AI estimating for HVAC replacements: what it needs from your history

What does AI estimating need from your history?#

AI estimating needs a history that shows what you proposed, what the customer chose, what the install actually took and whether it held up. Equipment catalogs and price books tell a tool what things cost; your history tells it how your company prices real houses with real ductwork, access problems and customer preferences.

Most HVAC companies hold parts of this history in ServiceTitan, FieldEdge, Housecall Pro or a sales app used by comfort advisors. The pieces are usually there. What is often missing is the link between them, so that one replacement can be followed from first visit to the first callback.

The readiness checklist for HVAC replacement history#

The readiness checklist covers eight record families. A company does not need all of them to be perfect, but past installs, options presented, win and loss outcomes, labor hours and callbacks decide whether an estimating tool can be tested at all.

Score each row honestly as complete, partial or missing for the past few seasons. Partial rows are normal; the point is to know which job types have enough clean history to test against.

Equipment changes add one more wrinkle. Refrigerant transitions and efficiency rule changes mean older installs used models that are no longer sold, so a tool comparing past jobs with today's options needs a way to map retired model numbers to current equivalents. Keeping the full model number on each job, rather than a bundle name, makes that mapping possible.

Why lost proposals teach as much as won ones#

Lost proposals teach an estimating tool where your prices or options did not work, which is half of pricing judgment. A history made only of signed jobs makes every price look acceptable.

For each proposal, keep every option the comfort advisor presented, the one chosen if any, and a short loss reason: price, timing, financing declined, chose a repair instead or went with another contractor. Even an imperfect reason field is more useful than none.

Watch for one quiet distortion. If advisors delete abandoned proposals to keep their pipelines tidy, the history overstates your close rate and hides which option tiers customers walk away from.

Labor hours and callbacks: the honesty check#

Labor hours and callbacks are the honesty check on an estimating history because they show what each job really took and whether it was done right. A proposal priced as a straightforward swap that turned into a long install with duct modifications is not a good example to learn from, even though it was signed.

Tie crew time to the job rather than to the day, and record the reason when an install runs long: duct changes, an attic access problem or an electrical upgrade discovered on site. Link callbacks and warranty visits from the first heating and cooling seasons back to the original install, so you can see which job types and crews generate return trips.

Callback reasons matter as much as callback counts. A return trip for a thermostat setting the homeowner misunderstood says something different from a refrigerant leak at a brazed joint, and only the second one belongs in an estimator's view of install risk. A short coded reason on each callback keeps the two apart.

Illustrative: an HVAC company backtests an AI estimator#

Illustrative: a fictional residential HVAC company with several comfort advisors and install crews is offered an AI proposal tool by a software vendor. The owner wants to know whether the tool's suggestions would have matched the company's own decisions.

The operations manager exports last season's closed replacement proposals with options, outcomes, equipment, crew hours and callbacks. The tool is run on each job's starting information only, and its suggested options are compared with what the advisor presented and what actually sold.

The tool tracks well on like-for-like system swaps and drifts on homes that needed duct or electrical work, where the company's history recorded add-ons as one lump price. The owner limits the pilot to like-for-like replacements and asks advisors to itemize add-ons from now on.

How to backtest before you rely on AI pricing#

A backtest runs the estimating tool on closed jobs and compares its output with what actually happened. It is the cheapest way to learn where a tool can be trusted and where it cannot.

Keep the backtest results. They show the vendor where its tool struggles, and they give you a record to point to if pricing questions come up with customers or managers later.

• 1Pick a recent season of closed replacement proposals, both sold and lost.
• 2Give the tool only what was known at the first visit: existing system, home details and customer requests.
• 3Compare its options and prices with what your advisor presented and what the customer chose.
• 4Compare its labor assumptions with actual crew hours.
• 5Sort mismatches by job type to see where the tool drifts.
• 6Limit live use to job types where it tracked your results, and recheck each season.

What your estimating history is worth beyond the estimator#

Your estimating history is worth more than its use in one software tool, because linked proposals, outcomes, labor and callbacks show how skilled people make replacement decisions. That is the kind of human-generated signal the SourceX Enterprise Data Value Framework rates highly, alongside domain expertise and recency.

Your price book, margin rules and supplier costs are a different matter. Many owners treat them as confidential and exclude or mask them in any outside use, including AI vendor integrations. When SourceX reviews estimate records, those fields are handled in the Rights and Preparation steps, and the owner approves what is included.