AI Door Takeoff: What the Software Actually Reads on a Division 8 Set

Your count came back clean. Now decide whether to put your name on a number a machine pulled from documents you already know contradict each other.

An AI door takeoff doesn’t read your documents one at a time. That’s the part demos skip, and it’s what separates a tool that returns a count from one that returns a bid you can defend. Reading a door schedule alone produces a list. Reading the schedule, the elevations, the partition schedule, the floor plans and the 087100 hardware specification against each other produces that list plus the places those documents contradict one another, which is where a Division 8 bid comes apart.

Arvind Veluvali, Fresco’s CEO, states the design constraint in one line: “We don’t look at any of your documents in isolation.”

Fresco runs a takeoff in four stages, or five when a project has repeating units, meaning a hotel or a multifamily building where one unit plan repeats across floors. What follows is organized by document, because the document is what an estimator recognizes.

What an AI Division 8 takeoff reads before it counts anything

Before the counting starts, the system returns a recap of the job: opening count, hardware set count, project type, location, architect. Then the features that decide whether the job is worth bidding at all. A key system. Access control. STC. Lead-lining. Adjacent glazing. Pair doors. Approved manufacturers for doors, frames and hardware. Door construction by type, hollow metal against wood.

That page exists so an estimator can rule a job out before opening a plan sheet. On a bid board where the constraint is hours rather than opportunities, a fast no is worth nearly as much as a fast yes.

The door schedule, and the columns an architect often leaves blank

The takeoff starts where an estimator starts. Fresco pulls every door schedule in the set, normalizes them into one table, and keeps the source pages alongside it. One large hospital package carried eleven separate schedules spread across ten pages.

Then it fills columns the architect didn’t provide: handing, leaf count, interior or exterior, STC, lead-lined, wall type, existing frame.

That list is the tell, because the architect didn’t put those columns in the schedule on this set, and most sets are the same. Handing is read off the floor plan. Leaf count comes off the door elevation. Wall type comes from the partition tag at that opening on the plan, checked against the partition schedule. A tool that only opens the door schedule cannot fill any of them, which is the practical difference between software that reads a schedule and software that reads a set.

This pass also flags what’s wrong inside the schedule itself, starting with door codes that repeat when a code should belong to one opening. That is one of several ways a count drifts from what’s actually built, and we covered the others in where Division 8 counts drift across match lines.

Door and frame elevations, where pairs get caught

Next Fresco pulls the types, meaning the door and frame elevations, and normalizes those into a table as well.

Pair detection is why this pass exists. A pair is one opening with two leaves, and a takeoff needs both quantities, which is why the counts export reports openings and leaf count as separate columns rather than one number. The schedule may or may not tell you which openings are pairs. The elevation shows it, because the two leaves are drawn, so pair status gets read there and the leaf count column on the schedule fills from it.

The partition schedule, for what the wall drives

Wall types come next, normalized the same way: code, description, total thickness, fire rating, STC rating and the stud information.

That last field is the one estimators ask for in different words. As Arvind puts it, people say jamb depth or throat size, and what they mean is stud info. So the table carries it explicitly instead of leaving it to be worked out at ordering time.

One caveat worth knowing before you rely on that column. Total wall thickness is pulled when the architect provided it. When a wall carries multiple layers of board and the set never states the total, that arithmetic is still yours.

Where an AI door takeoff marks off the floor plans

This is the pass that most resembles the manual work it replaces. Fresco marks off the floor plans opening by opening and compares what it finds against the schedule it built earlier.

Three failures show up, and they don’t cancel out:

• Doors drawn on the floor plans that never appear in the door schedule
• Doors in the schedule that don’t exist anywhere on the plans
• Doors that appear six times in the plans and twice in the schedule

A total can reconcile against the schedule exactly and still be wrong in two directions at once. Why schedules and plans disagree so reliably is its own subject, covered in the fake door schedule problem.

The more useful behavior is what happens with a door code the plans never show. Instead of reporting it missing and stopping, the system drops the estimator into the area where that door should be, on the logic that a CX door usually sits near another CX door, and lets them scan the neighborhood for unmarked openings. A search problem comes back as a location, which is a better handoff than a flag.

The hardware spec, where the set numbers stop matching

The last pass pulls hardware sets, off the 087100 specification or off the drawings depending on where that project put them, and reconciles them against the door schedule. Two hardware set conflicts dominate, and neither one moves the count.

A door assigned to two different sets. One opening carries hardware group 107.3 in the door schedule and 107.4 in the specification. An opening cannot have two hardware sets, so somebody decides which document governs. Both values get surfaced and the estimator picks.

A set that doesn’t exist. A door is assigned to set 107.8 in the schedule, and no set 107.8 appears in the hardware spec. The count is correct. The hardware behind it is missing entirely.

Neither error moves the opening total, which is exactly why both survive a review that was checking totals. Getting the sets out of the specification in the first place is a separate problem, covered in extracting a hardware schedule from an 087100 spec.

The system also applies the rules governing what hardware belongs on what door, a 42 inch wide door calling for a particular closer, a condition calling for a kick plate, and shows its reasoning instead of asserting the result. Arvind’s description of the source for those rules: “We take these directly off the rag books.”

Why the multi-document read is the whole architecture

The reason an AI door takeoff can do any of this is that the model holds the documents open at the same time instead of in sequence. Akhil Gupta, Fresco’s CTO, frames the advantage in terms of what a person physically cannot do: “The best human estimator is still limited by the powers of human perception.”

While the model has one drawing in front of it, it also has the specifications, the shop drawings, the submittals and the contracts. That is what lets a wall type on a partition schedule reach a door line on a schedule three sheets away, and it’s the capability a general takeoff tool trained on broad construction imagery has no path to, because it was never taught what a hardware set is.

What comes out the other end

The takeoff exports as three Excel files rather than one, because they answer three different questions.

Schedules. Every door schedule on its own sheet plus a consolidated view, with door prep, frame prep, wall type and handing calculated in.

Counts. Openings and leaf count broken out by every characteristic on the job, so how many hollow metal 3070s are on this project is a lookup rather than an afternoon.

Hardware. The sets themselves, plus a bill of materials that pulls a hinge out of the twenty or so sets it appears in and sums it, plus totals by manufacturer. That file is the one that matters at buyout, and a takeoff that stops at a door count cannot produce it. Reviewed data exports toward Comsense hardware entry as import-ready files instead of a schedule to re-key by hand.

Where this stops helping

Four limits, all of them Arvind’s own.

It isn’t fully automatic. The flags in the action center are work an estimator still has to review and resolve, and on a large package that runs 30 to 60 minutes. The extraction and formatting go away. The friction doesn’t.

Some openings are excluded by default. Closet doors and bathroom doors typically aren’t marked unless you ask, because on most commercial packages they aren’t the scope being bid.

Some questions have no second document to check against. Unit entry doors on a repeating-unit job are the clearest case. With 18 door types across a set of units, and nothing stating which door appears in which unit, there is no source of truth to compare against, so the flag gets raised at the unit level for a person to settle.

Unit mix tables are often wrong, which is why unit counts get read off the floor plans instead of taken from the table. Architectural and structural partition drawings that disagree get flagged the same way, though whether that comparison runs depends on the project.

What actually changes about the review

On that hospital package of more than 750 doors, the review list came down to the three or four openings with real problems. Arvind’s estimate is that the job runs about a week for an estimator working alone, and under an hour through Fresco.

The judgment stays exactly where it was: which document governs when the schedule and the spec disagree, and what goes back to the architect as a written qualification rather than an assumption buried in a number. Those calls belong to the person whose name is on the bid.

What changes is where the day goes. Making those calls takes a fraction of the time that finding them used to.

Key takeaways

• The errors that sink a Division 8 bid live between documents rather than inside any one of them.
• Handing, leaf count and wall type are answers a door schedule was never going to contain, because they belong to the floor plan, the elevation and the partition tag.
• A door assigned to two hardware sets, or to a set that doesn’t exist in the spec, leaves the opening count correct and the hardware wrong.
• A frame throat matches total wall thickness, not stud width, and a wall’s hourly rating is not the door’s label rating.
• A total that reconciles against the door schedule can still be wrong in two directions at once.

Frequently asked questions

See what it reads on your own set. Fresco’s AI takeoff platform is built for doors, frames and hardware specifically, and reads the whole package, not one document at a time.