I Brought AI Into My Family's Steel Fabrication Business, Starting With One RFQ Email

Yize — Mon, 26 Oct 2026 10:53:59 +0000

I work in the internet and AI industry, and my family happens to run a steel fabrication business.

After spending a lot of time with AI, automation, and content systems, I kept wondering what these technologies could solve inside a traditional factory.

Our business provides structural steel detailing, fabrication, and overseas delivery services. Internationally, we operate under the FindingSteel brand, with more information available on our structural steel fabrication and delivery services page.

I initially saw many possible applications. AI could help process RFQs, organize business knowledge, improve website content, and support Generative Engine Optimization (GEO).

Once I looked closely at the actual workflow, all of those ideas led back to the same problem.

The factory already had a large amount of data and experience. Drawings, quotations, material standards, equipment specifications, project records, and customer conversations were created every day. They were scattered across email, chat histories, shared folders, and the experience of individual employees.

There was plenty of information, but very little of it could be accessed consistently or handed directly to an AI system.

What happens after an RFQ reaches the factory

An overseas RFQ arrives at a steel fabrication company. The attachments include CAD drawings, a PDF specification, and an incomplete bill of quantities.

The customer wants a quote quickly, but the material standard, surface treatment, destination port, installation scope, and delivery schedule are not fully specified.

Sales reviews the files and forwards them to engineering. Engineering identifies missing details and sends a list of questions back to sales. Sales contacts the customer again. When the additional information arrives, procurement, detailing, and production still need to review it separately.

The files move between email, chat, and shared folders. The RFQ is eventually processed, but the reasoning behind the decisions rarely enters a reusable system.

Steel projects are difficult to standardize completely. The same type of member may be produced under different standards, material grades, connection methods, corrosion-protection systems, inspection requirements, and delivery boundaries.

Important knowledge is distributed across different roles. Sales knows the customer and previous quotations. Engineers understand risks hidden in the drawings. Production teams know the limitations of equipment and processes. Project managers know which details usually create problems on site.

By the time information reaches the website team, it has often been summarized several times. What remains is usually a list of capacity figures, equipment, certifications, and project names.

This information structure affects both internal efficiency and customer acquisition.

An internal AI system can read an old quotation, but it does not know whether the price is still valid. It can extract an equipment specification from an old page, but it may not know whether that number describes axis positioning, a cutting process, or the tolerance of a finished component.

Customers using AI to search for suppliers face the same gap. If a company website only says “experienced” and “high quality,” an AI system has little evidence for answering specific procurement questions.

It needs to know what the company can deliver, which projects the capability applies to, what information the customer must provide, and which commitments still depend on contract and project conditions.

AI can find relevant text. It cannot decide which business fact is approved for use.

Give every business fact a status

Many companies build a knowledge base by importing PDFs, web pages, and internal documents in one batch. The document count increases, but conflicting statements enter the system at the same time.

I believe the first step should be a business fact registry. Each fact should record at least six fields.

Field	Question it answers
Claim	What does the company want to communicate?
Scope	Which company, equipment, process, or project does it apply to?
Status	Is it approved, unverified, contradicted, or retired?
Evidence	Where is the contract, certificate, equipment record, or inspection report?
Owner	Who can confirm and update it?
Reviewed at	When was it last checked?

“Designed annual capacity of 40,000 tons” and “40,000 tons delivered in a specific year” are different claims. Equipment axis-positioning accuracy and finished-component tolerance should not be treated as the same measurement either.

The fact registry needs to preserve these boundaries. Content systems, RAG pipelines, and automation workflows should only use approved versions.

This step sounds basic, but it determines whether the rest of the AI system can operate safely. The more documents a company has, the greater the risk of reusing an incorrect statement.

Once an unverified fact enters automated writing, translation, and multi-channel publishing, it can quickly appear in several languages and on multiple external platforms.

Automate the information handoff first

A steel fabrication company should not let AI complete quotations independently at this stage. Material prices, project risks, detailing effort, transport conditions, and installation scope still require professional judgment.

The RFQ workflow already contains many tasks that are suitable for automation.

Receive the RFQ email and attachments
→ Identify drawings, bills of quantities, and technical specifications
→ Extract materials, sizes, quantities, standards, surface treatment, and destination
→ Compare the request with an RFQ checklist
→ Generate questions for missing information
→ Route the request to sales and engineering
→ Review scope, risks, and quotation conditions manually
→ Draft a response and save the processing record

AI handles extraction, classification, comparison, and drafting. Sales and engineering remain responsible for judgment and commitments.

The system can begin by converting an RFQ into a structured record.

{
  "project_type": "industrial_warehouse",
  "material_standard": null,
  "surface_treatment": "hot_dip_galvanizing",
  "destination_port": "Lagos",
  "installation_scope": "unclear",
  "missing_information": [
    "material_standard",
    "installation_scope",
    "required_delivery_date"
  ],
  "review_status": "engineering_review_required"
}

The value of this record goes beyond saving data-entry time. It gives the company a way to measure which information is most often missing, which questions repeatedly consume engineering time, and which project types create misunderstandings around quotation and delivery boundaries.

Those findings can improve RFQ forms, sales communication, and internal review checklists.

My current view of AI transformation in traditional manufacturing is fairly conservative. The first goal is to let AI turn scattered material into inputs that people can review quickly, then preserve the corrections and decisions made by those people.

Once this workflow is stable, the company can move on to project handoffs, quality documentation, export paperwork, website content, and GEO. A faster system built on unreliable information will only repeat old errors faster.

I would love to hear from developers, manufacturers, and operators working on similar problems. If you are experimenting with AI in manufacturing, RAG pipelines, workflow automation, or GEO, feel free to share your approach, lessons, or disagreements in the comments.

DEV Community: Yize

I Brought AI Into My Family's Steel Fabrication Business, Starting With One RFQ Email

What happens after an RFQ reaches the factory

Give every business fact a status

Automate the information handoff first