How Casework Manufacturers Are Cutting Quote Time by 95%

Commercial casework is a bidding business long before they cut any plywood. Someone opens a set of architectural drawings, often hundreds of pages, and counts every cabinet, countertop, elevation and callout. Then they read the specification document, which runs past 200 pages, the key goal is to spot the clauses that change the price.

Shops have done this by hand for decades and some still refuse to digitize these projects. It’s why quotes take days, and why manufacturers turn down bid invitations they’d have a fair shot at winning.

Of course with the AI revolution it is starting to change, and the numbers are larger than most people outside the industry expect.

The estimating job is very manual and time consuming

Manufacturers invest heavily in the factory buying CNC routers, edgebanders, and automated finishing lines. A modern casework plant can hold millions of dollars of equipment running with tight tolerances. The estimating department, meanwhile, often runs on PDFs and highlighters.

An estimator works through a drawing set sheet by sheet, marking every unit and tallying quantities by type. A set of moderate complexity takes two to six hours; a large one takes considerably longer. The work is also unforgiving. Miscount the base cabinets on sheet 240 and the error flows straight into the quote, where it becomes a lost bid or a lost margin.

How fast a small team can read documents ends up capping how many projects the whole business can pursue.

What is changed right now in estimation

Two capabilities matured at roughly the same time, and together they cover the estimating problem.

The first is object detection on drawings. A computer vision model trained specifically on construction documents can locate and classify cabinet units on a sheet. Generic image recognition falls short here; the model needs thousands of labeled examples, because drawing conventions vary between architectural firms. One architect’s base cabinet symbol looks little like another’s.

The second is language models reading specifications. Spec documents are long, repetitive and consequential. A model can search them in seconds and return the relevant clause with a page reference, instead of an estimator scrolling for half an hour to find the hardware requirement that drives the cost.

The estimator stays in the loop either way. What the software takes over is the mechanical part of the job: counting and searching.

The results in production

At the largest US manufacturer of commercial casework and architectural millwork, this approach has been running in daily production since early 2025.

The system reads architectural drawings, detects and classifies cabinet units, and exports structured takeoff data formatted for direct import into the quoting system. The results in production

The measured outcome:

• Drawing review dropped from two to six hours to roughly ten minutes.
• One estimator can process about 20 drawing sets per day, up from two to four.
• Specification lookup fell from over 30 minutes to under 30 seconds, returned with page references.
• The largest set processed to date ran to 470 sheets.
• The system reached break-even within one month.

Nobody’s job disappeared. Output per person increased by a factor of ten to twenty, which in a bidding business means answering more invitations rather than employing fewer people.

Why 90% accuracy was the stopping point for AI

Most technical write-ups skip this part, and it matters for anyone evaluating a system like this.

Detection accuracy on the custom AI takeoff system started at 76%. Systematic error analysis brought it to about 90%. Pushing higher was achievable and would have taken several more weeks of work. That work was skipped, deliberately. Why 90% accuracy was the stopping point for AI

At 90% accuracy, an estimator reviews the output in about ten minutes and corrects what needs correcting. At 97%, they’d still review it, because nobody submits a bid on unreviewed machine output. The review step survives at any accuracy level; it just gets marginally shorter.

Chasing the last few percentage points buys very little and costs a great deal. The useful target is the accuracy at which human review becomes fast.

Anyone assessing a vendor should ask two questions: what accuracy the system achieves, and how long the review takes. The second number governs throughput.

Where they AI systems break

The failure modes are predictable, so they’re worth listing plainly.

Models tend to merge adjacent cabinet units into a single detection, or miss units altogether. The root causes are image resolution limits and symbol density. A drawing packed with closely spaced units, scanned at low resolution, is where accuracy degrades fastest.

Drawing conventions also drift. When a new architect’s set arrives with unfamiliar symbols, performance drops until the model is retrained on those examples. A production system needs a retraining loop, ideally overnight, or accuracy erodes as the mix of clients changes.

One more lesson from benchmarking this work: the pipeline around the model, meaning preprocessing, resolution handling and how sheets are segmented before they reach the model, affects results more than the choice between leading models. Teams that fixate on model selection and neglect the workflow tend to be disappointed.

What this means for shops

The estimating bottleneck comes down to document reading, and document reading is now largely automatable.

For a casework manufacturer, the practical consequence is bid volume. If quoting a project takes a day, there’s a hard ceiling on how many opportunities the business can pursue. If it takes twenty minutes, that ceiling moves. In a market decided by who responds fastest with an accurate number, that shift settles the competitive question.

The technology has left the experimental stage. It runs in production, on real drawing sets, in a plant shipping product every day.