DEV Community

Hieu Luong
Hieu Luong

Posted on Originally published at himitek.com

Case Study: How an Electronics Factory Saved VND 1.2 Billion on Components with an AI Agent

1. Pain – When DRAM procurement still runs on manual labor

At 8 a.m. on Monday, Minh, the procurement director of an electronics factory in Bac Ninh, received 137 quotations for DRAM, NAND Flash and controller chips from 12 suppliers. Each supplier used different part numbers, currencies and delivery terms. His team had to open every Excel file, reconcile it against the BOM and request clarification by email.

This manual, chạy bằng cơm process concealed three risks: abnormal price increases, unfavorable payment terms and excessive dependence on one supplier. When DRAM prices move weekly, a report completed two days later may already be obsolete. The ERP records transactions, but it cannot recognize that differently formatted component codes refer to the same SKU.

2. Agitate – Half-finished optimization can erode the margin

Minh once selected the lowest quotation but overlooked its MOQ and expedited delivery fee. The supposedly cheaper order ultimately cost 6% more. Another supplier accounted for nearly 58% of memory-chip volume, creating a nghẽn cổ chai when its delivery schedule slipped.

With approximately VND 15.4 billion in annual component spend under review, an undetected variance of only 2% represents more than VND 300 million. That excludes four employees spending two days on every quotation cycle, production downtime and damage to delivery credibility. Hiring more people only vá víu the process. Attractive dashboards disconnected from purchasing decisions produce KPI ảo. This nợ kỹ thuật can eventually cause data operations to sập nguồn, encourage ăn xổi buying decisions and make the company mất tiền ngu on hidden terms.

3. Solve – A three-step AI Agent workflow without replacing the ERP

HimiTek starts with a DRAM pilot, retains the existing ERP and restricts the AI to recommendations rather than autonomous purchasing.

  • Step 1 – Normalize the data: The Agent ingests Excel files, emails and BOMs, then standardizes currencies, units, lead times and MOQs while mapping supplier codes to internal part numbers. The input checklist covers part number, manufacturer, capacity, speed, Incoterms, validity date and historical delivery failures.
  • Step 2 – Score suppliers and detect anomalies: Current prices are compared with historical medians. Suppliers are scored by price, lead time, on-time delivery and concentration risk. The following rule routes suspicious transactions to a human approval queue.
def review(quote, history_median, supplier_share):
    increase = (quote['unit_price'] / history_median - 1) * 100
    risks = []
    if increase > 8:
        risks.append('price_spike')
    if supplier_share > 0.45:
        risks.append('supplier_dependency')
    if quote['payment_days'] < 30:
        risks.append('unfavorable_terms')
    return {'approval_required': bool(risks), 'risks': risks}
Enter fullscreen mode Exit fullscreen mode
  • Step 3 – Lock down execution rights: OpenClaw Gatekeeper inspects every transaction before the Actuator sends commands to the ERP. The Reasoner and Actuator are isolated to reduce prompt-injection risk. Routing uses 9router v0.4.66 with LiteLLM dual-instance failover, automatic rate limiting, API-key rotation and hard budget caps, such as USD 5 per month for each virtual key, to stop runaway loops. Elevated tools remain locked by default, and purchase orders execute only after explicit authorization.

In this hypothetical case-study model, component expenditure fell by 7.8%, saving approximately VND 1.2 billion per year. The time required to analyze 137 quotations dropped from two days to 45 minutes. The savings came from detecting price variances, consolidating demand at the right time and rejecting unfavorable terms—not from vẽ dự án AI and forcing employees to support it.

4. CTA – Prove the savings before scaling

Start with one volatile component category for 30 days. HimiTek will measure three outcomes: purchase-price reduction, quotation-processing hours and supplier concentration. If the pilot does not produce verifiable savings, the factory can stop without replacing its ERP. If the ROI is confirmed, the workflow can then expand to additional BOMs and factories.

Top comments (0)