Manufacturing cost estimation & instant quoting from CAD (should-cost)

Quoting a custom part has traditionally meant an estimator reading a drawing, guessing setups and cycle times, and looking up material prices; buyers face the mirror-image problem of not knowing what a part should cost. AI extracts manufacturing features from the 3D model, derives a process route and times, and combines physics-style cost rules with regression on the shop's own past quotes and actuals. Result: RFQs answered in minutes instead of hours or days, and fact-based negotiation targets for purchasing.

Problem & value

• Seller side (job shops, contract manufacturers): estimators are a bottleneck; slow answers lose RFQs, and inconsistent estimates between people erode margin. Many shops quote only a fraction of the incoming requests.
• Buyer side (OEM purchasing, cost engineering): supplier quotes are opaque; without a bottom-up model, negotiation relies on price history or benchmarks. Should-cost models expose cost drivers (e.g. tight tolerances, avoidable setups) and feed design-to-cost loops in engineering.
• Marketplaces: platforms such as Xometry, Spanflug and Laserhub offer instant quotes for uploaded CAD files and route orders to partner shops.
• Reported effects (vendor claims, unverified independently): Paperless Parts states quoting a single component drops from up to 2 hours to 15 minutes and complex assemblies from up to 4 days to 4 hours [1]. In a simus customer statement, a Trumpf cost engineer reports cutting calculation effort by 98 percent and a BSH user reports halving the time per part and work plan [2]. Spanflug advertises up to 90% less effort for buyers and suppliers [3]. Treat all as marketing figures; measure your own baseline.

Variants

• Instant quote for contract manufacturers: customer-facing web portal, price and lead time in seconds (CNC, sheet metal/laser, injection moulding, AM, casting).
• Internal estimator assistant: the quote engine pre-fills a proposal; the estimator reviews it. Typical for complex parts and assemblies.
• OEM should-cost / design-to-cost: bottom-up cost models for negotiation and for engineers checking cost impact during design (simus and Tset-style tools, Siemens Teamcenter Product Cost Management, aPriori) [2][4].
• Cost plus CO2: the same process model yields a product carbon footprint estimate (Tset positions cost and CO2 together) [4].
• Private/maker use: hobbyists upload STL/STEP to marketplaces for price comparison; no own AI needed.

How it works

1. CAD (STEP/native) + drawing + material + tolerances + quantity
2. drawing/spec parsing (title block, tolerances, finish, certificates) [DOC-013]
3. geometry analysis: volume, bounding box, features, B-rep graph embeddings
4. manufacturability check + process selection (3-axis/5-axis, turning, bending, moulding ...)
5. time model: setups, cycle time, machine hours (rules + ML correction)
6. cost build-up: material (price index), machine/labour rates, tooling, overhead, margin
7. price with confidence interval per batch size
8. human review on low-confidence or high-value cases
9. quote to customer / ERP
10. feedback loop: actual times and costs calibrate the model

• Geometry: Open CASCADE-based kernels read STEP B-rep; learned encoders such as UV-Net and BRepNet operate directly on B-rep faces/edges for feature recognition and embeddings [5].
• Cost engine: the robust pattern is hybrid. A deterministic process model (feeds, speeds, setup rules, rates) gives explainable cost; gradient-boosted trees or neural nets learn the residual between model and the shop's actual times/prices. Pure end-to-end regression from geometry is less explainable and needs far more data.
• LLM/VLM role: read drawings and RFQ emails, extract tolerances, materials and standards, and explain a quote in words. The numeric price should come from the cost engine, not from the LLM.
• Human-in-the-loop: estimators approve quotes above value thresholds, for novel geometries (out-of-distribution flag) and for tight tolerances; overrides are logged as training data.

Data preparation

Needed data: 3D CAD files (STEP and native formats), 2D drawings (often PDF scans), material and tolerance specifications, historic quotes and orders, and above all post-calculation data (actual cycle times, scrap, material costs) with machine and labour rates. These typically live in the CAD/PDM or PLM system, the ERP, file shares, e-mail attachments (RFQs), the estimator's Excel sheets and paper drawings. Typical gaps: quotes without actual costs, scanned drawings without text layer, inconsistent part numbers and duplicates, tacit knowledge in the estimator's head, outdated rate tables and no consistent process-route labels. Customer and supplier data carry confidentiality duties.

Steps: (1) inventory sources and map them to the use case (PREP-001, PREP-002); (2) export ERP quotes, orders and post-calculations (PREP-005); (3) scan and OCR old drawings (PREP-006); (4) restructure estimator spreadsheets into tidy tables (PREP-011); (5) harvest labels from won and lost quotes and actual costs (PREP-020); (6) capture expert know-how and gold-standard quotes (PREP-016). Tools: CAD converters, OCR/document parsing, Python, a small data warehouse. Expect 50 to 60 percent of effort in data work for an own model, less with a product. See GUIDE-003.

Implementation paths

Which path for whom

• SMB (contract manufacturers): Path A, because mature instant-quoting and should-cost products exist at SMB prices and quoting is a standard process.
• Enterprise (OEMs, large manufacturers): Path C, because cost models are strategic, CAD data is trade-secret IP, and a model trained on own post-calculations beats generic benchmarks.
• Developer: Path C, because open CAD kernels and ML libraries allow integration into own ERP/PLM and custom cost models.

Path A — Buy (recommended for SMB contract manufacturers)

• Quoting software for job shops: Paperless Parts (US; automated part-file analysis for sheet metal, CNC, Swiss machining and AM [1]), Spanflug MAKE (German; calculation software for machining companies [3]), simus classmate (German; cost and emissions from a 3D CAD model in seconds [2]), Kaktus AI. Many ERP/MES vendors (e.g. for job shops) add quote modules.
• Should-cost / cost engineering for OEMs: aPriori, Siemens Teamcenter Product Cost Management, Tset (Germany; cost and CO2 platform plus expert services [4]), simus.
• Marketplaces (buy parts instead of own engine): Xometry (accepts STEP, SLDPRT, STL, DXF, 3MF [6]), Spanflug BUY (network of 8,000+ machines [3]), Laserhub (24/7 instant quote, ~100 producers [7]). Note that uploading CAD to a US platform is an IP and export-control decision.
• Choose A when: standard processes, you can export machine rates, and you do not want to maintain a geometry pipeline. Check EU hosting, NDA/IP terms and whether the vendor trains models on your data.

Path B — Build on cloud APIs

• Geometry stays classical (Open CASCADE, CAD-vendor APIs, Autodesk Platform Services) since frontier LLM APIs do not natively read STEP; render views or send extracted feature tables.
• Drawings/RFQs: a multimodal model (Claude, Gemini, GPT, Mistral) with JSON-schema outputs extracts tolerances, material and finish; EU regions or EU providers for IP-sensitive drawings.
• Cost model: gradient boosting (LightGBM/XGBoost) on features plus rules, hosted in a managed ML service in an EU region. Integration via ERP APIs and a workflow tool (n8n, Power Automate).
• Choose B when: you have a unique process or ERP and want to own the pricing logic.

Path C — Self-host (open source)

• Stack: Open CASCADE/pythonOCC or CadQuery/build123d for STEP reading, feature recognition and volume/area computation; trimesh for meshes; scikit-learn/LightGBM for costs; optionally UV-Net or BRepNet encoders (note: BRepNet's repository is CC BY-NC-SA, non-commercial [8]; verify the UV-Net licence before commercial use). Open VLMs (Qwen-VL family) via vLLM for drawing reading, DOC-013.
• Hardware: the numeric pipeline runs on CPU (HW1); a GPU (HW2) is only needed for drawing VLMs or training B-rep networks. CAD never leaves the company, which is the main reason to choose C.
• Ops effort: moderate; the hard part is the process knowledge, not the infrastructure.

Path D — Custom training

Worth it when you have thousands of historical quotes with work plans and actual times for a repeatable part family. Train per-process regressors (cycle time, setup time, scrap) on engineered geometry features, optionally with a B-rep graph encoder as feature extractor. Data needed: CAD files linked to quotes, work plans, actual machine times and final prices, plus won/lost outcome for price optimisation (see OPT-004). Weak data sets (<~500 parts per process) favour the hybrid rule-plus-residual model.

State of the art (2026)

• Hybrid engines are standard: feature recognition + rule-based process planning + ML correction. Vendors such as Paperless Parts, Spanflug and simus automate feature, time and cost analysis from the CAD file [1][2][3]; marketplaces add auto-quote for more processes (Xometry announced injection moulding auto-quote in October 2025 [6]).
• Learned B-rep representations (UV-Net, BRepNet; CVPR 2021) enable feature segmentation and shape embeddings that also power similar-part search (KNOW-020) [5][8]. They are research-grade building blocks, not turnkey cost predictors.
• LLMs/VLMs shifted drawing intake: tolerances, notes and certificates can be read from PDFs without templates, with errors on dense or handwritten drawings.
• Limits: novel geometries, assemblies, tight-tolerance or special-process parts, supplier- specific rates, and market capacity swings (cost is not price). Unit costs drift with material and energy prices, so connect to index data (PRED-008). Models must show a confidence range; silent extrapolation causes loss-making quotes. No independent public benchmark of quote accuracy exists that we could verify; require a pilot on your own historical parts.

Shared resources & combinations

• Drawing and RFQ intake: DOC-013 (technical drawing understanding) and DOC-015 (RFQ analysis) feed this model; same VLM/OCR server (HW2) as other document use cases.
• Similar-part search: KNOW-020 shares B-rep embeddings and the CAD pipeline; nearest historical parts also serve as price anchors.
• Procurement side: AGENT-005 compares and negotiates received quotes; the should-cost output is its benchmark.
• Pricing and inputs: PRED-008 (material/energy index forecasts) updates material cost; OPT-004 (price optimisation) turns cost into a win-probability-aware price.
• Shop-floor feedback: PRED-014 (process quality / virtual sensors) and MES actuals recalibrate time models.
• Shared blocks: tabular-ml, 3d-vision geometry pipeline, ERP connector, workflow engine.

Effort, cost & KPIs

• Path A: weeks to a few months (machine-rate and material set-up dominates); subscription pricing.
• Path B: 3-6 months for a first process family incl. ERP integration.
• Path C: similar to B plus operations; mostly CPU, low hardware cost.
• Path D: 6-12 months incl. data cleanup of historical quotes and work plans.
• KPIs: quote turnaround time, share of fully automated quotes, hit rate (won/quoted), mean absolute percentage error of predicted vs. actual cost, margin realised vs. planned, estimator hours per quote, supplier price vs. should-cost gap (buyer side).

People & change

Involve management and the sales or work-preparation lead for goals and pricing policy, the estimators as the key end users from day one (they hold the know-how and fear replacement), IT for CAD/ERP integration, and data owners for post-calculation data. Suppliers and customers must be informed when their drawings or prices are processed; check NDAs and contract terms with legal.

Works council: depends. The tool mainly supports quoting, but if it logs estimator overrides or quote times per person it becomes suited to monitor performance, so co-determination under §87(1) no. 6 BetrVG applies; changed work processes also require early information (§90(1) no. 3). A short works agreement (no individual performance evaluation) usually settles it. Personal data is limited (names in drawings or RFQs); involve the DPO for the AVV (Art. 28 GDPR) with any cloud vendor, a DPIA is normally not required.

AI literacy: the AI Act (Art. 4) requires measures to support staff literacy; train estimators to check suggestions, understand confidence ranges and spot out-of-distribution parts. Pilot with one product line and a few champion estimators, compare AI quotes with manual ones, and let the estimator keep final sign-off. Typical concerns: "it will replace me" (it removes routine, experts handle complex parts) and wrong quotes (margin guardrails, human release). See GUIDE-008, GUIDE-004, GUIDE-009.

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III category (no decisions about natural persons). AI literacy (Art. 4) applies to staff.
• GDPR: mostly business data; personal data only in customer contacts and sole-trader suppliers.
• IP and trade secrets: CAD files and drawings are core IP (Directive (EU) 2016/943 on trade secrets). Require NDAs, no training on customer data, EU hosting, access control.
• Export control / defence parts: dual-use or military drawings may not be uploaded to foreign platforms; route to Path C.
• Competition law: a shared should-cost or benchmarking service must not exchange competitor prices or rates; use anonymised, aggregated data only.
• Commercial risk: wrong quotes bind the shop; keep validity clauses, human release thresholds, drift monitoring and out-of-distribution flags.
• Negotiation use: should-cost is an estimate; present it as a model with assumptions.

Getting started

1. Collect 200-500 historical parts with CAD, drawing, quote, work plan and actual time/cost; define the first process family (e.g. milled aluminium parts).
2. Document machine rates, setups, material prices and overhead as the deterministic cost baseline.
3. Pilot 2-3 tools from Path A on your own historical parts and compare against actuals (MAPE).
4. Set release rules: automatic below a value/confidence threshold, estimator review above.
5. Feed won/lost results and actual times back every month to recalibrate.

Sources

1. Paperless Parts — quoting software, part-file analysis and vendor claims — https://www.paperlessparts.com/
2. simus systems — simus classmate, cost and emissions from 3D CAD, customer statements — https://www.simus-systems.com/
3. Spanflug — BUY/MAKE platform and figures — https://www.spanflug.de/
4. Tset — product costing and CO2 platform, cost engineering services — https://www.tset.com/
5. Jayaraman et al., "UV-Net: Learning from Boundary Representations", CVPR 2021 — https://arxiv.org/abs/2006.10211
6. Xometry — Instant Quote, accepted CAD formats, injection moulding auto-quote news — https://www.xometry.com/
7. Laserhub — 24/7 instant quote and network figures — https://www.laserhub.com/
8. BRepNet (Autodesk AI Lab), CVPR 2021, CC BY-NC-SA 4.0 — https://github.com/AutodeskAILab/BRepNet
9. UV-Net implementation — https://github.com/AutodeskAILab/UV-Net
10. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj