There is a particular kind of software output that has learned to sound like knowledge. It looks like this: Cost overrun risk: high (87% confidence). Orange badge, small chart, maybe a trend arrow. It is on every construction dashboard sold in the last three years, and there is nothing you can do with it. You cannot check it. You cannot argue with it. You cannot take it to the sub. The only two responses available are believe it or ignore it, and after the third false alarm everyone picks ignore.
A score is a feeling with a decimal point
Eighty-seven percent confidence in what, computed from which rows, compared against what? The number is unfalsifiable by construction. If the phase comes in under, the score was 87 and not 100. If it comes in over, the score was right. It cannot be wrong, which is exactly why it cannot be trusted — a claim that cannot fail is not telling you anything about the world.
It is also the wrong output for the job. A superintendent standing in a driveway at 6:40 a.m. does not need a risk temperature. He needs to know which line, how much, and whether to make a phone call before the truck leaves the yard.
What a sourced claim looks like instead
Same situation, stated the way anomaly detection states it:
Framing materials will land at 96% of current once the two pending receipts post. Current $38,500. Actual to date $28,200. Committed and approved $2,400. Two receipts photographed at the counter Thursday, still unposted, $6,340 combined. 28,200 + 2,400 + 6,340 = $36,940 against $38,500. Roof deck is not sheathed.
Every figure in that paragraph is a row you can open. The $6,340 is two photographs you can look at. The $2,400 is a purchase request with an approver's name and a timestamp. The $28,200 was never typed by anybody; it rolled up from transactions. The claim can be wrong — maybe one of those receipts belongs to the neighbouring lot — and that is the point. A wrong sourced claim gets caught in eleven seconds. A wrong confidence score gets caught in November.
The figures have to be the same figures
This only works when the AI reads the identical ledger the human reads. Not a nightly extract, not a warehouse copy, not a model's summary of the ledger. The 96% is computed from the four numbers on the line — baseline, current, committed, actual — and if you open that line yourself you will find those exact values sitting there. When the arithmetic lives somewhere the user cannot reach, “show the math” becomes a second story about the first story, which is worse than no story.
The one-action rule
A claim in BLT arrives with exactly one action. Not a menu of six. One.
Raise the framing current. Assign the flag. Post the two receipts. Whatever the single most likely correct next move is, that is the button, and it is the only orange thing on the screen. There are two reasons for that discipline. The first is that a notification with six buttons is a dashboard, and a dashboard is a thing people scroll past. The second is harder: choosing one action forces the system to actually have a recommendation. Six options is what you offer when you don't know, dressed up as flexibility.
Next to the action sits a plain link that opens the arithmetic in full — the rows, the dates, the receipt photographs. It is not a footnote. On most claims it is the more important of the two, because the first thing an experienced builder does with a machine's opinion is check whether the machine understood the job.
Vision states what it saw, not how sure it is
The same rule governs photos. Vision AI does not return a safety score of 0.73. It returns a sentence in plain language: fall protection not visible on the main roof surface, three workers on the deck, photo taken 07:14, phase roof deck. Then it routes the flag to a person with a name.
You look at the photo. Either the anchors are out of frame or they are not there. If they are out of frame you clear it in two seconds and BLT records that a named human cleared it and when. If they are not there, somebody is on a roof right now. Either way the exchange took one glance, because the flag told you what it saw rather than how it felt. A score would have given you nothing to look at.
The same applies to receipt reading. Extracted line items always keep the photograph of the paper alongside them, so checking the machine is a matter of looking at the thing it read, not trusting that it read well.
It never decides
This is worth saying without hedging. BLT AI does not move a budget line. It does not approve a purchase request. It does not pay an invoice, close a phase, sign an inspection, or clear a safety flag. It drafts, it computes, it routes, it asks. A human commits.
That is not caution for its own sake. Three concrete reasons:
- It cannot see the whole board. The reason you are holding that $4,000 back from the drywall sub is a conversation from last Tuesday that exists in no system. The machine reading the invoice thinks you are 12 days late paying.
- Authority is a real thing. Somebody's name goes on a change order. The person whose name it is decides. A ledger where the last actor on a $9,200 move is a piece of software is a ledger you cannot defend at a draw meeting.
- Being checkable is the whole value. An assistant that acts on its own has to be right every time or it costs you money. One that states, sources, and waits only has to be useful, and it is allowed to be wrong sometimes without breaking anything.
The test
Next time a tool tells you something about your job, ask it three questions. Which rows is this computed from? What were the actual figures? What is the one thing I should do about it?
If the answers are a chart, a percentage, and a suggestion to review your budget, you have been handed a mood. If the answers are eleven transactions, an addition you can do on paper, and a button that raises one line, you have been handed a fact and left in charge of what happens next. That is the only arrangement worth putting on a jobsite — and it is the same on every plan, including the free one.