In custom manufacturing — sheet metal, machining, stamping, plastic molds — slow quotation is one of the most overlooked profit leaks. A single RFQ drawing travels from sales to process engineering, then to a veteran estimator who prices it from memory. Two or three days pass. Some customers move on; others wait, and the quote comes back 10–20% off the mark.
This article is not about using an LLM to write marketing copy. It is about combining drawing parsing, similar-part retrieval, and a cost model so that custom-part quoting moves from “the senior estimator’s gut feeling” to “the system proposes a range, a human approves”. We close with measurable acceptance criteria and a pragmatic phased rollout.
1. Why quoting is slow: three hidden bottlenecks
Most companies assume quoting is slow because they are short-staffed. In practice, the time is consumed in three places, and hiring more people does not fix any of them.
Bottleneck 1: Drawing information is never structured. RFQ drawings arrive as vector PDFs, scans, DWG files, or photos of a hand sketch. Material grade, plate thickness, tolerance class, surface treatment, and heat treatment requirements are scattered across title blocks, technical notes, and section callouts. Sales cannot read them all; process engineers must inspect every sheet.
Bottleneck 2: Routing depends on individual experience. A veteran knows how many operations a part needs, how many setups, whether a simple fixture is required. A junior adds a safety factor and produces a wildly different number for the same part.
Bottleneck 3: Historical quotes are not comparable data. Quotes from three years ago sit in spreadsheets or paper files that nobody can search. Every new RFQ starts from zero.
| Step | Traditional time | Main bottleneck | What can be automated |
|---|---|---|---|
| Receiving and extracting drawing data | 20–60 min per drawing | Mixed formats, scattered fields | Parse material, dimensions, tolerance, surface finish |
| Determining the process route | 15–40 min per drawing | Depends on one person | Reuse routing from similar parts plus rule constraints |
| Cost calculation | 20–30 min per drawing | Manual material and labor lookup | Auto-fill material library and labor rates |
| Approval and customer response | 0.5–2 days in queue | Requires someone available | System proposes a range; human only signs off |
2. The four building blocks
2.1 Drawing parsing layer
The output is a structured JSON record: part type, overall dimensions, critical dimensions and tolerances, material grade, plate or wall thickness, surface treatment, heat treatment, quantity, and delivery requirement. There are two engineering paths:
- Vector PDF / DWG: extract the text layer and layers, then match title blocks and technical notes with rules. Highest accuracy, lowest cost.
- Scans and photos: OCR plus layout analysis plus a vision model to identify views and callouts, followed by human spot-checking. The error rate of this layer determines whether the whole system is trusted, so a human review loop is mandatory.
A practical lesson: do not chase 100% automatic parsing. Instead, highlight low-confidence fields so the process engineer spends one minute correcting two or three fields instead of reading the drawing from scratch.
2.2 Similar-part retrieval layer
Use hybrid retrieval — geometry features, material, and process tags as hard filters, drawing embeddings for similarity ranking. Return the top five historical similar parts with their actual cost and quoted price.
This is the highest return-on-effort part of the entire system: it turns a veteran’s experience into a searchable case library without training any model.
2.3 Cost model layer
Custom-part cost splits into five items: material, machining time, subcontracting, scrap and rework, and overhead plus margin. Each must become a computable formula rather than a fudge factor.
| Cost item | Formula basis | Data source |
|---|---|---|
| Material | Volume × density × unit price × (1 + scrap rate) | Material library plus daily price feed |
| Machining time | Operation time × machine rate | Routing rules plus measured times from similar parts |
| Subcontracting | Surface treatment / heat treatment priced by area or weight | Supplier price list |
| Scrap and rework | Historical defect rate banded by difficulty | MES and inspection records |
| Overhead and margin | Banded by customer tier and lot size | Company pricing policy |
2.4 Quotation explanation layer
The LLM does not guess prices here. Its job is to translate the computed result into a human-readable quotation note: why this part is 18% more expensive than last time, whether the increase comes from material or operations, and which assumptions the customer must confirm (for example, an unspecified tolerance defaults to a given ISO grade).
One hard rule: the model never sends a price to a customer. Every quote is confirmed by a human; model output is advisory only.
3. Defining accuracy so it can be accepted
The most common failure mode is a vendor promising “95% accuracy” without defining accuracy. Put these four metrics in the contract.
| Metric | Definition | Phase-1 target | Acceptance method |
|---|---|---|---|
| Quote deviation | |system price − actual deal price| / actual deal price | Median ≤ 15% | Back-test on 50 historical RFQs |
| Similar-part recall | Share of cases where a true similar part appears in the top 5 | ≥ 80% | Manual labelling of 100 samples |
| Field extraction coverage | Share of key fields (material, dimensions, tolerance, surface finish) successfully extracted | ≥ 90% | Sampling 200 drawings |
| Time per quote | From drawing upload to proposed price | ≤ 3 minutes | System telemetry, P90 |
Note the comparison baseline: custom-part quoting always involves negotiation, so comparing against the final deal price is more reliable than comparing against the veteran’s original number, which carries its own bias.
4. A four-phase rollout
Phase 1: Data inventory (2–3 weeks)
Collect the last twelve months of RFQs, drawings, quotes, and actual cost sheets, then check how many can be matched (drawing ↔ quote ↔ actual cost). This decides feasibility: with fewer than 300 matched sets, fix the data before building the system.
Phase 2: Parsing and retrieval PoC (4–6 weeks)
Skip the cost model for now. Deliver “upload a drawing → structured fields → top five similar parts”. Let the process team use it for a week and count how many fields they had to correct. Cheapest PoC, strongest signal on data quality.
Phase 3: Cost model in parallel (4–6 weeks)
The system proposes prices while the existing process continues unchanged, running in parallel for 4–8 weeks. Review the gap weekly, analyse the outliers, and iterate the cost rules. Do not rush to cut over; let the estimators build trust first.
Phase 4: Cut over and operate
The system proposes a range plus an explanation; the estimator approves and leaves an audit trail. Review deviation monthly and update material prices and labor rates.
5. Five common failure modes
- Parsing only, no cost loop. Extracted fields are an intermediate artifact. If they cannot become a price, business users will not adopt it.
- Historical data that does not reconcile. If drawings, quotes, and actual costs cannot be matched, the model learns noise.
- Letting the model set the price. LLMs are unreliable arithmetic engines. Prices must come from the cost formula; the model only explains.
- Ignoring the format mix. If 70% of RFQs are scans but the solution is designed for vector PDFs, it fails at go-live.
- No human correction path. If estimators cannot fix fields quickly, they abandon the tool after two uses.
6. Budget and vendor selection
A reasonable budget for this kind of system is USD 20k–70k (RMB 150k–500k), driven by three variables: drawing format complexity, the share of historical data that reconciles, and whether it must integrate with the existing ERP (writing quotes back, validating customer credit limits).
When evaluating vendors, check three things: whether they have delivered custom-manufacturing projects rather than generic OCR; whether they will commit to a measurable deviation target; and whether they support on-premise deployment — drawings are core commercial assets and most manufacturers will not upload them to a public cloud.
For on-premise architecture, see our private LLM deployment guide. If you are also building production-line visual inspection, the drawing and image pipelines can share infrastructure.
Conclusion
The hard part of custom-part quoting is not precision arithmetic. It is the three losses of information between drawing and price: drawings not fully read, routing kept in one person’s head, history that cannot be searched. AI compresses those three losses, letting estimators move from “read and calculate” to “review and negotiate”.
If you are evaluating such a project, start with 30 historical drawings: can the system extract material, dimensions, and tolerance, and can it find similar parts? If you need help running that check, contact us for an assessment template tailored to your industry.
