Every construction AI pitch you have read this year has been a list of things it can do. Here is the other list, which is shorter to write and more useful to have, because a tool whose boundaries you know is a tool you can actually put on a job. The claim underneath all of this: naming the limits precisely is not a disclaimer. It is the thing that makes the rest of the claims worth anything.
It cannot inspect what was never photographed
Vision AI reads frames. Fall protection, house wrap below grade, exposed conductors, standing water, debris on a deck — all of it comes from a photograph somebody took. Zero photographs produce zero flags, and zero flags is not the same thing as zero problems.
This is the most dangerous limit on the list, because the failure is silent and looks like success. A phase with 140 photos and four flags looks worse on a report than a phase with two photos and none. It is almost always the safer phase. Any system that lets an empty photo log render as a clean bill of health is lying by omission, and it is lying in the direction that gets somebody hurt.
So BLT counts the absence. Twenty-six hours clocked to roof deck this week, zero frames uploaded, phase active — that is a stated fact with an action attached, not a green check. It still does not know what the roof looks like. It knows that nobody does.
It cannot price work nobody scoped
The Baseline memory holds what phases have actually cost against recorded scope. Recorded is doing the work in that sentence. The client who says “while you're in there, put two more cans in each room” on a Saturday walkthrough has just added twenty-eight fixtures across fourteen rooms, plus the wire, plus the electrician's time, plus a lighting plan revision. If nobody writes it down, the ledger sees an electrical phase drifting over for no visible reason, and the smartest possible model reaches the same conclusion you would from the same data: costs are up, cause unknown.
The machine can flag the drift. It can tell you electrical is 22% past current with rough-in incomplete. It cannot tell you why, because the why happened out loud in a room, and nothing on this list of subsystems was in that room.
It cannot know a handshake
The lumber yard takes 8% off your orders because you have bought from them for eleven years and the owner's kid played ball with your kid. It has never been written anywhere. It is renewed by a phone call each spring.
BLT reads the invoice, sees $18,400 against $20,000 expected for the package, and reports that you came in under comparable. Correct on the arithmetic, blind on the mechanism. It does not know that the discount is a relationship, that the owner retires in March, and that your framing costs are going up 8% the moment his son-in-law takes over pricing. Everything soft in this business — who owes whom a favor, which sub will show up in a rainstorm, which inspector wants the panel schedule laminated — is invisible to a system that reads transactions. Those are frequently the facts that decide whether a job makes money.
It cannot replace a superintendent's judgement
A schedule can be drafted from the budget, the phases, and the crews. Whether to pour Thursday with 40% rain forecast and a finisher who has been off his game since his divorce is not a scheduling problem. It is a read on weather, concrete, and a person, made by somebody who has been wrong about all three before and remembers how it went.
The same goes for the hundred small orderings that make a week work — who to put on which lot, when to let a sub slide a day to keep them for the next four jobs, when the framing crew is tired enough that the mistakes are about to start. None of it is in the stream. Some of it should never be written down at all.
It cannot fix inputs that were wrong going in
Hours coded to the wrong phase produce a clean, confident, incorrect cost per square foot. A receipt photographed at the counter but assigned to Maple Ridge instead of Ridgeline moves $550.17 to the wrong job, and every projection on both jobs is now slightly false. OCR catches vendor, date and line items reliably; it cannot know which of your three active projects the person standing at the counter was buying for if they tap the wrong one.
This is why capture takes ten seconds and defaults to the phase you are clocked into — not to be pleasant, but because the error rate of the whole system is set at the counter. Nothing downstream repairs a bad input. It just carries it further with more decimal places.
It cannot predict a market it has not seen
Baseline is a memory, not a forecast. It knows what OSB cost you across nineteen closed phases in this market. It does not know about the mill fire, the tariff announced Tuesday, or the hurricane rerouting three weeks of freight. When BLT says a phase is running 38% above comparable, it is comparing against the past and stating the window it used. Ask it what lumber does in September and the honest answer is that it has no idea, and the useful version of the tool says that instead of generating a plausible number.
And it does not decide
Not a technical limit — a designed one, and the most important entry here. BLT AI does not move a budget line, approve a purchase request, pay an invoice, close a phase, or clear a safety flag. It states a fact, shows the arithmetic, and offers one action. A person with a name commits it, and that name is what appears in the change log at the draw meeting. For a GC carrying real liability, a ledger whose last actor on a $9,200 move is a piece of software is not a ledger they can defend.
Why the list is the point
Two reasons, one of them commercial and one of them practical.
The commercial one: a tool that claims everything gets tested on nothing. Builders are professionally suspicious of vendors, correctly, and the first over-claim spends the credibility of every accurate statement that follows. When a system tells you it cannot see something, the things it does tell you start carrying weight, because you now know the shape of what it is drawing from. Precision about limits is the cheapest trust available, and almost nobody in this category spends it.
The practical one is better. Look at the list again and notice how many of these are input problems rather than intelligence problems.
- It cannot inspect what was not photographed — so photograph the phase while it is open. Ten seconds, on the phone already in their hand.
- It cannot price unscoped work — so write the Saturday conversation into a change order before Monday.
- It cannot know the handshake — so put the 8% in a note on the vendor, where the person who inherits your estimating can find it.
- It cannot fix bad coding — so tap the right phase at the counter, which is where the whole accuracy of the system is actually decided.
Four of the six limits close from the field side, not the software side. The two that remain — judgement and the future — are the two that were always going to be yours, and the reasonable ambition for a machine is to hand you a complete, dated, sourced picture and then get out of the way while you use them. That division of labour holds across every trade and role that runs a build.
Anyone promising more than that has not stood on a jobsite in the rain, watching a decision get made on information nobody wrote down.