AI Adoption in Practice

Quotes, Part Numbers and Export Documents: What Metal Parts Makers Can Hand to AI First

Messy part master data, quotes that depend on a few senior staff, slow replies to RFQs, and invoices and packing lists retyped again and again. Four steps, with the data to prepare, the common pitfalls and how to check the results.

Many hardware, fastener and mold makers who look at AI start with the same idea: let AI do the quoting. The direction is fine, but the order is usually wrong. A quote draws on part numbers, materials, surface finishes, packaging and past prices. If that data is spread across three spreadsheets and one senior salesperson's memory, AI will only produce wrong quotes faster. Below are four steps, in the order we recommend.

01Step 1: Clean up the part master first

Sales calls a screw "M6x20 stainless hex", the warehouse writes "SUS304 HEX M6*20", and the customer's RFQ uses its own part number. People can tell these are the same item. Systems cannot. Reading RFQs, drafting quotes and filling export documents all depend on mapping these names to one internal part number.

What to prepare: the current part list from your ERP or spreadsheets, one to two years of shipping and quotation records, and any customer part number cross-reference you already have. At minimum the master should hold internal part number, item name in Chinese and English, specification (size, thread pitch, length), material, surface finish, unit, packing method and quantity per carton, net weight per piece, HS code, and active or discontinued status. Step 3 needs net weight and HS code, so fill them in now.

Three common pitfalls. First, trying to clean every part number at once and losing six months. Start with parts shipped in the past year and leave discontinued items for later. Second, the same meaning written several ways, such as "304", "SUS304" and "stainless 304". AI is useful here for a first pass: let it group entries and flag likely duplicates, then have someone who knows the products decide what to merge. Third, nobody maintains the master afterward and it drifts again within months. Name an owner and route every new part number through a fixed process.

How to check it: sample recent delivery notes and confirm that every line maps to exactly one internal part number, required fields are filled, and every suspected duplicate has a decision.

02Step 2: Let AI read RFQ emails and draft the quote; people set the price

Slow RFQ replies are often not about pricing. The delay is in the preparation. A customer lists twenty items in an email, some with drawings, some with a spreadsheet, some just saying "same as last time". Sales has to copy out each line, match part numbers and look up the last quote before any pricing starts. That preparation is work AI handles well.

Have AI read the email and attachments and produce a table: customer item name or part number, matched internal part number, specification, quantity, requested delivery, trade terms, and past transaction prices. Items it cannot match, or that lack information, are marked for review along with the questions to ask the customer. The unit price column stays empty or shows only a reference price. Sales or a manager sets the price based on material costs, capacity and the account.

What to prepare: the part master and customer cross-reference from Step 1, past quotes and order history, your current quote template, and your quoting rules such as minimum order quantities and packaging surcharges.

Common pitfalls: generative AI can produce content that looks plausible but is wrong. The U.S. National Institute of Standards and Technology calls this confabulation in its generative AI risk profile, NIST AI 600-1, and also warns that people tend to over-trust automated output. So every part match should show its basis, meaning which email, attachment and line it came from, so the person approving the quote can check it. RFQ emails also contain the names and contact details of customer staff, which count as personal data under Article 2 of Taiwan's Personal Data Protection Act. When choosing a tool, confirm where the data is stored and whether it is used for model training, and control account access.

How to check it: run the process on RFQs you have already quoted. Compare the items, part numbers and quantities AI extracted against what your staff prepared at the time, and log each error by type. Where errors cluster, fix the master or the cross-reference. Once in use, every quote is approved by a person before it goes out.

03Step 3: Generate invoices and packing lists from templates, then check each line

After an order comes in, the same data is often retyped several times: sales order, delivery note, commercial invoice, packing list, and sometimes a customer-specific format. Every retyping is another chance for an error.

Export documents carry legal requirements. Under Article 6 of Taiwan's Regulations Governing Customs Clearance, Inspection and Release of Export Cargo, a packing list must be submitted with the export declaration (with exemptions such as single-package shipments or uniformly packed goods of one kind). Article 8 requires the packing list to state the specifications and the gross weight, net weight and quantity of each package. Article 12 requires tariff classification to be declared truthfully according to the Customs Import Tariff and Classification of Import and Export Commodities of the ROC.

Fix your invoice and packing list templates, and pull item name, specification, quantity, unit price, gross and net weight, carton count, shipping marks and HS code from the order and the part master. AI compares the customer PO against the order and flags mismatches, such as a PO quantity that differs from the packed quantity, or a customer-required item description that differs from the master. The documents themselves come from the template rather than free generation by AI, which keeps formats and figures consistent.

What to prepare: recent export document samples, format requirements from key customers, weight and packing data in the master, and the corrections your customs broker asks for most often. A common pitfall is using theoretical net weight that does not match the scale, so gross weight should be filled in from actual packing results. Another is changing packaging or carton quantity without updating the master, which makes every later document wrong.

How to check it: before each shipment, the shipping team checks the documents against the PO, delivery note and packing record, and logs how often and why the broker or customer sends documents back for correction. That log drives the next round of fixes.

04Step 4: Feed every finding back into the master

After the first three steps, most teams find that quote and document errors trace back to the master data. Set up a feedback loop: new customer part numbers found during quoting go into the cross-reference, actual weights measured at shipping go back into the master, and discontinued or substitute parts are clearly marked.

One example we can share is a hardware materials business handling several thousand part numbers. It built its e-commerce, ERP and retail counter around a single product data source, which cut duplicate entry across systems and manual cross-checking. The point is that data is maintained in one place and every other system reads from it. AI and automation need that foundation to produce stable results.

05Where to start

For most factories we recommend Step 1, but not for the whole catalog. Pick the one or two customers with the largest share of revenue and build a clean master and cross-reference for the parts they ordered in the past year. A small scope shows results sooner and gives you real data to test Step 2.

If your part master is already in reasonable shape and export paperwork is the biggest burden on the shipping team, start with Step 3. Template-based documents are the easiest to verify. Leave quote drafting until the master is stable, and keep pricing decisions with people.

Before starting, agree on who owns the master, which past RFQs and documents you will test against, and what counts as passing. Then choose tools.

Sources

Figures and regulations in this article come from the official sources below. Please cite the originals.

  1. [1]出口貨物報關驗放辦法(全國法規資料庫)
  2. [2]個人資料保護法第 2 條(全國法規資料庫)
  3. [3]NIST AI 600-1: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
Portrait of Kai Hsu (KK)

Author

Kai Hsu (KK)

Founder & lead advisor, Twouring Co., Ltd.

LinkedIn →

Start with a 30-minute discovery call to confirm your next step

The call is free. We confirm the scope of the problem, the budget and the decision-maker, then give you initial recommendations and a practical next step.