Blog

qAI agents in manufacturing operations

AI agents in manufacturing operations: what a person keeps

AI agents in manufacturing operations carry work between your systems. Five workflows, the step a person keeps in each, and the step each job refuses.

4 min readAgent design, Orchestration

by Grisha Pertaia

illustration
  1. session
  2. credential
  3. notify
  4. score
  5. write
  6. intake
  7. route · redacted· client identifiers
  8. score
  9. notify

Most of the work between a manufacturer's systems is reading in one place and typing in another. A buyer reads a supplier reply and moves a date in the ERP. An estimator reads a drawing and builds a quote. A quality lead collects certificates from three folders into one pack. The systems are fine. The gap between them is filled by people.

That gap is where AI agents in manufacturing operations earn their place. The question an operations lead asks next is the right one: if an agent carries the work, which step does a person keep? This post answers it for the five workflows on the manufacturing page. The answer is different in each, and the reason is always the same: a person keeps the step that commits the company.

The rule behind every answer

Each job gets its own credential, for that job only. The credential lists what the job may do and expires when it ends. Anything outside that list is refused, and the refusal goes on the record with the rule it hit. A step you mark as sensitive waits for the person you name, who sees the step and the limit it crossed. Approve and reject both go on the record.

So for each workflow there are two questions. Which step waits for a person? Which step is refused outright, even when it looks helpful?

Purchasing and supplier follow-up

A person keeps: the new delivery date. The buyer approves it, because a date change moves what the plant plans and what finance expects.

The job refuses: sending a drawing to an address that is not on the supplier record. An address that arrives in an email attachment is not a supplier record.

This one is built. You can run it in the sandbox on demo data, and the supplier follow-up workflow lists each step.

Quoting and RFQ handling

A person keeps: the price and the delivery date. The estimator sets both and releases the quote, because together they commit your company to a customer.

The job refuses: holding time in the production schedule. A quoting job may read the schedule. Promising capacity belongs to your planners. See RFQ and quoting. It is designed and has not run in the sandbox yet.

Engineering change

A person keeps: whether the change goes ahead, and from which order or serial it applies. The change board approves the notice, because that decision sets which parts are built to which drawing.

The job refuses: writing over a released revision. The job may add the new revision. The old one stays, because it is the evidence of what was built. The steps are on the engineering change workflow, which is designed.

Quality document packs

A person keeps: the signature on the pack. It is your company's statement to the customer that the parts conform, so it belongs to the quality lead.

The job refuses: writing or copying a certificate. A certificate stands for a test on one lot. The agent collects certificates, matches each one to its lot and flags the gaps. It never makes one. See the quality document pack workflow, which is designed.

Reconciliation between operations and accounting

A person keeps: each adjusting entry. The controller who signs the close approves each one, because an entry changes what the books say for the period.

The job refuses: posting into a period that is already closed. The job drafts into the open period and leaves any reopening to a person. The reconciliation workflow is designed.

Where a wrong action costs money

Read the refusals again as a list of costs. A drawing sent to the wrong party. Capacity promised twice. A part built to an overwritten drawing. A certificate nobody tested. Figures changed in a reported period. Each one is a single action that an agent with broad access could take in good faith, because a document asked it to or because it looked like the fastest way to finish.

That is why the limit is set per job, not per agent. An agent that follows up suppliers on Monday and reconciles receipts on Tuesday holds two different credentials, each with its own short list.

What this does not do

It does not replace your ERP or any other system. It works between them. It does not decide which steps are sensitive: your people do, and Ares enforces the list. It does not clean your data. If your data is not ready, the assessment finds one workflow where it is ready enough and says what to fix for the next.

It can run on a machine in your building, and nothing leaves unless you allow it. Local models differ from the largest cloud models, and we say which runs where.

If an agent in your ERP already works, keep it. Its output still has to reach your other systems, and that gap is what we build.

Keep reading