An owner in this industry described what most contractors are doing as the toy version of AI. That's the right phrase. Asking ChatGPT to write a follow-up email is the toy version. The applicable version starts when the AI can see your numbers.
Because generic AI advice skips the part that matters: it knows nothing about your company. Ask it "how do I improve my close rate" and you get a listicle. Ask it "why did my close rate drop 9 points last quarter" and it can actually help — but only if it can see your estimates, your lead sources, and your rep numbers.
Here's the whole approach, in three steps.
Step 01 — Export three files, not one
The common mistake is uploading one export from your estimating software. That tells you what you quoted. It can't tell you where the lead came from or what the job actually cost to produce — so it can't answer any question worth asking.
"By the time we get a bid accepted there's so much manual work behind the scenes. We're now in a four-time data entry process. We put it in the phone system, we put it in CompanyCam, we put it in Monday, and we put it in PaintScout — and they're all separate."
— Owner, $4M residential shop
That's the actual reason one export can't answer anything. The story of a single job is spread across four systems that were never introduced to each other.
Twelve months each. Then create a Project — a workspace that holds the files and remembers context between conversations — and drop all three in.
Step 02 — Make it link them before it answers anything
These three systems don't talk to each other and nothing shares an ID. Linking them is the actual work, so put it in the prompt:
I'm attaching three exports from three systems:
1. LEADS — every inquiry from the last 12 months
2. QUOTES — every estimate we sent
3. PRODUCTION — every job we completed, with estimated
and actual hours and material
Nothing shares an ID. Your first job is to link them.
- Match leads to quotes, and quotes to completed jobs.
- Match on name and address. Expect misspellings,
abbreviations, unit numbers, and "Bob" vs "Robert".
- Tell me how many matched confidently, how many are
fuzzy, and how many didn't match at all.
- Show me the failures. Don't guess to fill gaps.
Then, using only the linked records:
1. Close rate by lead source, by rep, and by job size.
2. Which job types actually make money — estimated vs
actual, by type.
3. Cost per lead by source, through to gross profit
per source.
4. The biggest thing here I'm probably not looking at.Step 03 — Read the match rate first
Most people skip to the report. Don't. How much of it linked up is the most valuable number in the exercise.
If 40% of your quotes can't be traced back to a lead, you don't have a reporting problem — you have a capture problem, and no dashboard will ever fix it. You just found out exactly what to fix at intake, in an afternoon, for twenty dollars.
"I don't use the dashboard because every time, I'm like — no, that's not right."
— Owner, $8M painting company
He didn't have a dashboard problem. He had systems that never agreed with each other, and nobody had ever measured how badly. The match rate measures it.
That's also why "don't guess to fill gaps" belongs in the prompt. Left alone, AI will hand you a clean, confident report built on data that was half holes. You want the holes visible.
When should I connect live data instead of exporting?
When re-exporting starts to annoy you. Not before.
At that point you connect the AI directly to the systems you already pay for, so the linking happens on its own. Most software will issue you an API key if you ask — it's just a long password that lets a tool read your data. Keep it read-only until you've watched it work for a few weeks.
And if you don't know how to connect something, ask the AI to walk you through it step by step. It's a good teacher about itself. When the explanation doesn't land, say "explain that visually" and you'll get a diagram instead of another wall of text.
What's the best first automation to build?
Estimate follow-up. Every morning it checks the last 90 days of quotes, finds the ones going cold — sent two weeks ago, opened three times, no reply — drafts a message for each in your voice, and texts you the list. You read it over coffee and reply "send 1, 3 and 4, skip 2."
"It's lost revenue. We already paid for the leads, and we're just not doing anything with them."
— Owner, on his aging estimates
Follow-up is the first thing that falls off when a rep gets busy, and it falls off silently. You're usually not losing money on jobs. You're losing it between them.
This is buildable today. It is not buildable in an afternoon by someone who's never done it.
What won't AI fix?
- Selling. In residential especially, that's a human function. Automate what the salesperson hands off — scope, hours, material budget, is there a dog, where do we park — not the close.
- A process that doesn't work yet. Prove it manually, train somebody else on it, then automate it.
- Messy data. AI will surface the mess. Fixing it at the point of capture is still on you.
- Anything financial, unsupervised. A client of mine set the rule better than I would have: automate as much as possible, but human-check all of it, because you don't mess with the money.
If you take a process that doesn't work and make it fast, you just get a faster mess.
Should I do this myself or hire someone?
- Reading and reporting on your own data — do it yourself. You'll learn more about your business in a weekend than a consultant will tell you in a month.
- Anything that writes data, moves money, or touches the customer — get help. Not because it's hard to build, but because it's hard to build so your team can't break it. The value isn't the automation. It's the guardrails around it.
"I want a system that's very hard to screw up. You put a great person into the role with minimal training and it's just — it's got training wheels on it."
— Jason Connors, Owner, Spray-Tex
Training wheels is the whole job. Anybody can wire two tools together. Building it so the new hire can't break it on their second week is the part you pay for.
Same as a sprayer. Hand one to somebody off the street and you get a mess. Hand it to a painter who's run one for twenty years and they finish in half the time. Same tool.