
Superintelligence for Metal Fabrication Shops: Orders With Technical Review
How an AI agent organizes orders and lead times for production in metal fabrication shops, and why every measurement needs technical review before confirming.
Want to apply this in your company?
Book a call with Juan →An AI agent in a metal fabrication shop can organize incoming orders, sequence production orders according to machine capacity, and suggest lead times based on the shop floor queue. What decides whether this workflow is useful is technical review: measurements, tolerances and specifications are only confirmed with the customer after an engineering or production lead has checked what the agent prepared. This article shows how the workflow is designed, where AI helps and where it should stop.
"Superintelligence" is a term used heavily in sales communication, but what a fabrication shop needs to run today is a narrow system with a defined task, known data and written limits. It is worth separating the promise from actual use before talking about hiring. For that distinction, see Narrow AI, General AI or Superintelligence: A Scale for Managers.
Key takeaways
- The agent sorts orders and production orders by priority, machine availability and agreed date, but it does not decide on its own what will be produced.
- Measurements, tolerances and materials go through human technical review before any confirmation reaches the customer.
- The deterministic workflow (records, queues, statuses) is kept separate from the part that requires judgment, which is where AI comes in.
- Every lead time suggestion records the data used, the drawing revision and who approved it.
- A sensible start is a pilot on one production line or one type of part, with a success metric agreed in advance.
Where production planning usually loses time
In most fabrication shops, orders arrive through different channels: a WhatsApp message with a photo of a sketch, an email from a sales rep, a spreadsheet from a construction site, or a verbal request written on paper. Each channel carries measurements in a different format, and someone in production has to transcribe them, check them against the drawing and only then build the order. The lead time is often estimated from memory, based on bottlenecks everyone knows, such as the press brake or welding.
This work is not difficult, but it is repetitive, and repetition produces transcription errors and lead times that are too optimistic. A well designed agent reduces the typing and the queue calculation, and leaves people with the part that needs experience: deciding whether a measurement makes sense for that material and that process. For a broader view of where to start in industry, see How to Start With AI in Manufacturing: Maintenance and Quality First.
How the agent organizes orders and production orders
The workflow we recommend separates reading, calculation, proposal and writing. The agent reads and proposes. Writing to the technical system and confirming with the customer is done by an authorized person. A possible sequence of steps:
- Receive the order from the intake channel and extract the essential fields: customer, item, quantity, measurements, material, finish and requested date.
- Compare the item with the product register and with the drawing in the current revision, flagging any divergence instead of correcting it alone.
- Check the open queue and the machine schedule in the production system, without changing the plan.
- Propose a sequence and an estimated lead time, explaining the logic in two or three sentences, for example: "enters after order 1234 on the press brake, with inspection the same day".
- Send the proposal to technical review, without sending any message to the customer.
- After approval, open or update the production order and record who approved it, when, and based on which drawing revision.
Notice that the last step is the only one that changes the system, and it only happens with explicit approval. That separation is what makes the workflow auditable later.
Technical review: what to check before confirming measurements
Technical review exists because a misread measurement turns into scrap, rework and a chain of delays. The agent can speed up the checking, but it should not replace the judgment of someone who knows the process. A useful checklist for the reviewer:
- The measurement and tolerance match the approved drawing in the current revision, not a photo or an older version.
- The requested material and finish are in stock or can be purchased within the expected time.
- The planned process (cutting, bending, welding, machining) exists in the shop and has capacity available that week.
- Orders with non-standard measurements or without a drawing go to engineering, not to the production queue.
- The suggested lead time accounts for inspection, shipping and any painting or outsourced treatment.
A rule that tends to work well: the agent never changes the technical register. It flags the divergence, shows where the information came from and waits. This prevents a misread message from becoming an official record in the shop.
Illustrative scenario: an order with a conflicting measurement
Illustrative scenario: a shop producing bent parts receives a support bracket order through WhatsApp, with a measurement shown on a photographed sketch. The agent extracts the fields and compares them with the drawing registered for that code. The measurement given differs from the current revision, which could be a transcription error or a real change requested by the customer. Instead of building the order, the agent sends the divergence to the engineer, with the photo, the drawing and a specific question. Meanwhile it does not confirm a lead time to the buyer and only replies with a receipt message, within the limits agreed. The engineer confirms that the customer sent a new revision. Only then does the agent recalculate the lead time against the current queue, the production lead approves it, and the order is opened with a complete record.
The value of this scenario is not raw speed. It is having a traceable path for the case where the information does not match, which is exactly the case that costs the most.
Data, records and safety limits
This kind of workflow needs a decision log. For each suggestion, keep the incoming request, the extracted fields, the drawing consulted with its revision, the proposed lead time, the reviewer and the outcome. Good observability practice, described in the OpenTelemetry documentation, is a useful reference for thinking about traces and metrics for the agent, even if the first implementation is simpler.
There are two specific risks. The first is untrusted input: the text of an order can contain instructions that try to change the agent's behavior. Treat all received content as data, never as instructions, following the OWASP guidance on applications that use language models. The second is personal data from buyer contacts, which must be handled under the LGPD, with a clear purpose and restricted access. For risk management more broadly, the NIST framework is a useful reference for structuring the discussion with leadership.
How we apply the method on a production line
Our method comes down to mapping, limiting, piloting and handing over. In production, we start by mapping how orders arrive today, how long each step takes and who reviews what. Next we limit the agent to reading, proposing and drafting, with no writing to the technical register and no direct contact with the customer. The pilot runs on one type of part or one line, with every suggestion logged and a metric agreed before it starts, such as the share of proposals sent back by review or the time between order and proposal. At the end, the company decides whether to expand, adjust or stop, based on what the log showed.
Frequently asked questions
Does superintelligence for metal fabrication already exist in practice? In practice, what exists are narrow AI systems aimed at specific tasks such as reading orders, calculating queues and drafting proposals. A shop does not need a general system to benefit from this, and the promise of a superior intelligence should not replace an analysis of what the tool actually does. Always ask for a description of the tasks, the data used and the limits.
Can the agent confirm a lead time directly to the customer? We recommend not, at least at the start. Confirming a lead time depends on reviewed measurements, available material and real queue capacity, and those are exactly the points technical review checks. A safe approach is for the agent to reply only with a receipt and say the proposal will be confirmed shortly.
Do you need an ERP or production planning system to begin? Not necessarily, but you do need a reliable source for the queue and the machine schedule. A well maintained spreadsheet can serve in a pilot, as long as it has an owner and is updated regularly. The agent reads what exists and flags when the information is out of date, which also helps organize the process itself.
How do you know if the pilot worked? Agree on the metric before starting, not after. Useful indicators are usually the time between order and proposal, the number of divergences found by review, and the share of proposals approved without adjustment. If the metric does not improve, or if review spends more time correcting than the agent saves, the right decision may be to adjust the rules or stop.