Confidently wrong is the expensive kind of wrong.
In this trade a wrong answer isn't a bad email. It's a misdiagnosed compressor, a warranty claim you can't defend, a gas appliance serviced from a guess, and a second truck roll to fix what the first visit got wrong. Every AI tool can answer. The question is what it does when it shouldn't.
HVAC professionals tested ChatGPT on trade questions and reached a verdict: the answers “look convincing... but often aren't actually correct.”1 A technician standing in front of a unit cannot check a confident paragraph against anything: no page number to open, no revision to compare, no way to tell a right answer from a fluent one.
And the liability is yours: a tribunal held Air Canada responsible for what its chatbot told a customer, rejecting the argument that the bot was somebody else's problem.2
1. HVAC Today, National Comfort Institute, reporting an HVAC-Talk thread on ChatGPT. hvactoday.com/chatgpt
2. Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, 2024. Negligent misrepresentation. American Bar Association summary
What refusing actually looks like
Asked: What pretreatment do we recommend for a site running above 5 NTU?
Not in the knowledge base. The service manual gives 5 NTU as the feedwater turbidity limit. It does not recommend a pretreatment method above that limit, and I will not infer a specification from a limit. I have filed this as a gap for product engineering.
Gap #7 filed · 7th ask · routed to product engineeringWhen the knowledge base doesn't settle a question, Answerlore doesn't produce its best guess in a confident tone. It says “not in the knowledge base,” names the team that owns the answer, and files the question as a gap.
We tested this on a real manufacturer's service manual: asked a question the manual doesn't settle, Answerlore declined to infer the answer from diagrams and logged the gap instead. It also caught the manual contradicting itself about its own version number. Honesty about documents turns out to include honesty about their defects.
Answerlore never answers from the model's general knowledge. Every factual claim traces to a file in your knowledge base, and every file traces to a source document and page. If the chain is missing, the answer is “not in the knowledge base.” Not because the model can't produce words, but because in support, provenance is the product.
Tools that never say “I don't know” can't tell you what your documentation is missing.
A refusal files a gap. Gaps carry a count of how often the question is asked, so your expert answers the most-asked one first, once, in writing. The knowledge base grows in the exact order your customers are asking. Answerlore's gap log is that list, ranked.
| Question | Hits | Owner | |
|---|---|---|---|
| can a warranty transfer a second time if the first was within 30 days | 4 | RMA desk | P1 |
| recommended feedwater pretreatment for sites above 5 NTU | 7 | product eng. | P1 |
| does E11 ever fire from a clogged pre-filter | 3 | product eng. | P2 |
“A product that refuses looks worse in a bake-off against one that always answers. We accept that.”
If you want to test us fairly: ask questions your manuals answer, ask ones they don't, and watch which tool tells you the difference.
What this costs us · Answerlore
Ask it something your manuals don't answer.
The refusal is the feature. Try to make it guess.