Manufacturing
AI nesting and quoting for sheet metal shops: what history it needs
By SourceX Editorial · Updated
Short answer
AI nesting and quoting for sheet metal shops needs linked history, not just old quotes: part geometry and quoted price, ERP actual times by operation, nest reports with sheet yields and remnants, and whether each quote won. If operators clock in by batch rather than by job, treat those times as averages and fix clock-in habits before training anything.
Key takeaways
- AI quoting learns from quotes linked to job actuals and outcomes; a folder of old PDF quotes teaches it list prices, not costs.
- Nest reports with sheet utilization and remnant use are the best source for material cost per part.
- Batch clock-ins and setups mixed with run time are the most common reasons AI time estimates drift.
- Lost quotes, reason codes and estimator overrides carry the judgment a model needs to price competitively.
- Customer prints stay customer property, while your times, yields and outcomes are your shop's operating record.
What history does AI quoting actually learn from?#
AI quoting for a sheet metal shop learns from three linked record sets: the quote, the job that followed and the outcome. A quote alone shows what you charged; the job record shows what the part really cost in laser time, brake setups, welding and finishing; the outcome shows whether you won, made margin or reworked the part.
The link between those sets matters more than the volume of any one of them. A shop with fewer quotes that carry a work order number through to the router, the nest report and the invoice gives a model more to learn from than a shop with a large archive of unconnected PDFs.
- Quote records: part number, revision, material, gauge, quantity, price, lead time and estimator.
- Job records: routing, actual time by operation, setups, scrap, rework and ship date.
- Nesting records: nest files and reports with sheet size, utilization and remnant use.
- Outcome records: won or lost, reason code, invoice amount and any NCR or customer return.
Inputs, estimates and where they fail#
Each input teaches an AI quoting or nesting tool something different, and each has a typical failure that shows up as estimates your estimators stop trusting. The table maps the common inputs to what they support and where they break, so you know which records to clean first.
| Input | What the AI estimates from it | Where it fails |
|---|---|---|
| Quote history with part files | Price by part family, quote turnaround | Prints quoted but never run; customer discounts read as cost |
| ERP actual times by operation | Laser, punch, brake, weld and finish time | Batch clock-ins, setups blended into run time |
| Nest reports and sheet yields | Material cost per part and expected utilization | Mixed sheet sizes or common-line cuts not labeled |
| Material price history | Cost at the quote date | Contract pricing mixed with spot buys and surcharges |
| Bend counts and brake setups | Brake time and tooling changes | Bends live only in drawings, not as data fields |
| Win and loss records | Price likely to win by customer and part type | Missing reasons; lost quotes purged from the system |
| NCRs and rework | True cost of difficult geometry | NCRs not tied back to the job or the quote |
Why nest reports and remnant records matter#
Nest reports are the most reliable record of material cost per part because they show how parts actually shared a sheet. Whichever nesting package you run, such as SigmaNEST, ProNest, Radan, Lantek or the laser builder's own software, check whether it saves a report for each nest with sheet size, parts, utilization and skeleton scrap, and how long those reports are kept; many shops overwrite nests or delete them once a job ships.
An AI nesting tool also needs the constraints your programmers apply by habit: grain direction for brushed stainless, part priority by due date, common-line cutting rules and which remnants are worth keeping. If those rules live only in a programmer's head, capture them as written notes before a pilot, because the tool cannot infer them from finished nests alone.
Keep the job or work order number on every nest report. Without it, a quoting model cannot connect a part's share of a sheet to the price you charged, and nest history becomes a separate archive instead of training signal.
How clean does ERP time data need to be?#
ERP time data needs to separate setup from run time and attach each to a single job, or AI time estimates will average away the differences between easy and hard parts. The most common problem in fab shops is operators clocking several jobs at once on the press brake, which spreads one block of time evenly across unlike parts.
Check the data before blaming the model. Pull actual times for a few repeat part numbers and compare runs. Wide swings on the same part usually point to clock-in practice, not shop performance, and machine monitoring logs from laser or punch controllers can serve as a cross-check on cutting time.
| Check | What good looks like | What to do if it fails |
|---|---|---|
| Setup versus run | Separate labor entries for each | Tag history as blended and fix the clock-in screen |
| One job per clock-in | Time attaches to a single work order | Treat batch entries as averages, not part times |
| Routing matches reality | Operations in the ERP match the floor | Update routings before using them as features |
| Equipment changes marked | A date for each new laser or brake | Split history before and after the change |
Lost quotes and estimator overrides carry the judgment#
Lost quotes are as useful as won ones because they show where your price sat outside the market. Many shops delete them or never record a reason, which leaves a model that only knows what customers accepted. A simple reason code such as price, lead time, capability or no response adds a lot of signal for little effort.
Estimator overrides matter for the same reason. When an estimator adds a markup for a tight-tolerance flange or a customer with heavy paperwork, that adjustment is expert judgment. Store the system price and the final price side by side, with a short note, so the reasoning survives the estimator's retirement.
Illustrative: an enclosure shop tests AI quoting#
Illustrative: a fictional sheet metal shop that builds electrical enclosures and brackets wants faster quotes for repeat customers. It runs a job-shop ERP, a separate nesting package on its fiber lasers and a shared drive of RFQ emails and customer prints. The estimating lead pulls several years of quotes and finds that laser times are clean, but press brake times were clocked by batch for most of that period.
The shop pilots AI quoting on laser-heavy flat parts first, tags brake history as averages, and adds reason codes to every new lost quote. It also changes the brake clock-in screen to one job at a time. Estimators review every AI draft, and their overrides are saved as notes. Quote turnaround improves on flat parts while bent assemblies stay with the estimators until cleaner brake data builds up.
Who owns this history, and how SourceX approaches it#
Customer prints, models and specifications generally remain the customer's property and confidential information, while your quotes, actual times, nest yields and outcomes are records of your own operations. Before uploading anything to an AI quoting vendor, check your customer NDAs and the vendor's terms on whether it may train on your files.
The same linked history can matter beyond your own quoting. AI developers building estimating and planning tools look for real quote-to-job-to-outcome records, and the SourceX Enterprise Data Value Framework rates drivers such as domain expertise, human-generated signal, recency and rights. SourceX works through the SourceX five-step transaction of Supply, Rights, Preparation, Approval and Delivery, with customer drawings and identifiers excluded during Preparation.
Frequently asked questions
How much quote history does AI quoting need?
There is no fixed minimum. Coverage matters more than age: enough won and lost quotes across your common part families, materials and gauges, with actual times for jobs that ran. History from before a major equipment change, such as a new laser, should be tagged rather than mixed in, because it describes a different shop.
Can AI nesting work if our nesting software is separate from the ERP?
Yes, as long as nest reports carry the job or work order number. That key lets you connect sheet utilization to the job's routing, actual times and invoice. Without it, nest history and ERP history stay in separate silos and neither tool learns material cost per part.
Will AI quoting replace our estimators?
In most shops it drafts quotes and estimators review them. The estimator still catches unusual tolerances, finishing requirements and customer quirks. Saving each override with a short reason improves later drafts and preserves estimating knowledge when experienced people retire.
Do we have to upload customer drawings to an AI quoting tool?
Most tools need part geometry to estimate cutting and bending, so check customer NDAs first. Then read the vendor's terms: whether files are used only for your account, whether they train shared models, and how files are deleted when you leave.
Is our quote and job history worth anything outside the shop?
It can be. Linked records that show how experienced estimators priced parts, and what the jobs really cost, are the kind of human-generated operational data some AI developers look to license. Customer drawings are excluded, and any value depends on depth, linkage and rights, and is known only once a buyer engages.
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.