What AI is actually good at on a construction job — and what it isn't

There are two loud camps on this right now and both are wrong. One says AI is about to run your business. The other says it's a toy that makes things up. If you're a contractor trying to decide whether any of this is worth your time, neither is useful.

Here's the version from people who build it and use it daily.

Four things it does better than a person

1. Reading documents nobody has time to read. This is the big one and it isn't close. A 400-page spec book has maybe fifteen pages that change what you bid. Finding them is a full day of skilled work that no one has a full day for, so in practice it gets skimmed. Software reads all 400 pages the same way every time, at two in the morning, without getting tired at page 180. It doesn't understand the project better than you do — but it will never skip a section because it was the end of a long week.

2. Comparing two documents that should agree. Plans against specs. Bid against contract. Submittal against what was specified. Discrepancy-hunting is mechanical, exhausting, and the single highest-value thing you can do before signing — and it's the first thing that gets dropped when a bid is due Thursday. Machines are relentless at it in a way humans can't sustain.

3. Turning messy input into structured records. A foreman talking into his phone for ninety seconds on the drive home is a better daily report than the form he won't fill out at 6pm. Taking that and producing something with dates, quantities, equipment, and delay notes attached — that's a genuinely solved problem now, and it's where a lot of the practical value hides.

4. Not getting bored. Underrated. Most operational failures aren't from people who can't do the work. They're from people who have done the same check 400 times and stopped really looking on 401. Consistency on repetitive checks is the machine's actual superpower, and it's the least exciting thing about it.

Three things it does badly

1. Judgment about people and risk. Which crew can handle a tight pour. Whether this GC pays. Whether the guy on the phone is telling you the truth about site conditions. That's pattern recognition built from years of consequences, and it isn't in any document. Anyone selling you AI that makes staffing or trust calls is selling you something they haven't thought through.

2. Knowing what it doesn't know. More on this below — it's the important one.

3. Anything needing current physical ground truth. What the site actually looks like this morning. Whether the subgrade is soft. Whether the material on the truck is what the ticket says. If a human didn't observe it and record it, the system doesn't know it, and no amount of cleverness fixes that.

The failure mode that costs money

If you take one thing from this: these systems are far more likely to give you a confident wrong answer than to tell you they don't know.

Ask for a quantity that isn't in the document and a poorly-built tool will produce a number anyway. It'll be formatted correctly, it'll be plausible, and it will be invented. The same way it'll tell you a spec section says something reasonable when that section doesn't exist.

This is not a rare glitch. It's the default behavior of the underlying technology, and it is the whole reason careful construction software is harder to build than a demo suggests.

There's a subtler version that has bitten us in our own tools, and it's worth describing because it looks like nothing. A search runs across a limited window — say the last fourteen days — finds nothing, and reports "no documents found." Technically true. Completely misleading, because the spec book was uploaded in March. Nobody lied; the tool just answered a narrower question than the one that was asked, and the honest-sounding answer sent someone off to re-upload a file that was already there.

So when you evaluate any of this, ignore the demo. Ask three questions:

• Does it show its source? Every number should point at the page it came from, and you should be able to click it and land on that page. If it can't cite it, treat it as a guess.
• What does it do when the answer isn't there? Ask it something the document genuinely doesn't cover. A good tool says it couldn't find it. A bad one fills the silence. This is a five-minute test and it tells you more than the whole sales call.
• Is absence loud? If a check finds nothing, do you get told "nothing found" or do you get told nothing at all? Silence that means "all clear" and silence that means "the system broke" must not look the same. This is the one people discover too late.

What this means practically

The honest framing is that AI is very good at the reading, checking, and recording layer of a construction business — the part that's high-volume, low-judgment, and currently eating your evenings. It's poor at the part you actually get paid for, which is deciding what to do.

That's not a limitation to apologize for. It's a good trade. The reading and checking is precisely the work that doesn't get done properly because there's never enough time, and it's where the expensive mistakes are made. Handing that to a machine that will genuinely never skip page 180 is a real advantage — as long as you've verified it tells you when it comes up empty.

We build our software around those three questions because we got the answers wrong ourselves first. That's not a sales line; it's just how you end up with tools that hold in a real business rather than in a demo.