SAV · dealers · OEM hotlines
Deflect tickets without deflecting liability
HVAC support automation that escalates with proof
Dealer networks and OEM hotlines drown in repeat questions on the same twenty codes. Automate the lookup and first safe checks — but force abstention when the caller’s model is unknown.
Not a generic “HVAC GPT” — answers carry manufacturer references or a ticket escalation, never invented torque specs.
The business problem
Your tier-1 queue is costly when answers lack OEM sources
Support leaders measure AHT and deflection. Engineering measures API latency. Neither metric captures the cost of a chatbot that confidently ships the wrong EEPROM procedure — until the OEM escalates.
Repeatable volume
Many inbound calls are “what does code X mean?” on brands you officially support — structured lookup can deflect these before senior tech time.
Channel sprawl
Phone, Zendesk, WhatsApp, and marketplace messaging each got a different bot vendor — none share the same OEM corpus.
Brand risk
When the bot is wrong, the OEM logo is on the apology email — not the SaaS vendor’s.
Real scenario
Daikin U4 on a dealer chat — abstention saves a visit
A heat pump distributor routes web chat through their bot. User sends “U4 on Daikin Altherma, serial photo attached.” Bot calls resolve_model; serial maps to two possible sizes.
diagnose_v2 abstains with “ambiguous model — ask for type plate photo field X.” Ticket escalates to human with the abstention reason logged — no invented hydraulic procedure.
- Abstention is non-billable and JSON-stable for CRM routing
- search_technical_docs handles “what is U4?” without diagnostic commitment
- Senior queue receives citation snippets, not raw LLM prose
Redacted excerpt for illustration only. The live API returns full manufacturer, manual, page and section — see the HVAC-Bench protocol.
Illustrative citation
Gree · GMV6 service manual · p. 201
Bench scenario: hvac-bench/support-e6-gree
{
"status": "sourced",
"error_code": "E6",
"manufacturer": "gree",
"manual": "GMV6 service manual (redacted)",
"page": 201,
"recommended_action": "Verify communication cable shielding",
"billable": true
}{
"status": "no_exact_source",
"error_code": "E6",
"message": "SKU not in strict provenance scope — escalate to tier-2",
"billable": false
}Honest limits
Automation boundaries
Not a replacement for regulated advice
Gas, flue, and high-voltage safety remain human-in-the-loop — AskMarcel surfaces OEM checks, not legal compliance sign-off.
Language coverage
Corpus is strongest on FR/UK OEM packs; verify brand coverage for your export markets on /preuve.
Voice is your layer
We provide structured API/MCP responses — telephony ASR/TTS and queue logic stay in your contact centre stack.
Two perimeters
Support ops vs digital product
Perimeter A — Support & CX leadership
You own CSAT and escalation cost
You need deflection with a clear escalation path when the corpus is silent — and reporting that separates “answered” from “answered safely.”
- Route abstentions to tier-2 with structured reason codes
- Monthly quota fits tier-1 volume pilots
- HVAC-Bench answers procurement’s “why not ChatGPT?”
Perimeter B — Digital & integration team
You wire bots to backends
You want one HTTP/MCP integration behind Zendesk, Genesys, or a custom Node middleware — not per-brand scrapers.
- REST v1 for fast code meaning; v2 when model_id is known
- MCP when the bot runtime is already agent-native
- Webhooks and CRM mapping stay in your repo
Integration flow
From ticket created to escalated with citations
- 1
Normalise the inbound code
Map free text (“ erreur e7”, “E.07”) to brand + code via /v1/search or MCP search_technical_docs.
- 2
Resolve equipment if possible
CRM fields or user prompts feed resolve_model — skip diagnose if model unknown.
- 3
Answer or abstain
Successful lookup returns cited meaning; v2 diagnostic returns next check or abstention — both are valid bot outcomes.
- 4
Attach evidence to the ticket
Store JSON citations in the ticket for tier-2 — reduces “customer already talked to the bot” frustration.
Evidence
Lookup accuracy where automation pays off
On exact code lookup scenarios — the bulk of tier-1 volume — AskMarcel reaches 97% accuracy with valid manual references in HVAC-Bench.
Review HVAC-Bench rows97%
Exact code lookup
94%
Manual page retrieval
400k+
Corpus codes indexed
Automate tier-1 without automating mistakes
Test deflection on your top ten codes with a Developer key, then scale with a pilot tied to your ticket export.