The Real Cost of Industrial Quotation AI: Lessons from Taiwan

Recently in Taiwan, we built an AI quotation assistant for an industrial distributor. The biggest cost in the project was not model API calls. It was an unmanaged Excel sheet that had circulated across departments for years.

The client manages contracts and finance from Taipei, engineering data from Hsinchu, and warehouses in Taichung and Kaohsiung. Field sales reps operate in Vietnam and Thailand, while component suppliers are based in Shenzhen, Dongguan, Japan, and Taiwan.

The company distributes connectors, sensors, and automation controllers. The initial goal sounded simple: "Build a trilingual AI assistant. When an RFQ arrives, the system should match part numbers, check stock, calculate volume pricing, and generate a quotation with datasheets."

The CEO asked if we could deliver in four weeks.

I told him that connecting the model takes one week, but cleaning internal data determines the rest of the schedule.

---

Week One: Part Number Conflicts

During the first week, a sales rep in Vietnam received an urgent RFQ with 23 line items.

For one optical sensor, the customer used an internal code VN-BN-4421. The ERP listed the manufacturer code. Warehouse staff in Taichung had shortened the code to save typing time. The sales Excel sheet combined this sensor with a similar-looking model in the same row.

Both models had identical dimensions, but opposite electrical logic:

• One used PNP output (active high). The other used NPN output (active low).
• One used an M12 connector. The other used an M8 connector.

Sending the wrong part meant a short circuit when plugged into a PLC, halting a customer factory line.

In early testing, the model used semantic search and produced a clean quote. An application engineer in Hsinchu noticed the wiring difference during manual review and stopped the quote before delivery.

The model guessed because the company had no canonical record linking codes to physical parts.

---

Four Separate Systems: Fragmented Pricing and Stock

We discovered four separate pricing standards:

1. Taipei Finance: Maintained USD price books and foreign exchange rates.
2. Southeast Asian Sales: Used converted local currency sheets with custom discounts.
3. Taichung Warehouse: Kept old promotion rates in spreadsheet notes.
4. Procurement: Relied on informal supplier messages.

Inventory data showed similar gaps. The ERP showed 500 units in stock, but 300 were already committed to another contract. Supplier lead times covered factory dispatch, not customs clearance or shipping.

Origin rules posed another risk. Goods shipped from Taiwan do not gain preferential tariff status if they were simply transshipped through bonded zones. Language models cannot deduce trade compliance without explicit rules.

---

Engineering Guardrails and Data Governance

We paused automatic quote sending and established clear boundaries:

1. Product Identity Graph: Every item links manufacturer codes, internal SKUs, and customer references. Suffixes cannot be omitted.
2. Pipeline Separation: Vector databases handle static datasheets. Live APIs handle inventory and deterministic pricing.
3. Mandatory Sign-off: Human specialists must approve substitute parts, origin claims, and non-standard discounts.

When asked to remove approval steps, we shared the PNP/NPN incident. The company agreed that preventing faulty shipments outweighed saving seconds on quotes.

---

Shadow Mode Results

We ran a two-week shadow mode where sales teams quoted manually while the AI produced background drafts for audit.

Over six weeks and 100 live RFQs:

• Quotation Time: Dropped from 40 minutes to about 10 minutes.
• Error Interception: The system caught one expired certificate, one packaging error, one incompatible part, and one questionable origin claim.

Quotes still require manual confirmation before sending. Text generation is quick, but pricing commitments carry commercial responsibility.

---

Summary

In cross-border supply chains, the primary engineering challenge involves reconciling operational habits across departments and cleaning unmanaged spreadsheets.

Before deploying an AI quoting tool, three questions must be answered:

1. What is the canonical record for identifying a product?
2. Who holds legal authority to commit to price and delivery?
3. Can you audit which data and approval generated a quote?

Forward-Deployed Engineering focuses on organizing these operational details so AI can deliver reliable business value.