Inbound RFQ-to-quote agent for make-to-order manufacturers

Job shops and contract manufacturers lose days answering requests for quotation: an estimator opens the email, studies drawings and CAD, judges manufacturability, guesses machining time and checks material and subcontractor prices. An agent does the first pass: it extracts specs, flags DFM issues, estimates cost from geometry and comparable past parts, and drafts the quote in the ERP. The estimator reviews and approves. Faster, more consistent answers win more orders.

Problem & value

• Today: RFQs arrive as unstructured mail with PDF drawings, STEP files and Excel BOMs. Quoting is senior-estimator work, so it is a bottleneck: slow replies lose orders, and tacit knowledge sits in a few heads. Quotes are often inconsistent, and the actual costs seldom flow back into pricing.
• AI value: parallel first-pass handling of every RFQ, same-day response, systematic DFM feedback, feedback loop from actual job costs, and estimators who focus on complex or strategic quotes.
• Evidence (vendor claims, unverified): CloudNC states 5-9 minutes per RFQ with its Quote Agent versus 2-4+ hours manually [1]; DigiFabster claims a CAD-driven analysis and about 99% accuracy depending on the technology [2]. Treat these as marketing figures and benchmark on your own parts.

Variants

• CNC job shop (milling/turning): geometry features (holes, pockets, threads), setups, tool time.
• Sheet metal / laser / bending: flat pattern, nesting, bend count; mostly 2D-driven.
• Casting, forging, injection moulding: tooling cost plus piece price; long lead times.
• Additive and prototype services: marketplace-style instant quote on a web portal (Xometry-like) [3].
• Assembly / BOM quoting: multi-part RFQs with purchased parts and supplier price requests.
• Engineer-to-order: low automation; the agent only structures the request and finds similar past jobs.

How it works

1. RFQ inbox (email/portal)
2. classify & extract (customer, quantity, delivery, material, standards, attachments)
3. parse drawing (VLM/OCR: tolerances, surface finish, GD&T, notes) + CAD (STEP
4. features, volume, bounding box)
5. DFM check (rules + LLM: thin walls, deep pockets, unmachinable tolerances)
6. similar-part retrieval from past jobs (geometry + text embeddings)
7. cost model (cycle time, setups, material, tooling, outsourced steps, margin)
8. clarification questions to customer if data is missing (draft)
9. draft quote + lead time in ERP (capacity and material availability)
10. estimator approval
11. send
12. after the job: actual vs. estimated cost written back for retraining

• Spec extraction: a vision-language model reads drawings; the CAD kernel (not the LLM) computes geometry facts. Keep numbers from geometry deterministic.
• Cost estimation: gradient-boosted or regression models on features and past actual times, combined with rule-based machine and routing logic; output a range, not a point value (see PRED-032).
• Agent layer: an orchestrator calls tools (CAD parser, cost model, ERP lookup, stock check, supplier request) and writes the draft quote. It never sends a quote or commits capacity without approval.
• Human-in-the-loop: estimator approves every quote at first; later, auto-release only for repeat parts below a value threshold with high model confidence.

Data preparation

Data needed: incoming RFQ mails and attachments (PDF drawings, STEP/CAD, spec sheets), past quotes and their outcomes, routings, machine hour rates, material prices, supplier quotes and delivery times. In an SMB this lives in the ERP, in Outlook mailboxes, on file shares with grown folder structures, in Excel calculation sheets and partly on paper. Typical gaps: drawings as scans or inconsistent PDFs, past quotes without a link to the actual cost and outcome, cost rates that exist only in one employee's spreadsheet, duplicate part numbers and missing material or tolerance fields.

Concrete steps: export quotes, orders and actual times from the ERP ([PREP-005]), tidy the calculation spreadsheets into tables ([PREP-011]), digitise old drawings ([PREP-006]), tag parts with material, process and complexity ([PREP-010]), harvest win/loss and actual-cost labels from closed orders ([PREP-020]) and collect good example quotes ([PREP-016]). Document extraction is covered in [DOC-014]. Tools: Paperless-ngx or the DMS, OCR/parsers, n8n for exports, a simple database.

Rough share of effort: 40 to 50 percent. Start with 20 to 50 well-documented past quotes rather than the whole archive.

Implementation paths

Which path for whom

• SMB: B. The quote must use own cost rates, machine data, routings and ERP, and few products fit a specific shop; building on cloud APIs in an EU region on top of existing ERP and mail is more realistic than a standard product. Buy (A) only if your parts fit a vendor's instant-quote model.
• Enterprise: B. Quoting is a core competitive process tied to ERP/PLM; own orchestration and cost models avoid vendor lock-in.
• Developer: C. Open-weight document and vision models with an open-source workflow stack (e.g. n8n) can run in the shop's own network when drawings are trade secrets.

Path A — Buy (off-the-shelf)

• AI quoting for machine shops: CloudNC Quote Agent (UK; RFQ inbox, DFM, quoting; launch stage in 2026) [1].
• Instant-quote / MES-ERP suites: DigiFabster (CAD-based instant quoting, CRM/accounting integration) [2]; Xometry's quoting engine is API-first and mainly serves its own marketplace [3]; aPriori (digital manufacturing cost management, enterprise-grade) [4]; also Paperless Parts and Hexagon-/Siemens-based estimation modules.
• EU/DACH ERP/MES with quoting modules: ams.Solution, Proalpha, abas, Odoo (EU), plus CAD-viewer based tools. Check EU hosting, ERP connector, handling of confidential customer drawings.
• Choose A when: processes are standard (CNC, sheet metal) and a vendor supports your ERP.

Path B — Build on cloud APIs

• Models: frontier multimodal LLMs with structured outputs for drawing and email extraction (Claude, Gemini, GPT, Mistral in EU regions); a reasoning model for DFM review.
• CAD: Open CASCADE (pythonOCC / build123d / CadQuery) to read STEP and compute features; commercial kernels (CAD Exchanger, Datakit, Spatial) if needed.
• Orchestration: n8n or a custom agent service; ERP integration via API (SAP, Odoo, Microsoft Dynamics, Proalpha); cost model in scikit-learn/LightGBM trained on your job history.
• Choose B when: you have unusual processes or want your own cost logic as competitive asset.

Path C — Self-host (open source)

• Models: open VLMs for drawings (Qwen-VL family), Docling/PaddleOCR for PDFs, an open-weight LLM (Qwen, Llama, Mistral families) served via vLLM; pgvector for similar-part search.
• Stack: Open CASCADE + FreeCAD libraries for geometry, LightGBM/XGBoost for cost, n8n or Camunda.
• Hardware: HW3 for a ~30-70B model and VLM; HW2 is enough for a small pipeline.
• Choose C when: drawings are export-controlled, defence-related or strictly confidential.

Path D — Custom training

Train your own cycle-time and cost regressors on 1,000+ past jobs with actual times; fine-tune a drawing reader only for unusual drawing conventions. Geometric deep learning (feature recognition, 3D embeddings) pays off at high volumes or where many families of parts exist.

State of the art (2026)

• VLMs now read most mechanical drawings (title block, tolerances, notes) reasonably well, but dimension-to-feature association, GD&T and dense multi-view drawings remain error-prone; review is needed.
• Geometry-based feature recognition (holes, pockets, threads) is mature in commercial CAM/quoting tools; AI adds toolpath-informed time estimation (CAM-in-the-loop) as in CloudNC [1].
• Marketplaces (Xometry and others) have shown instant quoting at scale for standard processes [3]; the shift in 2026 is bringing it to individual shops' own RFQ inboxes via agents.
• Limits: one-off, complex or tight-tolerance parts have large estimation error; costs depend on shop-specific machines, fixturing and tooling that models do not see; prices drift (materials, energy). Historical data embeds old inefficiencies and win-bias (you only see actuals for won jobs).

Shared resources & combinations

• Same stack: PRED-032 (cost models from CAD), DOC-013 (drawing understanding), KNOW-020 (similar-part search), DOC-015 (tender/RFQ analysis), SCI-024 (CAD copilot).
• Upstream/downstream: AGENT-001 (inbox triage) feeds RFQs; after order, AGENT-004 (order entry); supplier prices for purchased parts via AGENT-005; OPT-009 (process optimisation) tunes routing.
• A VLM/LLM server on HW3 and the ERP connector can be shared across these.

Effort, cost & KPIs

• Path A: weeks to 3 months incl. master data (machine rates, materials); SaaS subscription, indicative range reported by vendors/aggregators from low thousands to tens of thousands USD/year [2].
• Path B: 4-9 months for production quality; main cost is integration and cleaning historical data.
• Path C/D: add GPU hardware and ML engineering; calibration of cost models is continuous.
• KPIs: RFQ response time, share of RFQs quoted, quote hit rate (win rate), estimator hours per quote, estimate vs. actual cost error (MAPE), margin realised, number of DFM issues caught before order.

People & change

Involve management for goals and pricing authority, the head of sales/estimating as process owner, and the estimators who know the real cost drivers: they supply gold-standard quotes and later release every quote. IT and the data owners of ERP and drawing archives are needed for exports and access. Customers and suppliers are affected indirectly: their drawings are confidential, so check NDAs and tell them how files are processed (GUIDE-010).

Works council: depends. If a works council exists, the tool is usually not about employee monitoring, but if it logs which estimator changed or released which quote, it is suited to monitor behaviour and §87(1) no. 6 BetrVG applies (BAG: objective suitability suffices). Inform it anyway under §90(1) no. 3 as work processes change (GUIDE-004). Personal data is mostly business contact data; the DPO should confirm the AVV for any cloud API (GUIDE-005).

Run a pilot with one product group and two or three estimators for four to eight weeks (GUIDE-009). Typical concern: "the machine will replace my judgement"; address it by keeping the estimator as the one who approves price and delivery date. AI literacy: support staff skills with a short training on checking extracted values (GUIDE-006, GUIDE-008).

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III use. Art. 4 AI literacy applies. If RFQ replies are sent by a chatbot-like system directly to people, mind Art. 50 transparency [5].
• GDPR: mostly B2B data; contact persons in mails are personal data, so use a DPA with the provider and EU hosting where possible.
• Confidentiality/IP: customer drawings are trade secrets; NDAs may forbid sending them to third-party clouds or using them for training; check export-control (dual-use) status of parts.
• Commercial risk: a wrong quote can bind the company; keep approval gates and validity clauses.
• Mitigations: estimate ranges, human sign-off, deterministic geometry calculations, audit trail of inputs and assumptions, prompt-injection filtering on incoming attachments.

Getting started

1. Export 200-500 won and lost quotes with actual job costs; check data quality per part family.
2. Pick one process family (e.g. turned parts) and list required quote fields and DFM rules.
3. Pilot a product (Path A) or a build with shadow mode: agent drafts, estimator quotes as usual; compare.
4. Measure error and hit rate; add auto-release for repeat parts only after stable accuracy.
5. Write actual costs back to the ERP and retrain quarterly.

Sources

1. CloudNC — Quote Agent — https://www.cloudnc.com/quote-agent
2. DigiFabster — instant quoting solution — https://digifabster.com/products/instant-quoting-solution/
3. Xometry — Instant Quoting Engine — https://www.xometry.com/
4. aPriori — manufacturing cost management — https://www.apriori.com/
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. CloudNC — why quoting is the bottleneck in manufacturing — https://www.cloudnc.com/blog/why-quoting-is-the-bottleneck-in-manufacturing
7. Open CASCADE Technology — https://dev.opencascade.org/