It told me fifteen. I buy it for seven.

It told me fifteen. I buy it for seven.

The model will never hand you a blank space. That's the whole problem.

Paul Cobb
August 24, 2026

Ask a model what a roof costs and it will tell you fifteen thousand dollars. I buy that roof for seven.

It isn't lying. It's averaging a world I don't buy in: homeowner pays a general contractor found through a search ad. Real number, wrong customer.

I said that out loud last week to a room of operators. Then somebody in that room showed me the same failure wearing a face I hadn't recognized as the same failure.

He'd uploaded photos of a house and told the model he was leaning toward a wholetale. Great idea, it said. He asked what to do with the kitchen. Don't put money in the kitchen, it said. He asked what it thought about putting new doors on that same kitchen. It liked that too.

His words: "It never tells me no."

That's the same bug as my roof. The model will not hand you a blank space. Ask it for a price with no price history and it fills the hole with retail. Ask it for a judgment with no resistance and it fills the hole with agreement. Either way you get something confident back, because confident is the only setting.

Two moves came out of that room, one for each hole.

1. Bring your own numbers, or it brings the internet's.

I keep an export of my own material and labor costs, and I keep it reasonably current. It goes into the project before I ask a single question. Then the instruction is blunt: use these rates, don't deviate, and if a number isn't in here, flag the hole instead of filling it.

The flagged hole is the entire point. A gap I can see costs me a phone call. A plausible number I can't see costs me the job.

The vacuum also hides well at volume. Forty line items on a scope you've never priced yourself all look equally plausible, and that isn't rigor, it's camouflage. Cap the output at the handful of categories you cut checks for. Fake precision needs room to hide, so don't give it any.

Same move works well outside rehab. Pull the comp set from your agent, hand it over as a file, and tell it to reason over that set specifically. You get a much better read than turning it loose on public listing sites, because you've replaced its priors with your evidence.

The general rule, and it applies to underwriting, scoping, and pricing alike: the model is only as good as the data you refuse to let it guess at.

2. Make it argue, because it will not volunteer.

Nobody in that room had to be told the cheerleader problem was real. Everybody had lived it. What was useful was watching one member solve it without meaning to.

He'd asked for sourcing help on interior sheetrock, quantities and measurements included. First pass came back with a list of options, all of them rated under four stars. Instead of taking the list, he pushed back: reevaluate, and give me only four stars and up. Bought off the second list. That job came in eighteen hundred dollars under the quote he'd been working from.

He didn't run a clever prompt. He just declined the first answer. That's the skill.

Three that work, in rising order of teeth:

Are those the only options, or the best options? Show me what you left out and why.

Attack this plan. Argue the case against it as hard as you'd argue for it.

What would have to be true for this to be the wrong call? Give me the three most likely ways I'm wrong.

Better than remembering: set it once at the settings level, so pushback is the default posture. Something like this in the custom instructions box does it.

Default to disagreement before agreement. When I propose something, give me the strongest case against it before you help me do it. Never call an idea good without naming what would have to be true for it to work. If you don't know something, say so rather than estimating.

Where this bites.

Pushback prompts make a model less agreeable. That is not the same as making it right.

The model doesn't sign the check, carry the holding cost, or eat the overage. You do, every time, and that asymmetry is the entire reason you stay in the loop. Both moves above buy you a better draft. Neither one transfers the decision, and any tool that seems to be offering to take it off your hands is the one to watch.

The limits showed up all over that room. One member's renders had been carrying his jobs, right up until a render's shading didn't match the real material and his installer stopped work because he thought he'd been handed the wrong color. Another member ran two walkthrough videos through the same workflow: the quiet one transcribed clean, the noisy one came back with about one word in five correct.

None of that is an argument against any of this. It's an argument for knowing which step you're on. Capture and narration, fine. Arithmetic, not without your numbers underneath it. I keep a deterministic engine under my rehab estimates for exactly that reason, and the model never touches the math.

The play. Before your next scope or estimate, load your own cost history first and add one sentence: if a number isn't in what I gave you, flag it rather than fill it. Then, whatever it hands back, decline the first answer once. Those two moves cost about ninety seconds and they close most of the gap between fifteen and seven.

The longer version of why this one number resists automation, and why I've stopped treating that as a problem, is up on the Journal: Why AI Still Can't Price My Rehabs.

Thanks to the operators who brought real examples to this month's session, including the failures. Those were the useful part.

Know somebody whose AI agrees with everything they say? Forward it. And hit reply with the last thing your model told you that you knew was wrong. I read every one.

Written by Paul Cobb, real estate operator (Louisville / Kentuckiana), forVEX builder. I run the plays I write about. More depth: forvex.app/journal