ServiceTitan's AI variant builder matches by word, not function
ServiceTitan’s AI can attach variant materials to a service in seconds, which is genuinely useful for a big build. The catch is that it matches by keyword, not by function. It reads the words in a material name and grabs things that sound related, whether or not they belong on the job.
So it speeds the build and quietly fills it with parts that do not go together, and a human who knows the trade still has to check every one.
What actually happened
We were building a circulation pump for a boiler. The AI dutifully attached variant materials: a condensate pump, and a three-quarter coupling. It matched the word “pump.” But you do not put a condensate pump on a boiler, those are different parts for different jobs, and the coupling was not what that connection needed either.
Nothing in the system flagged it. The variants looked filled in, the build looked done, and it was wrong.
Any tech reading that can spot the problem in one sentence. The software cannot, because it is not reasoning about boilers and pumps, it is matching strings.
Why this happens
The AI is doing keyword similarity, not trade logic. “Condensate pump” and “circulation pump” share the word “pump,” so to a model looking at text they look related. To a tradesperson they are unrelated parts that live in different systems.
The model has no concept of what a boiler is, what a circulation pump does, or why a condensate pump has no business there. It is fast and confidently wrong, which is the most expensive kind of wrong, because it looks finished.
This is not a knock on ServiceTitan. Keyword matching is how any text-based AI works, and the platform handed you a genuinely fast accelerator. The one piece it cannot supply is trade judgment, and that was always going to be human.
The lesson: the concept has to live in a human’s mind
This is the same truth as the rest of pricebook work: it is not the material, it is the concept. You can let an AI dump 300,000 catalog items or auto-attach a thousand variants, and you still cannot do anything useful with it unless someone holds the concept of what actually goes together (your pricebook has 400 services, the good ones have 80 is the same idea from the other direction).
The AI is a speed tool for someone who already knows the answer, not a replacement for knowing it.
So use it, then check it. Let the AI propose variants to save typing, and have someone who has done the trade delete the condensate pumps it hangs on boilers.
That review step is not optional, it is the whole difference between a book techs can trust and a book full of confidently wrong entries. A book techs do not trust is a book they stop using.
What this means for your book
If your pricebook was built fast with AI-attached variants and never reviewed by someone who knows the trade, assume some of those variants are wrong in ways that will surface in a driveway. The fix is a pass by human eyes that know a boiler from a furnace, which is part of every build we do, and the kind of thing the free Pricebook Health Audit helps you spot.
The full picture of how a book should be structured, with or without AI help, is in our complete guide to a ServiceTitan pricebook that sells.