We delivered an AI estimating engine for a custom cabinetry shop that turns client photos, designer sketches, and site measurements into detailed, accurate quotes in under an hour — replacing a two-week estimating backlog that was costing them jobs to faster competitors

A custom cabinetry shop in North Charleston with six craftsmen and a reputation for beautiful work had a problem that had nothing to do with woodworking: they could not quote fast enough. The owner was the only person who could estimate a job, because estimating custom cabinetry is genuinely hard — it requires translating a designer's sketch or a homeowner's Pinterest screenshot into board feet of specific hardwoods, sheet goods, hardware counts, finishing labor, shop hours, and installation days, all while accounting for the difference between a straightforward paint-grade laundry room and a curved walnut kitchen island with inset doors and custom corbels. He carried the entire estimating logic in his head, built from twenty-two years of jobs. The result was a two-week quoting backlog. Designers and contractors would send him a project, wait ten days for a number, and in the meantime a faster competitor — usually one quoting from a price-per-linear-foot rule of thumb — would win the job with a number that was sometimes wrong but always first. He estimated he was losing three to four good jobs a month not on price or quality but on response time. He also suspected, correctly, that his own quotes had drifted: material prices had moved significantly over two years while his mental math had not, and a post-mortem we ran on his last thirty completed jobs showed he had underquoted hardwood-heavy projects by an average of nine percent while overquoting paint-grade work enough to lose bids he should have won. We built him an AI estimating engine trained on his shop's actual history. The foundation was his own data: we ingested three years of completed jobs — the original quotes, the final invoices, the material purchase records, and the shop's time-tracking logs — and the AI learned the real relationships between job characteristics and true costs. Not industry averages. His shop, his craftsmen, his suppliers, his actual hours. It learned that inset doors take his shop 1.4 times the labor of overlay doors, that curved work carries a premium his old quotes never fully captured, that jobs on the peninsula carry an installation-day penalty for parking and access that jobs in Summerville do not. The intake layer accepts whatever the client actually has. A designer's dimensioned drawing gets parsed directly — the AI reads the elevations, counts the boxes, identifies door styles, and extracts dimensions. A homeowner's photo of an existing kitchen with a tape measure in frame gets interpreted with the AI asking targeted follow-up questions about the two or three things it cannot see: interior fittings, drawer counts, appliance panel requirements. A Pinterest screenshot gets classified by construction style and complexity tier, with the AI generating a clarifying-questions checklist the owner can send back to the designer in minutes instead of scheduling a call for next week. From the parsed scope, the engine builds the estimate the way the owner would, but in minutes: a full material takeoff with current pricing pulled from his suppliers' latest price sheets (updated monthly, so lumber price moves flow into quotes automatically instead of eroding his margin silently), hardware line items down to the hinge count, finishing labor calibrated to the specified finish system, shop hours broken out by fabrication stage, and installation days adjusted for site conditions. Every estimate comes with a confidence rating and flags the assumptions it made, so the owner reviews a structured draft instead of building from a blank page. His role shifted from producing estimates to approving them — a fifteen-minute review instead of a three-hour session he could only fit in after the shop closed. The quote document that goes out is a professional, itemized proposal with the shop's branding, scope descriptions written in client-friendly language, allowances clearly stated, and options presented as priced alternates — the walnut version versus the white oak version, soft-close upgrade priced as a line item — which turned out to matter commercially, because designers started forwarding his quotes directly to clients instead of retyping them, and the priced alternates consistently pulled projects up-market. The engine also learns continuously. When a job closes, the actual hours and materials feed back in, and the model adjusts. Two months in, it flagged that his new CNC operator had cut fabrication time on slab-door jobs by twelve percent — and suggested his quotes on that work were now leaving margin on the table that could either be kept or used to bid more aggressively. Ninety days after launch, his average quote turnaround had dropped from ten business days to same-day for standard work and forty-eight hours for complex projects. His close rate on quoted work rose from roughly one in four to better than one in three, and the jobs he won were more profitable because the pricing drift was gone — hardwood-heavy work now carried its true cost. He told us the strangest part was psychological: for twenty-two years, every quote request had felt like homework hanging over his evenings. Now the backlog is a review queue he clears with his morning coffee, and his craftsmen have a fuller schedule than the shop has ever run.