Case study · Werk24 × Saphirion
A German precision-components manufacturer now prices RFQs 2.4× more accurately than its expert estimators — fully automated, in about a minute per part
The challenge
Every incoming RFQ needed a cost estimate before sales could reply. Senior engineers read each drawing by hand, squeezed between their real engineering work. Quotes took hours on a good day and days on a normal one.
- A bottleneck nobody could remove:Cost estimation depended on a handful of senior engineers, so quoting speed was capped by their availability.
- Inconsistent numbers:Two estimators priced the same drawing differently — and that eroded trust in every quote.
- Systematic padding:“Safety” margins inflated roughly 80% of quotes, and padded quotes lose deals.
The goal was concrete: route incoming RFQs through a fast lane — quicker than before and more accurate.
Why now
- Faster RFQ response:Answer RFQs before competitors finish estimating, and win more of them.
- Margin protection:Over- and under-pricing both cost money. Calibrated estimates protect margin on every job.
- Scarce expert time:Senior engineers add more value on parts that need judgment than on routine quotes.
The Werk24 × Saphirion solution
Drawing in, price out — no human in the loop
The moment a drawing arrives, a three-stage pipeline runs end to end and writes the estimate straight into the customer’s CRM. Werk24 reads the drawing; Saphirion prices the part.
- Input: an RFQ arrives with a 2D drawing — PDF or TIFF, exactly as suppliers send them
- Extraction (Werk24): dimensions, tolerances, GD&T, material, and processes — structured data in seconds
- Cost model (Saphirion): a proprietary model turns those features into a price per piece, about a minute end to end
Extract only what drives the cost
The pipeline doesn’t read a drawing like a human. It reads it like a cost engineer, because it learned from measured reality.
- Ground truth is measured: the cost model is calibrated on real manufacturing costs from past production, not on old estimates
- Cost drivers are known: those actuals show which drawing details truly move cost — geometry, tolerances, material, processes
- Extraction is targeted: Werk24 pulls exactly those drivers and filters out the rest before pricing
How the benchmark ran
- Week 1: the customer shared production drawings with their actual manufacturing costs attached, under NDA
- Weeks 2–4: Saphirion tuned the cost model to the customer’s process profile while Werk24 extracted the cost drivers from every drawing
- Week 5: estimates were compared head-to-head against actual production costs and against the customer’s own expert estimators
All figures come from the customer’s production data (March 2026), benchmarked against actual manufacturing costs. The customer is anonymized at their request.
2.4× more accurate — in every cost bracket
Average error vs. actual cost
The pipeline’s mean error — less than half the 18.5% from expert estimators, so 2.4× more accurate.
Systematic bias (median)
Almost perfectly calibrated, versus +10.9% padding from expert estimators.
Average error per part
Down from €35.05 — more than halved, with the biggest gains on high-value parts, where errors hurt margin most.
Parts priced better than the experts
The AI beat the manual estimate on more than three in four parts — including the sub-€100 parts once the estimators’ strongest range.
By developers, for developers
Clean APIs, predictable JSON, confidence scores, and webhooks. Go from PDF to production in hours, not weeks.
pip install werk24 # Install Werk24
werk24 init # Get your API key: Pay as You Go, no minimum
werk24 health-check # Verify connection
Run the same benchmark on your data
Five weeks, your parts, your actual costs — and the same report, with your numbers in it. Decide on rollout with evidence, not promises.