Ask a builder what their company owns and you get trucks, tools, a yard, maybe a crew that has stayed eleven years. Nobody says the twenty-two closed jobs sitting in a filing cabinet and a QuickBooks file. Which is a shame, because that is the asset that decides whether next year's bids make money, and for most companies it is stored in a form no person and no machine can actually use.
What a closed phase actually contains
Not a total. A closed phase is a set of facts that were true about a specific piece of work in a specific market on specific dates:
- The baseline — what the phase was priced at when the job kicked off, locked and unable to drift.
- The current plan when it closed, and every move between the two with a name, an hour, and the reason the person typed.
- The actual, rolled up from real transactions rather than typed by anyone — the $550.17 receipt from Home Depot #4821 with its five line items, the sub invoice, the rental.
- The hours, coded to that phase, with a location on the row.
- The photographs, timestamped, showing the work at the state it was in on each of those days.
- The flags raised against it, when they were raised, and how long they sat before anyone owned them.
- The scope it was all against, in writing, including the two change orders and excluding the third thing the client asked for on a Saturday.
Hold that shape thirty times and you no longer estimate framing. You look up what framing has cost you.
Why almost nobody has it
Not because builders are careless. Because the record gets shredded across four systems the moment the job ends, and no one is paid to reassemble it.
The actuals land in accounting under GL codes that were designed for a tax return, not a phase. Labor goes to payroll as a weekly total per person, so the hours that went into roof deck are permanently mixed with the hours that went into siding. The photos are on a superintendent's phone, and he got a new phone in August. The scope lives in a PDF emailed in February. The change orders live in a text thread. The reason the framing line moved $3,000 lives in one person's head, and that person now works for a competitor.
Six months later somebody asks what framing costs on a 2,400 square foot two-story in this market and the honest answer available is: about what it cost last time, plus something for lumber. That is not a number. That is a feeling with a dollar sign in front of it.
Machine-readable is the whole trick
“We keep good records” usually means a folder named by street address containing forty PDFs. That is a record a person can read on a slow afternoon in January. It is not a record anything can compare.
Machine-readable means the phase is a row with typed fields: scope, square footage, market, start and close dates, baseline, current, committed, actual, hours, vendor mix. It means the actual was never keyed in by hand, because a hand-keyed actual is a guess wearing a suit — it comes from a person deciding which bucket a $1,240 plumbing invoice belongs in, three weeks after the fact. In the BLT ledger the actual rolls up from transactions that each carry their own receipt image, vendor, and date. Nobody types the number, so nobody can round it.
It also means the change history survives the close. A phase that finished 9% over baseline is one story if the moves were “client added a dormer” and a completely different story if they were “lumber up since baseline” twice. Same overage. Opposite lesson. Only one of them should change how you bid the next one.
What it does to the next estimate
Here is the arithmetic that becomes possible on the day you close your third comparable phase. Framing, three closed jobs in the same market inside fourteen months: $34,900 on 2,310 square feet, $38,200 on 2,480, $41,600 on 2,650. Per square foot: $15.11, $15.40, $15.70. That is a $0.59 spread across three builds and a visible drift upward. Now price a 2,520 square foot lot. Fifteen-forty a foot is $38,808, and you know the direction the last two moved.
Compare that to the alternative, which is taking last year's number and adding 8% because that feels about right. Eight percent is not a market observation. It is a hedge against not knowing, and you pay for it either by losing the bid to somebody who did know or by winning it at a margin you find out about in November.
The same memory is what lets BLT AI say something specific mid-build instead of something vague. “Foundation is 38% above comparable” only means anything if comparable is a set of real closed phases with real scope attached. That is what the Baseline subsystem holds, and it is built out of finished jobs, one at a time, starting with yours.
The other four times it pays
Estimating is the obvious one. It is not the only one.
- Draw requests. A lender asking what the $84,000 draw covers gets phase-level actuals with photographs dated inside the period, not a cover letter.
- Change order disputes. The client remembers agreeing to one thing. The record has the scope, the date, the photo of the framing before the wall moved, and the $2,150 invoice that followed. Arguments end faster when both sides are reading the same dates.
- Warranty and callbacks. Two years on, a leak at a valley. The closed phase has the roof deck photos from the day before it was covered, which is either the cheapest thing you own or the most expensive thing you didn't keep.
- Knowing which work to stop bidding. The most useful thing thirty closed phases will tell you is that one category of job has quietly lost money three times running. Nobody notices that from memory, because the jobs that hurt are the ones you remember least accurately.
Say plainly what it does not give you
Three closed phases is a starting point, not a statistic. The spread on a small sample is wide, and a builder who treats $15.40 a foot as a law will be wrong the first time a lot has bad soil. BLT AI states what it has — how many comparable phases, over what window, in what market — and lets you decide whether that is enough to price against. It does not manufacture confidence out of three data points.
The record is also only as honest as the capture. A phase where half the hours went to the wrong code produces a clean-looking cost per foot that is quietly wrong. Scope creep that never got written down turns a well-run phase into a mysterious overrun. The machine cannot recover what nobody entered, and it will say so rather than fill the hole with an average.
Start with one
You do not need to reconstruct twenty-two jobs. You need the next phase you close to close properly: baseline locked at kickoff, every change carrying a reason, receipts photographed at the counter and read into line items against the phase rather than typed later, hours coded on the day, photos taken while the work is open. Then lock the period and let it become the thing the next one is measured against.
One phase, done that way, is worth more to your estimating than a year of folders named by street address. And if your estimating lives in the back office, this is the change that makes their number defensible instead of diplomatic.