My role: Designed and coded both agents.

Executive Context#

A steel-trading owner ran sales and deliveries largely from memory and a desk full of paper. Two things kept costing him money. Customers who asked for a quote were often not followed up in time. And plates were sometimes held up at entry or delivery because the test certificate did not match.

The agents now process approximately 30 leads and 70–80 plates a day and save the owner 1–2 hours a day.

The Actual Problem#

There were three leaks, and none of them was about the steel itself.

Late follow-up cost the price. A company would ask for a product; two or three hours later a middleman asked for the same material. The middleman buys in bulk, so the rate was almost the same, and the original customer never got a better price. They did not become a long-term customer.
Regional demand went unnoticed. When one region ordered heavily within a day or two, other companies nearby probably needed the same material. He could not follow up with them, so margin-rich repeat sales were missed.
Certificates halted deliveries. The plate quality was fine. The test certificate sometimes was not, and that stopped plates at entry and at delivery. The cost showed up hour by hour as delay.

Diagnostic Approach#

AI Necessity: The useful tasks were structured lead capture and document exception checks inside the existing trading workflow.
Data Sufficiency: Leads needed purchase history, nearby-buyer context and current or upcoming yard stock. Plate checks needed both mark photographs and certificates.
Methodological Fit: Voice input from the trader captured leads; local OCR read plate marks and certificate content, with an LLM agent supporting the workflow.
Infrastructure Readiness: A local web application ran on the trader’s desktop with local GLM-OCR and an LLM accessed through an API.

Strategic Intervention#

The first agent turns his voice input into a structured lead entry, then checks recent purchases, nearby buyers and yard availability. It clubs related leads so the right people are followed up while the interest is fresh. It is not an automated calling agent.

The second agent reads plate-mark photographs and mill test certificates as plates arrive, attaches the right certificate to each plate, and flags mismatches for review. When a piece is cut or delivered, the correct certificate goes to the client.

The first useful result came within two days. A working prototype took two weeks, and the final version took three more.

Decision Ownership#

The owner approves next steps such as merging or changing leads before the marketing team acts. People review plate exceptions. The agents support the existing workflow rather than make final commercial or quality decisions.

Outcome#

Metric	Before	After	Context
Owner time	Manual follow-up and plate checks	1–2 hours/day saved	Reported by the owner
Follow-up, conversion	Late and missed follow-ups	Improved	Margin gain not yet quantified
Daily volume	—	~30 leads, 70–80 plates	Both agents in daily use; changes continuing
Time to first value	—	~2 days	Prototype in 2 weeks, final version in 3 more
Running cost	—	~$30/month	The owner tops up this much LLM usage

He describes the business gain as sizeable but cannot yet separate it from other changes, so no margin figure is claimed. No validation-accuracy percentage has been supplied either.

Strategic Takeaway#

Small agents can earn their place by taking recurring work out of the owner’s day while leaving consequential approvals with people.

Giving 1–2 hours a day back to an SMB owner with agentic AI