AI in a metal fabrication shop: five use cases that pay for themselves

Metal fabrication shops resemble each other on one point: the value rests on a handful of people. The estimator who knows the prices by heart. The project manager who knows which press brake is free. The welder who guesses the right joint preparation without opening the procedure. As long as those people are there, everything runs. The day one of them leaves, the knowledge leaves with them.

At a steel structure fabricator in the Eastern Townships, the picture was typical: roughly 100 GB of drawings, specifications and history sat dormant in network folders, unfindable at the moment the estimator needed them. The full case study sets out that engagement; this article draws the lessons that apply to most shops.

Quotes go out at the pace of the expert preparing them, and every public tender arrives with a hard deadline. An assistant that indexes your past quotes changes the posture: the estimator asks “what have we quoted that is comparable to this walkway?” and gets the relevant files, the prices used and the particulars of the contract, with the source cited down to the page.

The enterprise brain principle applies in full: the assistant prepares a basis for the estimate, the estimator adjusts, decides and signs. The speed comes from the search; the judgment stays human. The gains are to be measured at your site, quote by quote.

Counting parts by hand on drawings is slow, and it is the surest way to miss a plate or get a length wrong. An extraction agent reads the DXF files and produces a structured bill of materials: profiles, lengths, quantities, weights.

The output never goes straight into production. It arrives as a checkable table, with a reference back to the original drawing, and someone from the shop validates it before it feeds purchasing or cutting. That is the rule of the game: AI counts, people confirm.

CSA W47.1 and W59, CWB certification requirements, the conditions of the RBQ licence issued by Quebec’s building authority: these frameworks govern every public contract, and nobody reads them cover to cover in the heat of the moment. A conversational search engine indexes your procedures, your work instructions and your compliance documents, then answers by citing the document and the page. No invented answers: if the source does not exist, the assistant says so.

The most lasting effect lies elsewhere. By formalizing the questions and the answers, you capture the veterans’ knowledge before they leave: the foreman who asks “what joint preparation for this profile?” enriches, answer after answer, the company’s memory. Knowledge that retires is poorly replaced; it is much better captured.

The same certified welder is scheduled on two contracts in the same week. The press brake is booked twice. The crane arrives before the steel. These conflicts are usually discovered when the floor grinds to a halt; a constraint-optimization agent spots them in the schedules while there is still time to act.

The agent proposes resequencing scenarios; the project manager chooses. Nobody hands the plant over to an algorithm: you hand it the chore of cross-referencing dates and resources, the one nobody has time to do by hand.

It is the least spectacular use case, and often the first delivered: the production meeting is transcribed then summarized, and the decisions, the owners and the action items come out structured. In the engagement described, the meeting and procedure agents made up the first wave, the one that builds trust.

The value goes beyond the writing saved: decisions stop evaporating. Three months later, “who decided to change the bolt supplier, and why?” has a documented, dated answer, attached to its context.

Deploying everything at once is the surest way to fail. The engagement described planned five agents delivered in waves, with a decision point at every milestone: measure, adjust, continue or stop. The first measurable value was targeted at around two months, carried by the meeting and procedure agents.

On the hardware side, a single server installed in the plant is enough: a GPU with 48 GB of memory runs open models in the 24-billion-parameter class, under the Apache 2.0 licence, so with no royalty and no per-token subscription. Our comparison of models to host yourself sets out the current options.

One rule is not negotiable: no sealed drawing is produced automatically. The engineer’s seal remains a human act, as does weld certification. AI prepares the deliverables; the expert validates them and answers for them.

In the package prepared with this client, public support covered roughly 45% of the implementation invoice, combining ESSOR and NRC IRAP. ESSOR stream 1B reimburses 50% of eligible expenditures up to $20,000, and stream 1C, 50% up to $50,000, for SMEs with 250 employees or fewer and revenue of at least $2.5M; applications can be filed until March 31, 2027.

NRC IRAP is aimed at SMEs with fewer than 500 employees, the C3I credit covers 15% to 25% of computer hardware acquired before 2030 depending on the region, and Services Québec funds team training up to 75%. Our overviews of Quebec and federal support set out the conditions and the documents to prepare.

Start with the inventory: where do your drawings, specifications, work instructions and past quotes sit? Then pick a first low-risk use case (meeting summaries are a good candidate), set an indicator before you start and give yourself a real decision point at the first milestone.

Cogio designs this kind of enterprise brain: custom agents, hosted on your premises, where AI prepares and people decide. To find out what such a project would produce in your shop, let us talk.