prompt. Avoca and the other vendors expose this list during pilot setup; the operator's job is to populate it with shop-specific landmines. Daily review of the AI's call transcripts for the first 60 days surfaces drift. Weekly review after that. The 4 p.m. CSR review of all AI-booked calls catches the day's near-misses before they harden into tomorrow's complaints. Avoca and the other vendors are getting better at preventing this out of the box, but no vendor's defaults match every shop's actual policy on day one.

Failure Seven: FTC Endorsement Guidelines on AI-Drafted Review Responses

The marketing manager turns on NiceJob, Podium AI Employee, or Birdeye AI Employee for automated review responses. The AI responds to a negative review by acknowledging "we'll refund the full amount" — which the shop's actual policy does not commit to. The customer screenshots the response and posts it to a homeowners' Facebook group with the caption "they promised a refund, never gave it." The shop is now on the FTC's radar for misleading commercial practice and on the local consumer-affairs office's radar for unfair business practice. In 2026, the FTC's endorsement guidance specifically covers AI-generated review responses; the shop's "the AI wrote it" defense does not exist.

The fix: AI drafts review responses. The marketing manager or owner reviews every one before it posts (or applies a 24-hour holding window with notification to a human reviewer). The system prompt explicitly forbids commitment language ("we will refund," "we guarantee," "this won't happen again"); the AI is taught to acknowledge, route to a manager, and invite offline conversation — never to commit on behalf of the shop in a public reply.

Failure Eight: TCPA Exposure on AI-Initiated Texts and Calls

Hatch nurture sequences and AI-recovered missed-call outreach text customers. The TCPA governs automated texts and calls. Customers who explicitly opted out, customers whose phone numbers are wrong, customers whose numbers are now on the Do Not Call registry, customers on tribal lands with different consent requirements — each is per-violation statutory exposure of $500-$1,500. A Hatch sequence that sends 200 texts a week with even a 1% rate of opted-out or wrong-number recipients is $1,000-$3,000 per week in potential exposure.

The fix: the AI receptionist and the nurture-text tool both enforce DNC, opt-out, and wrong-number suppression at the source — not as a checkbox the marketing manager can override. The shop's monthly TCPA audit reviews flagged opt-outs and confirms suppression worked. AI does not initiate contact with customers who have not consented; the shop's CRM holds the consent log and the AI tool reads from it.

Failure Nine: Two-Party Consent on Recorded Calls

The shop deploys CallRail Conversation Intelligence and Avoca and Rilla. All three record audio. The shop operates in California, where two-party consent is required. The shop's pre-call disclosure says "this call may be recorded for quality assurance" — which is the standard one-party-consent script. In a two-party-consent state, that standard script is not legally sufficient. The shop is exposed on every recorded call: $1,000-$10,000 wiretap statutory exposure per call, plus civil liability, plus the political-PR damage when a single customer raises it.

The fix: the shop's pre-call disclosure language is reviewed by counsel against the state-by-state two-party-consent map and uses explicit consent language ("by remaining on the line, you consent to recording for quality and training purposes; if you do not consent, please advise the agent") in two-party-consent states (California, Pennsylvania, Florida, Illinois, Massachusetts, and the rest of the patchwork). AI tools that listen — CallRail, Avoca, Rilla — have to play the right disclosure for the right state, which means the vendor's deployment configuration covers this and the operator confirms it during pilot.

Failure Ten: The Multi-Error Cascade

The most expensive trades AI failure mode in 2026 is not any single error. It is the cascade. The CSR books a call through Avoca with an unverified service area. The dispatcher routes it. The tech arrives at the wrong house, returns to the shop. The AI-drafted apology email to the customer references a discount the owner never approved. The customer escalates. The Birdeye AI Employee response to the resulting one-star review commits to a remedy the shop won't honor. The customer posts the screenshot. The marketing manager scrambles. Hours of senior time burn. The Friday recap to the owner is dominated by the cascade rather than the lift.

This is the failure mode that compounds because each error feeds the next. The fix is the 30-second verify habit applied at every handoff. AI books → CSR verifies. CSR confirms → dispatcher verifies. AI emails → manager verifies. AI responds → marketing manager verifies. The verify habit is the brake on the cascade. Without it, one error becomes ten by Friday.

The Five Ways AI Gets Shops Fined in 2026

If the failures above are the daily exposures, here are the five named regulatory pathways most likely to produce a 2026 fine or lawsuit. Memorize these. Build verify discipline against each.

One: Reg Z violations. AI-drafted financing language with wrong APR, term, or fee disclosure. Federal exposure, $500-$5,000 per violation plus attorney fees plus CFPB complaint trail.

Two: FCRA adverse-action failures. AI-drafted soft-pull-decline notice that omits required disclosures. $100-$1,000 per violation plus actual damages.

Three: TCPA on automated calls and texts. Hatch sequences and AI receptionist outbound to non-consented or DNC numbers. $500-$1,500 per call/text.

Four: Two-party consent wiretap exposure. AI-listening tools recording without compliant disclosure in two-party states. $1,000-$10,000 per call plus civil liability.

Five: FTC endorsement and unfair-practice exposure. AI-drafted review responses or marketing content that misleads, fabricates remedies, or fails endorsement disclosure. Variable penalties plus state-AG complaint exposure.

None of these is hypothetical. Each has 2025-2026 case law or settlement precedent in adjacent industries (mortgage, auto, telecom, lending). Trades shops are the next vertical for plaintiff-firm attention as AI deployment expands. The verify discipline is the cheap insurance.

What This Lesson Fixes

The 12% embedded number in ServiceTitan's 2026 report includes shops that deployed AI, captured visible early wins, ran into one of the failures above, and quietly rolled it back. Each rollback in 2026 has the same shape: tool worked, workflow was thin, verify discipline was absent, the failure landed, the owner pulled the plug, the AI initiative is dead at that shop for the next 12-18 months. The shops that survive AI are not the shops with the best vendors; they are the shops with the most disciplined verify habit and the most explicit failure-mode awareness.

The list above is not exhaustive. New failure modes emerge as AI agents take more action surface — agentic dispatching, agentic financing initiation, agentic permit filing — each will have its own version of failures eight, nine, and ten. The discipline that protects against all of them is the same: name the failure mode before it lands; build the verify habit around the highest-exposure handoffs; review the failure log weekly; refuse to let the AI act unsupervised on regulated artifacts. The shop that operates this way captures the wins listed in the prior lesson and avoids the losses in this one. That is the shop the rest of this program builds.

Key Takeaways

• Hallucinated part numbers and equipment specs — AI never orders parts; verify against supplier catalog before order.
• Fabricated warranty and rebate terms — verify against manufacturer term sheet and current utility/state rebate page; never paste AI numbers into customer-facing warranty language.
• AI-drafted financing and Reg Z exposure — financing numbers come from Wisetack / GreenSky / Synchrony portal output, copy-pasted verbatim; AI never touches regulated APR, term, or fee disclosures. FCRA adverse-action notices use lender templates only.
• Electrical and plumbing code hallucinations — AI never cites code in customer-facing artifacts; licensed professional cites jurisdiction-specific code from current adopted version.
• Customer data leaks — PII (SSN, full credit card, jobsite photos with people) never enters consumer-tier AI; only enterprise tools with signed BAA and data agreement.
• AI voice agent over-promising — forbidden-phrase list in system prompt, daily transcript review for 60 days, then weekly; 4 p.m. CSR review of all AI-booked calls catches drift.
• FTC exposure on AI review responses — AI never commits remedy on shop's behalf in public reply; system prompt forbids commitment language; manager reviews before post or 24-hour holding window.
• TCPA on AI-initiated calls and texts — DNC, opt-out, and wrong-number suppression enforced at source; monthly TCPA audit.
• Two-party consent — state-specific disclosure language reviewed by counsel; AI tools configured per-state; pilot confirms it.
• The multi-error cascade is the worst failure mode — 30-second verify at every handoff is the brake.
• Five named 2026 fine pathways — Reg Z, FCRA, TCPA, two-party consent wiretap, FTC endorsement / unfair practice. Each has 2025-2026 case-law precedent in adjacent verticals.