Home Blog Machine Learning From 6 Hours to 10 Minutes: How AI Transformed Estimating at the Largest U.S. Casework Factory One of largest U.S. manufacturer of commercial cabinetry operates a highly automated factory: CNC machines, automated material handling, and robotics. One stage of the workflow was slowing down the overall process – estimating. Each estimate took the team two to six hours, most of it spent manually counting cabinets on construction drawings. Given that the company handles hundreds of thousands of cabinets annually, the team wanted to speed up and automate this step. Here’s how the new system works. Their estimating team spent two to six hours on every quote. With hundreds of thousands of cabinets shipped each year, that adds up fast. Here’s how we got it down to under ten minutes. The slowest step in a fast factory Quote requests came in as large architectual drawings. The estimator’s job sounds simple: read the drawings, count the cabinets by type, enter the numbers into the quoting system, send the bid. In practice, every step was manual. Cabinet symbols varied between architects. Drawings had to be reviewed page by page, with relevant sections highlighted by hand. Counts were tallied manually. Then the data was entered into the quoting system. Why this needed custom AI? They had already tried off-the-shelf takeoff software, but none of it fit, because the work is really three jobs: reading drawings, searching spec documents, and feeding clean data into the ERP. So the project needed a custom build. Three AI types, working as one system 1. Computer Vision for reading the drawings We trained a custom vision model on more than 10,000 labeled cabinet examples. It detects each cabinet, classifies it by type, counts them, and marks up the drawing with annotations the estimator can review at a glance. Different architects draw cabinets in subtly different ways. A model trained on real, varied examples picks up those patterns instead of breaking on them. 2. LLM for searching the spec documents Each project comes with a 200-page spec document detailing the architect’s requirements for materials, hardware, finishes, and dozens of other details. When an estimator needed a specific answer from it, they used to spend 30+ minutes searching. We built a tool that lets them ask a question and get the answer in under 30 seconds, with the exact page reference. Once the AI counted the cabinets and pulled the answers from the specs, the results still had to land in the right place: the manufacturer’s internal ordering system, in their format, using their internal product codes. 3. Custom integration for fitting their workflow We connected the AI directly to their system, so the estimator’s spreadsheet fills itself. How accuracy improved over time The AI’s accuracy didn’t grow smoothly. It moved in jumps, and each jump came from fixing a specific thing. We started at 76%. Looking at the data, we realized the problem wasn’t the AI. Three different people had labeled our training examples slightly differently, so the AI was learning the inconsistency, not the cabinets. We re-labeled everything with one expert and clear definitions. Accuracy jumped to 82%. Next, we started tracking exactly what kinds of mistakes the AI was making. The most common one: confusing elevation views with plan views. We added more training examples specifically for that case. Accuracy went to 87%. By weeks 7–10, we had a proper system in place for spotting errors and improving the model systematically. Accuracy reached 90%. Could we have pushed for 97%? Yes. It would have taken another six to eight weeks for diminishing returns. At 94%, the estimator reviews the AI’s output in 10 minutes instead of doing the work from scratch in four hours. That math wins. Going further would have cost weeks for marginal gain. The results, in numbers Time per estimate: 2–6 hours → under 10 minutes Accuracy: 90%, measured against a ground-truth test set ROI: break-even in under one month Headcount: no layoffs. The estimating team didn’t get smaller — they got 10–20x more output per person Scalability: the system keeps improving as estimators flag edge cases. When a new architect started using non-standard symbols, the estimator flagged 20 examples, the system retrained overnight, and it now handles that architect’s drawings automatically What this means for your operation The headline numbers: 2–6 hours per quote down to under 10 minutes. Accuracy at 94%, measured against ground truth. Break-even in under a month. The estimators didn’t shrink in number — their role shifted from counting symbols to reviewing the system’s output and making judgment calls. If your team is doing significant document work by hand ( drawings, specs, quote prep) there’s almost certainly an AI-assisted version of that workflow that’s already feasible. The hard part isn’t the AI. It’s mapping your workflow precisely enough that the AI can be built around it. Related articles //December 26, 2024 //LLM Experiments with different LMMs //June 16, 2026 //News COXIT Co-Founder Volodymyr Hresko Named to Woodworking Network’s 2026 Wood Industry 40 Under 40 //December 26, 2024 //Prompting Prompt Engineering through Structured Instructions and Advanced Techniques Let's collaborate Tell us a bit about your project or challenge, and we'll get back to you shortly. Volodymyr Hresko Co-Founder & COO Reach out directly [email protected] CompanyThis field is for validation purposes and should be left unchanged.Full name First Email DetailsBy submitting the form, you agree to Coxit’s Privacy Policy.
Home Blog Machine Learning From 6 Hours to 10 Minutes: How AI Transformed Estimating at the Largest U.S. Casework Factory One of largest U.S. manufacturer of commercial cabinetry operates a highly automated factory: CNC machines, automated material handling, and robotics. One stage of the workflow was slowing down the overall process – estimating. Each estimate took the team two to six hours, most of it spent manually counting cabinets on construction drawings. Given that the company handles hundreds of thousands of cabinets annually, the team wanted to speed up and automate this step. Here’s how the new system works. Their estimating team spent two to six hours on every quote. With hundreds of thousands of cabinets shipped each year, that adds up fast. Here’s how we got it down to under ten minutes. The slowest step in a fast factory Quote requests came in as large architectual drawings. The estimator’s job sounds simple: read the drawings, count the cabinets by type, enter the numbers into the quoting system, send the bid. In practice, every step was manual. Cabinet symbols varied between architects. Drawings had to be reviewed page by page, with relevant sections highlighted by hand. Counts were tallied manually. Then the data was entered into the quoting system. Why this needed custom AI? They had already tried off-the-shelf takeoff software, but none of it fit, because the work is really three jobs: reading drawings, searching spec documents, and feeding clean data into the ERP. So the project needed a custom build. Three AI types, working as one system 1. Computer Vision for reading the drawings We trained a custom vision model on more than 10,000 labeled cabinet examples. It detects each cabinet, classifies it by type, counts them, and marks up the drawing with annotations the estimator can review at a glance. Different architects draw cabinets in subtly different ways. A model trained on real, varied examples picks up those patterns instead of breaking on them. 2. LLM for searching the spec documents Each project comes with a 200-page spec document detailing the architect’s requirements for materials, hardware, finishes, and dozens of other details. When an estimator needed a specific answer from it, they used to spend 30+ minutes searching. We built a tool that lets them ask a question and get the answer in under 30 seconds, with the exact page reference. Once the AI counted the cabinets and pulled the answers from the specs, the results still had to land in the right place: the manufacturer’s internal ordering system, in their format, using their internal product codes. 3. Custom integration for fitting their workflow We connected the AI directly to their system, so the estimator’s spreadsheet fills itself. How accuracy improved over time The AI’s accuracy didn’t grow smoothly. It moved in jumps, and each jump came from fixing a specific thing. We started at 76%. Looking at the data, we realized the problem wasn’t the AI. Three different people had labeled our training examples slightly differently, so the AI was learning the inconsistency, not the cabinets. We re-labeled everything with one expert and clear definitions. Accuracy jumped to 82%. Next, we started tracking exactly what kinds of mistakes the AI was making. The most common one: confusing elevation views with plan views. We added more training examples specifically for that case. Accuracy went to 87%. By weeks 7–10, we had a proper system in place for spotting errors and improving the model systematically. Accuracy reached 90%. Could we have pushed for 97%? Yes. It would have taken another six to eight weeks for diminishing returns. At 94%, the estimator reviews the AI’s output in 10 minutes instead of doing the work from scratch in four hours. That math wins. Going further would have cost weeks for marginal gain. The results, in numbers Time per estimate: 2–6 hours → under 10 minutes Accuracy: 90%, measured against a ground-truth test set ROI: break-even in under one month Headcount: no layoffs. The estimating team didn’t get smaller — they got 10–20x more output per person Scalability: the system keeps improving as estimators flag edge cases. When a new architect started using non-standard symbols, the estimator flagged 20 examples, the system retrained overnight, and it now handles that architect’s drawings automatically What this means for your operation The headline numbers: 2–6 hours per quote down to under 10 minutes. Accuracy at 94%, measured against ground truth. Break-even in under a month. The estimators didn’t shrink in number — their role shifted from counting symbols to reviewing the system’s output and making judgment calls. If your team is doing significant document work by hand ( drawings, specs, quote prep) there’s almost certainly an AI-assisted version of that workflow that’s already feasible. The hard part isn’t the AI. It’s mapping your workflow precisely enough that the AI can be built around it. Related articles //December 26, 2024 //LLM Experiments with different LMMs //June 16, 2026 //News COXIT Co-Founder Volodymyr Hresko Named to Woodworking Network’s 2026 Wood Industry 40 Under 40 //December 26, 2024 //Prompting Prompt Engineering through Structured Instructions and Advanced Techniques
