AI Agent for Manufacturing: Automation That Reads and Writes Your ERP

What is an AI agent in a manufacturing setting?

An AI agent is a language model wired to tools. In a plant, the tools are API calls: look up a work order, check inventory for a part, read a supplier PDF, create a purchase requisition, send an email. The model decides which tool to call and in what order, within rules you define.

Deloitte's 2026 research found 40% of manufacturers report deploying GenAI and agentic AI, close to the 42% reporting machine learning and deep learning. Many of those deployments are still small, which is how an agent should start.

A concrete example: a customer emails asking where order 4471 stands. The agent identifies the customer, finds the sales order in the ERP, checks the linked work orders and the shipment record, drafts a reply with the current status and expected ship date, and places the draft in a queue for the customer service rep. The rep reads it, edits if needed and sends.

Which manufacturing tasks suit an agent?

Tasks with a clear trigger, a limited number of systems and a checkable result. Order-status requests from customers, supplier acknowledgement matching, expedite emails for late purchase orders, and first-pass review of certificates of conformance fit this profile.

Tasks that depend on tacit judgment, such as deciding whether to accept a rush job that disrupts three other orders, are better handled by a person with the agent preparing the facts.

Another good fit is the exception digest. Each morning the agent compares open purchase orders with promised dates, open work orders with available material, and open quality holds, then sends the planner a ranked list of what needs attention. Nothing is written to the ERP, so the risk is low and the value shows quickly.

• Order-status replies pulled from ERP work order and shipment records
• Supplier PO acknowledgement checked against requested dates and quantities
• Material certificate review against spec requirements
• Daily exception digest for planners: late jobs, shortages, open quality holds

Where do agents fail on the shop floor?

They fail on bad master data, on ambiguous instructions, and on actions with no undo. If part numbers differ between the ERP and the customer's drawings, the agent will match wrongly. If an agent can create a purchase order, it can create a wrong one at volume.

Design for this: restrict tool permissions to the minimum, require approval for writes that spend money or commit inventory, cap actions per hour, and log every tool call with its inputs.

Testing should use historical cases. Replay last quarter's customer emails and supplier acknowledgements through the agent in a sandbox, compare its drafts with what staff actually sent, and review every disagreement. This shows failure patterns before any customer sees an output, and it gives you a baseline accuracy figure to track after launch.

How much does it cost to run an agent?

The model usage cost for a narrow workflow is typically small relative to labor. The larger cost is build and maintenance of integrations. Deloitte lists high cost as the top AI barrier, cited by 43% of manufacturers, so the sensible approach is one workflow, measured, before adding the next.

Track hours saved, error rate against the manual process and the share of cases the agent hands off. Those three numbers tell you whether to expand or stop.

Budget also for prompt and tool maintenance. When a supplier changes a document layout or the ERP adds a required field, someone must update the workflow. A short monthly review of logged failures keeps this manageable.

Can an agent work with an older ERP or MES?

Yes, if there is some interface. REST and SOAP APIs are straightforward. For systems with only a database or file exports, the agent works through a service that wraps those into safe, narrow functions. The agent never gets raw database access.

Authentication follows the same rules as any integration. Use a dedicated service account, give it the narrowest ERP role that works, store credentials in a secrets manager and rotate them on a schedule.