Case Study: How a Mechanical Factory Saves 120 Hours per Month with an AI Quotation Comparison Agent

1. Pain — Spending 90 minutes on one quotation package can still result in the wrong material

At 7:45 a.m., Hung, procurement manager at a mechanical factory in Binh Duong, receives 12 quotations for bearings, steel plates and milling cutters. Supplier A sends a PDF, supplier B uses Excel, while supplier C submits a photo through Zalo. The same bearing appears under several names: 6205-2RS, 6205 2RS or a Vietnamese description meaning a 6205 bearing sealed on both sides.

Three employees must open every file, copy prices into a comparison sheet and verify specifications, VAT, delivery dates and payment terms. This rice-powered, manual process takes an average of 90 minutes per package. One misread steel thickness or an overlooked exclusion of shipping can turn the apparently cheapest quotation into the most expensive choice.

The risk is not a lack of data but inconsistent data. Years of patched-together spreadsheets create technical debt, while managers often see only the final amount. A specification error may therefore remain hidden until the purchase order has already been issued.

2. Agitate — Slow approval becomes a bottleneck for the production schedule

When procurement becomes the bottleneck, late materials lead to idle machines, waiting workers and delayed customer orders. If one employee takes leave or burns out from overload, the comparison sheet may remain untouched for two days. The number of quotations processed can still produce an attractive dashboard, but it is a vanity KPI when the factory buys the wrong item or overlooks the option with the lowest total cost.

At 120 manual hours per month and an assumed fully loaded labor cost of VND 90,000 per hour, the factory spends approximately VND 10.8 million merely reading and re-entering data. If monthly purchasing volume is VND 2 billion, identifying unfavorable differences of 3–5% can retain VND 60–100 million. Hiring more people for a quick fix or creating yet another half-optimized spreadsheet only increases overhead. It is avoidable dumb money lost to controllable errors.

3. Solve — Deploy a controlled procurement AI Agent in three steps

• Step 1 — Pilot one material group: Select 50–100 historical quotations for bearings or steel plates. Consolidate PDFs, spreadsheets and images in one location. Define mandatory fields including material code, specification, quantity, unit price, VAT, shipping cost, delivery date and payment terms.

Step 2 — Normalize data and raise flags: The AI Agent extracts information, maps equivalent descriptions to a canonical material code, and compares total cost, delivery time and commercial terms. A minimum validation rule can be implemented immediately:

def flag_quote(q, baseline):
    alerts = []
    if q['spec'] != baseline['spec']:
        alerts.append('Specification mismatch')
    if q['unit_price'] > baseline['unit_price'] * 1.05:
        alerts.append('Price exceeds baseline by over 5%')
    if not q.get('payment_term'):
        alerts.append('Missing payment terms')
    return alerts

The output is a recommendation table containing source files, ranking reasons and explicit warnings—not an unverifiable chatbot response.

• Step 3 — Restrict execution rights: HimiTek routes requests through 9router v0.4.66 and dual-instance LiteLLM failover. OpenClaw Gatekeeper separates the AI reasoner from execution scripts, blocks elevated tools by default and requires approval from an authorized employee before a purchase order can be created. Each virtual key can receive a hard budget cap, such as USD 5 per month during the pilot, together with rate limiting and automatic API-key rotation to prevent runaway loops from exhausting the quota.

Following the pilot, processing time per quotation package falls from 90 to 20 minutes. The factory saves approximately 120 working hours per month, shortens its purchasing cycle by 60% and reduces purchasing costs by 3–5%. People continue to approve material orders with financial or production impact; the AI handles repetitive reading, matching and exception detection.

4. CTA — Measure ROI before expanding

Do not spend six months drawing up an oversized AI project. Start with one material category and the latest 50 quotations, then track three metrics: minutes per package, specification-error rate and price differences detected. HimiTek can build a Gatekeeper-controlled comparison pilot so the factory sees the hours and money retained before extending the workflow across procurement.