AI for Welding Shops: Quotes, Job Tracking and Material Certs

An AI agent for a welding or fabrication shop is a governed role that reads drawings and requests, drafts quotes from the shop's own rates and past jobs, follows each order from material receipt to shipment, and files material certificates against the parts they cover. Its value comes from memory that compounds: corrected estimates, supplier documents and inspection outcomes accumulate in one record, so each new quote starts from what the shop already knows. The useful first step is not the agent itself but the rules it will follow, starting with material records, drawing revision control and the named person who signs off on inspection.

Key takeaways

• Link material traceability first: heat numbers, certificates and the parts they cover must connect before an agent touches quoting.
• Build quotes from your own cutting, welding and finishing rates, and keep every assumption visible to the estimator who approves it.
• Give drawing revisions one rule: the agent may flag a newer revision, but it never re-prices a quote from it without a person deciding.
• Keep inspection sign-off with a named person; the agent prepares the packet, checks what is missing and tracks open items.
• Measure the work by fewer reworked quotes, fewer chased certificates and fewer jobs waiting on paperwork, not by how many documents the agent touches.

Why a welding shop is a memory problem

Most fabrication shops run on knowledge that lives in several places at once. The estimator remembers that a particular joint on thin-wall tube takes longer to fit up than the drawing suggests. The office keeps rates in a spreadsheet that was last reviewed when a supplier changed prices. Job details sit in emails, paper travelers, photos of fixtures and PDF certificates saved to whatever folder was open that day. When someone leaves or a busy month arrives, the shop loses the reasoning behind its numbers, and quotes drift.

A compounding agent addresses that gap by keeping a structured record of what the shop has actually done. It reads past quotes alongside the final job cost, the certificates received for each material lot and the inspection results. Over time it can answer practical questions: what did a similar bracket cost to make last time, which supplier certificates were incomplete, and which drawing revision a shipped order was built to. The important word is governed. The agent reads and writes only the records it has been granted, and every change to a rate or a certificate link is traceable to a person or a defined process.

This is the same retrieval pattern used in many business applications: the model answers from documents the organization supplies rather than from general knowledge. The architecture is described in the [Google Cloud retrieval-augmented generation guidance](https://docs.cloud.google.com/architecture/rag-capable-gen-ai-app-using-vertex-ai), and the trade-off between retrieving records and changing a model is laid out in the [AWS comparison of retrieval and fine-tuning](https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html). For a shop, retrieval is usually the better starting point because the records change every week and must stay correctable.

What a compounding agent would change in quoting

Quoting is where most shops feel the pressure first. A request arrives as a drawing, a sketch, a photo or an email with a parts list. Someone has to identify materials, thicknesses, joint types, processes, quantities and finishes, then price the labor, consumables, outside services and shipping. Done carefully, that takes an experienced estimator a good part of a morning. Done quickly, it leaves gaps that surface as change orders.

A well-scoped agent would take that intake and produce a draft. It would extract the parts list into a structured sheet, match each part to similar completed jobs, pull the current rate table and propose a price with its assumptions listed beside each line. The estimator then accepts, edits or rejects each assumption. Nothing is sent to a customer until a person approves it. The draft also records which drawing revision and which rate table version it used, so the number can be explained later.

The quality of this depends on what the agent is allowed to reference. Past quotes are useful only when the final job cost is attached, otherwise the agent learns from estimates rather than outcomes. The rate table must have an owner and a date. Shops that handle shop drawing packages should also set rules for how drawing-based quotes are compared, which is covered in more detail in the Be Human guide to cabinet shops and shop drawings, since the same discipline applies to steel fabrication.

How job tracking changes after the order is won

Once an order is accepted, the shop needs a reliable answer to one question: where is this job right now and what is blocking it? A traveler on paper answers that only for the person holding it. A spreadsheet answers it only if someone updates it at the end of each shift. An agent can keep the record current by reading the inputs the shop already produces, such as material issue, cut confirmation, fit-up completion, weld completion, required examinations, finishing and dispatch.

The agent should not invent status. It should only move a job forward when a defined input arrives, and it should flag a job that has sat in one state longer than the shop expects. A flag might say that a part has been cut but no weld record exists, or that a finished assembly has no matching certificate for its steel lot. Those flags go to the production lead as a short list, not as a stream of alerts. This is the sort of nonconformance tracking that quality teams already manage, and the Be Human article on AI for quality management describes how records of that kind are kept and reviewed.

Material certificates: what to file, what to match, what to flag

Material certificates are where traceability is won or lost. A mill test certificate or supplier certificate of conformity states the grade, chemistry, mechanical results and a heat or lot identifier for a batch of material. To be useful during a customer audit, that certificate must be linked to the material actually received, the parts cut from it and the jobs that used those parts. The chain is simple to describe and tedious to maintain by hand.

An agent can take on the filing work. When a certificate arrives, it reads the grade, size and heat number, compares them with the purchase order and the receiving record, and checks that the document is complete and legible. It then links the certificate to the material lot and to any job that has issued that lot. If the heat number on a certificate does not match the tag on the bar stock, the agent holds the link and flags the mismatch. It does not guess which certificate belongs to which bar.

There is also a security point. Supplier PDFs are external documents, and text inside them should be treated as data to be read, not as instructions the agent should follow. The [OWASP guidance on prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) explains why content from outside the business needs separation from the agent's own instructions. In practice, the agent should extract fields into a fixed format and have no authority to change a rate, approve a lot or alter a record on the strength of anything written in a document.

Illustrative scenario: a shop receives a certificate for a bar lot whose heat number differs by one digit from the tag on the rack. The agent does not link the certificate to the parts already cut. It places the lot in a hold state, lists the three jobs that might use it, and asks the receiving clerk to check the physical tag. The lot is released only after the clerk records the correction.

Drawing revisions: define the rules before any automation

Drawing revision control is the rule most shops have informally and few have written down. A quote may reference revision C of an assembly drawing. The shop floor may cut from revision B because that copy was printed last week. A customer may send revision D after the job has started. Each of these situations is ordinary. Each becomes expensive when nobody can say which revision governed which decision.

Before any agent touches drawings, the shop should settle a short set of rules, written as a checklist that the team can read in one sitting:

• Every drawing has one register entry with its number, current revision letter, issue date and the person who released it.
• Every quote and every job record states the revision it was built against.
• Superseded revisions are archived and kept reachable, never deleted or overwritten in place.
• When a newer revision arrives for an open quote, the agent lists the differences and routes them to the estimator, without re-pricing on its own.
• When a newer revision arrives for an open job, production is told before further cutting or welding against the old copy.
• Only the drawing owner can release a revision, and the agent cannot.

These rules turn drawing changes into visible decisions. They also define where the agent is allowed to help, which is comparison, routing and record keeping, and where it is not, which is releasing or re-pricing. Shops that want a fuller view of how records like these are governed can read the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework), which organizes AI oversight into mapping, measuring and managing risk in the setting where the system is used.

Inspection sign-offs and where the agent stops

Inspection is the point where the agent's role has to be most precise. Depending on the product and the customer, a job may need visual checks, dimensional checks, weld examinations or witnessed tests, each with its own acceptance criteria and its own signer. The agent can assemble the inspection packet: the drawing revision, the welding procedure references, the material certificates linked to the assemblies, the inspector's checklist and any open nonconformances. It can check that every required entry is present before the packet goes to a person.

What the agent should not do is mark a weld or dimension as accepted. The sign-off belongs to a named inspector or responsible person, and that name appears on the record. This is the approval boundary to define in the first engagement. A simple test is to ask, for each action, whether a wrong result could reach a customer without a human decision in between. If yes, the action is a person's. If no, such as chasing a missing signature or noting that a certificate has not been filed, the agent can handle it.

Be Human designs and operates this kind of setup as a managed service. A first engagement defines one useful role, such as quote drafting or certificate filing, the memory it reads and writes, the tools it may call, its approval boundaries and an observable success criterion. Ongoing operation covers monitoring, recurring evaluations against past quotes and certificates, permission reviews and controlled changes when a rate table, a supplier format or a drawing process changes. Work is remote, in English or Portuguese. The aim is an agent that the shop can explain, correct and switch off, not one that runs unseen.

For evaluation, the practice of testing an agent against known past cases is the most useful habit to borrow. The [OpenAI evals guide](https://platform.openai.com/docs/guides/evals) describes how to write checks against expected outputs, which for a welding shop means taking twenty or so finished quotes and certificates, running the agent on the intake documents, and seeing whether it reaches the same parts list, the same heat matches and the same open flags.

Where to start: the first task worth handing over

Most shops should not begin with quoting, even though it is the most visible task. Begin with the records the other tasks depend on: a complete material register, a drawing register with revision rules, and a list of jobs with their certificates. Once those are reliable, certificate filing is usually the safest first agent role, because its output is checkable and its errors are easy to catch. Quote drafting can follow once the rate table has an owner and past jobs carry their final costs.

Take the free Be Human AI diagnostic to find the first task worth handing to an agent in your shop.

Frequently asked questions

Can an AI agent read our mill test certificates and link them to material lots? Yes, for the reading and linking work, provided the certificates and the receiving records share an identifier such as a heat or lot number. The agent extracts grade, size and identifier, compares them with the purchase and receiving records, and links only when the fields agree. When they do not agree, it holds the link and flags the mismatch for a person to resolve against the physical tag.

Does the agent decide whether a weld or a part passes inspection? No, and it should not be designed to. The agent can prepare the inspection packet, confirm that required entries are present and track open nonconformances. The acceptance decision stays with a named inspector or responsible person, and that person's name is recorded on the sign-off.

What should we settle before any AI work on quoting? Settle the rate table and the drawing revision rules first. Each rate needs an owner and a date, and each quote needs to record the rate version and the drawing revision it used. Past quotes are most useful when they are matched to final job costs, so that the agent learns from outcomes rather than from estimates alone.

How do we know whether the agent is helping? Compare the work before and after on a small set of measures you already track, such as how many quotes were reworked after approval, how many certificates were chased late and how many jobs waited on paperwork before shipment. Run recurring checks against past cases so that a change in the agent's behavior is noticed early, rather than discovered during a customer audit.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.