Where it started
3D Coupe runs an active CNC floor. Every project goes through a nesting report: parts, material, cut lengths, sheets used, offcuts. The information is complete and reliable, but unreadable to a system and interpretable only by the person who programs the machines.
The direct consequence: every price request went back through the programmer, even the simple ones. Sales could not produce a quote on their own, and time estimates rested on experience rather than a written method.
The challenge
Make the sales side self-sufficient on estimating, with machining times computed rather than eyeballed, without touching the machines or adding a minute of data entry on the floor.
What we built
The mandate opened with a full diagnostic: the real flow was mapped from the client drawing coming in to the program going out to the machine, and twenty-two automation avenues were assessed one by one, each with its difficulty, cost and integration time. Management picked the starting point on that basis, not on a hunch.
What was then built is a chain of five specialized agents. A nesting report lands in a watched folder and the chain fires on its own: PDF extraction, part normalization, cutting-tool matching against the shop's tool chart, time and material calculation, then a consolidated quoting report emailed to the person responsible.
The heart of the system is not the AI, it is the formula. Number of passes by thickness, real cut length, feed rate per tool, setup time that tapers with the number of parts per sheet, surface consumed net of offcuts. The calculation is written down, reproducible and verifiable line by line. The agents apply the method, they do not invent it.
Nothing was wired directly into the machines. The chain works off the reports the nesting software already produces, so there was no risk to production during rollout.
How it unfolded
Mapping
The real flow documented with the shop, from drawing intake to G-code.
Chart and formula
Tool chart cleaned up, time calculation method written down in black and white.
Agent chain
Extraction, normalization, tool matching, costing, final report.
Tests and calibration
A dozen real cases across clients, materials and thicknesses, with instructions tightened at every gap.
The results
Estimating a 2D part went from 58.4 minutes to 14.5 minutes on average across the tests. Sales drop in a file and get a costed report back, without going through the programmer for routine work.
The least visible gain is the most durable: the calculation method is now written down. Two people costing the same project land on the same number, and a new hire starts from the right basis instead of relearning by trial.
Put against the shop's real volume, the math is simple: six quotes a week, about five parts each, so roughly thirty parts to cost every week. At 43.9 minutes saved per part, that is close to 22 hours a week and more than 1,100 hours a year taken off a job nobody enjoyed.
Every project processed also builds a history of real times, which will sharpen estimates and open the later phases.
Estimating a 2D part
“What convinced us is that he did not show up with a ready-made solution. He took the time to understand our plant before touching anything. The result was good enough that we are going again with Avenor on a much broader mandate: production planning, offcuts, ERP integration.”
— Martin Picard, co-owner and operations director, 3D Coupe (translated from French)
The plan, delivered and ahead
Phase 1 was not picked at random. Twenty-two automation avenues were assessed, each with its difficulty, cost and integration time, then ranked by what paid back fastest with the least risk. Estimating came out first.
The rest was not thrown away. The avenues set aside for phase 1 form the plan for the later phases, each with the data that has to accumulate first. That is exactly what the current chain is already doing in the background.
Price a quote from what the shop already produces
- A file drop that triggers everythingThe nesting report lands in a watched folder. The chain starts on its own, processes files one by one and emails back a consolidated report.
- A tool chart cleaned up and usableTools, feed rates and depths of cut gathered into a single reference that the whole calculation hangs off.
- A standardized calculation formulaPasses by thickness, real cut length, setup tapering per sheet, surface consumed net of offcuts. Written down, therefore checkable.
- A quoting report ready to attachCut time, setup time, average time per part, sheet count and material utilization rate, in a single PDF.
- A dozen tests on real projectsVaried clients, materials, thicknesses and quantities. Every gap observed was used to fix the instructions rather than excuse the result.
Make the reference data trustworthy before wiring it to the ERP
- The full reference extracted and analyzedBefore wiring anything to their ERP, the data had to hold up. All 2,902 records in the reference were extracted and analyzed.
- 496 records audited, 410 defects documentedMaterial and tool records were gone through one by one, and every defect was written down rather than quietly fixed.
- Two naming standards deliveredPlanning built on a reference where the same material exists in three copies produces a wrong plan, confidently.
Close the loop, from the client drawing in to material tracking
- Client drawings validated at intakeA quarter of incoming files have to be reworked. A public template plus a conformity agent at the drop point settles it before it reaches the shop.
- Smart offcut trackingEvery remnant is identified, measured and stored with its reference, then proposed automatically when a new project can use it.
- The ERP link, in developmentThe interface to the management system API is in development. Orders, stock and production history will feed the estimate, and the estimate will feed its results back. That link is what unlocks the two projects behind it: production planning driven by real data, and smart offcut management.
- Assisted 3D programmingAutomatic generation of a large share of 3D toolpaths, with the programmer validating and finishing rather than starting from scratch.
- Maintenance before the breakdownOnce machine data is collected, wear drift can be spotted before it stops production.
Five agents in a chain, one quoting report out
Each agent has a single responsibility and hands a structured result to the next. That separation makes the chain testable: when a number is wrong, you know exactly which link to fix, and you fix it without touching the others.
The client file's journey is mapped all the way to the quote going out
Phase 1 stops at the costed report. The rest of the journey is already modelled: what comes in, what gets checked, what is billable when a drawing has to be reworked, and the exact point where a person approves before a price reaches the client.
Phase 1 was built to stand on its own, but also to set up what follows: every project processed leaves behind real times and utilization rates that will make the later phases easier to cost and easier to defend.
What the company gets out of it
Sales price work without waiting on programming
For routine work, the salesperson produces the quote alone. The programmer stays on the projects that actually need the expertise.
A written calculation method, not a habit
Passes, feed rate, setup, surface net of offcuts: the formula is documented and can be checked by hand.
Expertise no longer rests on one person
The tool chart and the best operators' settings live in the system. A new hire starts from the right parameter.
A production history that builds itself
Every project processed leaves its real times behind, the raw material of the later phases.
The thinking behind this mandate is laid out here: Why starting small with AI works better.