What is AI takeoff software?
AI takeoff software reads a bid set and produces the quantities an estimator would otherwise pull by hand. The useful version does more than recognize shapes on a sheet. It holds several documents against each other at once and reports where they disagree, because that disagreement decides whether the number is safe to bid.
That distinction is the whole category. A tool that counts symbols on a floor plan gives you a count. A tool that reads the floor plan alongside the schedule, the specification and the elevations gives you a count plus a list of the places the documents contradict themselves. On a Division 8 package the second list is where the money is.
Why the same model performs differently in every trade
An AI model is only as good as the material it learned from, and construction drawings are among the hardest material in the world to get hold of at scale. They’re proprietary, they’re not published, and they look like nothing else. Akhil Gupta, Fresco’s CTO, puts the constraint plainly: “Every single construction drawing is proprietary and it looks very very different than anything else in the world.”
That’s why performance varies so much by trade. A model trained broadly across construction learns what is common to every discipline, which is largely geometry. Finding and counting a symbol is well within that. Recognizing that a door assigned to hardware set 107.3 on the schedule is assigned to 107.4 in the specification is a different task, and it depends on having been trained on what a hardware set is. Akhil’s read from customer experience is blunt: every customer who arrived from a generalist platform had found it fell flat.
Doors are the sharpest version of the problem. A single opening carries a door or a pair, a frame, and a hardware set that may run to a dozen line items. The opening itself carries the hardware group number, the fire label where it’s rated, and the handing. Most of that detail lives in a different document than the one showing the door. Fresco trained models specifically on Division 8 symbology and Division 8 documents for that reason. As Akhil describes it: “We had to train our own models to understand the door hardware division very specifically, and also understand the symbology in plans that is specific to door hardware.”
What the software reads on a Division 8 set
Fresco works through the package in four stages, or five when a project has repeating units like a hotel or a multifamily building.
First it returns a recap: opening count, hardware set count, project type, location, architect, and the features that decide whether a job is worth bidding at all. A key system. Access control. STC ratings. Lead-lining. Adjacent glazing. Pair doors. Approved manufacturers for doors, frames and hardware, and whether substitutions are allowed.
The reading is simultaneous rather than sequential, and that’s the part a person can’t match. While the model has one drawing in front of it, it also has the specifications, the shop drawings, the submittals and the contracts. Akhil’s framing of why that matters: “The best human estimator is still limited by the powers of human perception.”
One large hospital package carried eleven separate schedules spread across ten pages. Fresco surfaces every one of them, puts its own analysis beside them, and keeps the source pages on screen alongside it.
Ninety percent of a bid package is fine. The other ten decides the bid
Fresco’s CEO Arvind Veluvali frames the job around the part that isn’t: “90% of an architectural bid package will be accurate generally, right? It’s the 10% that’s not accurate or not consistent that’ll sink your bid.”
So the software assembles the straightforward ninety percent and hands the estimator the rest as a short list of decisions. In Arvind’s words: “we assemble like the 90% of the puzzle that is relatively simple to assemble ... and we show the estimator straight up like, ‘Hey, here are the seven or eight non-conforming pieces.’”
That reframes the common objection. Estimators ask what happens when the schedule and the plans are wrong, on the assumption that contradictory documents are where the software breaks. Contradictory documents are the condition it was built for.
What it does to the clock
Arvind describes a 4,000-door takeoff that took an estimator two weeks by hand and ran in under four hours through Fresco, and a twenty-door job that runs thirty minutes manually and about three with it. Across roughly sixty customers and thousands of takeoffs, the figure customers report back is 80 to 90 percent faster.
Reviewed takeoff data moves downstream toward Comsense hardware entry and the eMullion integration as import-ready output, with a spot-check before handoff rather than a full re-key.
Where the estimator still decides
Fresco shipped a fully automatic version first and pulled back from it deliberately. The reason was structural: a takeoff can only be as accurate as the documents behind it, and most bid sets aren’t clean. Better than eight in ten packages carry architect issues an estimator has to rule on.
So the flags go to a person. Which document governs when the schedule and the specification disagree, what to qualify in writing back to the architect. Those calls stay with the estimator whose name goes on the bid. Fresco’s job is to find them fast, not to decide them for you.
If you want the longer version of how these documents behave, the Division 8 Estimator’s Guide walks through three published 087100 specifications and how differently three owners published the same section. For the trade-level view of the same workflow, see the Division 8 takeoff software page. Both are linked in the related pages.
What to take from this
- The errors that decide a Division 8 bid live between documents, not inside any one of them.
- A model’s usefulness on a trade follows the material it was trained on, which is why performance varies so much by division.
- Contradictory documents are the condition this software was built for, not the case where it fails.
- The estimator still decides which document governs. What changes is how long it takes to find the decisions.
Frequently asked questions
What does AI takeoff software actually do?
It reads a bid set and returns quantities, then flags the places the documents contradict each other. On a Division 8 package that means reading the door schedule, floor plans, hardware sets, elevations and the 087100 specification together rather than one at a time, and reporting the openings where those sources disagree.
What can AI takeoff catch that a general takeoff tool can't?
Errors that live in the relationships between documents, and errors inside a document that only matter if you know what you're reading. A general tool trained on broad construction imagery can find and count symbols. What it typically won't surface is that one door is assigned to two different hardware sets, or that the set the schedule references doesn't exist in the specification, because those errors are only visible to a system trained on what a hardware set is. Fresco flags those and routes them to the estimator instead of picking an answer on its own.
Why does Fresco stop and ask instead of finishing the takeoff?
Because the fully automatic version didn't hold up on real bid sets. The first release produced a finished takeoff with no human step, and real bid sets contain contradictions that no tool should resolve on its own. So the flags go to the estimator, who makes the calls on what the documents actually mean, and the software removes the hours of rote reconciliation that used to surround those calls.
What does the time difference look like on a large package?
A 4,000-door package that took an estimator two weeks by hand ran in under four hours through Fresco. Smaller jobs compress similarly: a twenty-door takeoff that runs thirty minutes manually runs in about three.
Can AI takeoff handle incomplete or contradictory drawings?
That's the design target rather than the edge case. Takeoffs are often priced off drawings the architect hasn't finished, and better than eight in ten sets carry issues an estimator has to resolve. Fresco surfaces those as a working list instead of leaving them to be discovered during buyout.