Manufacturing gets pitched AI more aggressively than almost any other sector, and most of it targets the wrong layer. Predictive maintenance on vibration data, computer vision on the line, autonomous scheduling: real technologies, long implementation cycles, and rarely the first thing a factory should buy. AI agents solve a smaller and duller problem, which is why they pay back faster. They handle the paperwork, the coordination, and the chasing that sits around production and consumes people who should be doing something else.
§ 01The boundary that decides everything
Every serious conversation about agents in a factory starts by drawing one line. Below it sits operational technology: PLCs, SCADA, machine control, safety interlocks. That layer is deterministic, certified, and designed so that a wrong value cannot become an injury. Above it sits the information layer: ERP, MES, quality records, maintenance tickets, shift plans, supplier email.
Agents belong above the line, without exception. A language model is probabilistic by construction, and probabilistic systems do not belong in a control loop where the failure mode is physical. This is not caution for its own sake. It is the difference between a wrong output costing a correction and a wrong output costing a person. Any vendor willing to blur that line is telling you something important about how they engineer.
§ 02What works above the line
The deployments that survive their first quarter share a shape: high volume, rule heavy, judgment at the edges, and a human who confirms before anything is committed.
- Supplier and customer communication. Incoming order confirmations, delivery date changes, and specification questions arrive as unstructured email. An agent reads them, extracts the relevant fields, checks them against the ERP, drafts a reply, and escalates anything that does not match. Purchasing stops being an inbox job.
- Quality documentation. Non-conformance reports, certificates, and inspection records need to be filed consistently and found later under audit pressure. An agent structures what an inspector writes in plain language, files it in the right place, and flags missing evidence before the auditor does.
- Maintenance triage. Requests arrive by phone, paper, and message. An agent normalizes them into tickets, assigns priority from history and asset criticality, notifies the right technician, and chases open items nobody closed.
- Shift and workforce coordination. Absences, shift swaps, qualification checks, and overtime rules are pure logic wrapped in constant exceptions. An agent proposes coverage that satisfies the rules and hands the decision to a supervisor. This is the pattern behind the manufacturing deployment documented in our production cases.
- Shop floor knowledge. Procedures, drawings, and past fixes exist somewhere in the network and nobody can find them at 3 AM on nights. An agent answers from the documents you already have and cites where the answer came from, which matters more than the answer itself.
§ 03What does not work, and why
Machine control is the obvious one and it stays off the table. Beyond that, three patterns fail quietly enough to be worth naming.
- Full autonomous scheduling. Production scheduling looks like an optimization problem until you learn that a specific operator is faster on a specific machine, that one customer always calls to expedite, and that half the constraints live in a planner's head. Agents are good at preparing a schedule and explaining tradeoffs. Committing it without a planner is how you learn what was not written down.
- Anything requiring data that does not exist. If downtime reasons are recorded as a free text field that everyone fills with the word other, no agent will produce insight from it. The honest first project is sometimes capturing the data properly, and that is not an agent project.
- Processes that change every week. Agents encode rules. If your process is genuinely different each month, the agent will always be one revision behind and the maintenance cost will exceed the savings.
§ 04Integration is the project
In manufacturing more than any other sector, the cost is not the AI. It is reaching your systems. A modern ERP with a documented API is a few days of work. A twelve year old MES with a database nobody has permission to touch, a planning process that runs on a shared spreadsheet, and a supplier portal that only exports CSV by email is where weeks go. Integration is typically 60 to 70 percent of build effort, and the discovery call exists mostly to find out which of those two worlds you live in.
One practical consequence: read-only access is the correct starting point. An agent that reads production data, prepares proposals, and writes nothing carries almost no risk and still removes most of the manual work, because in most processes the reading and preparing is the slow part. Write permissions come later, one process at a time, with approval gates in front of each. The reasoning behind that sequencing is in our engineering notes.
§ 05What a first deployment should look like
Pick one process with high volume, low blast radius, and visible errors. Supplier email handling and maintenance triage are the usual first choices because a mistake is noticed the same day and costs a correction, not a shipment.
Run it with a human approving every action for the first few weeks. Then measure one number: how often the person changes what the agent proposed. If that number is high, the rules are wrong and you have learned it cheaply. If it is low and stable, you have earned the right to remove approval from the routine cases while keeping it on the exceptions. Expanding before that number is stable is the single most common way these projects fail.
§ 06Cost and timeline
A single agent covering one process end to end costs 3,000 to 5,000 USD and takes 2 to 3 weeks from discovery to a monitored deployment. A coordinated fleet spanning purchasing, quality, and maintenance costs 15,000 to 50,000 USD over 6 to 12 weeks. Manufacturing projects sit at the upper end of any tier when the systems are old, which is most of the time. The full breakdown of what drives price is in our cost guide.
──
If you want to know whether a specific process in your plant is a sensible first agent, that is a 30 minute discovery call. We will also tell you when the answer is that your data is not ready yet, which is a cheaper thing to hear now than after a build.