Most manufacturers have run an AI pilot by now. Far fewer have an agent doing real work in production. The gap is rarely the model. It is the choice of use case, the connection to systems of record, and whether risk and quality teams trust what the agent does.
Start where the volume is
The fastest payback comes from work that is high-volume, repetitive and governed by rules people can already explain. In a manufacturing business, that work is everywhere:
- Chasing suppliers for purchase order acknowledgements and revised delivery dates
- Tracing an engineering change to every affected part, drawing and open order
- Triaging quality deviations and linking them to batch, lot and supplier history
- Checking supplier certificates and documents against what was ordered
None of these make a dramatic demo. All of them consume hours of skilled time every week, and all of them follow patterns an agent can learn and repeat.
Connect the agent to the systems of record
An agent that can only suggest is a chatbot. Value arrives when the agent can read the BOM, open the purchase order, draft the email and prepare the update for approval. That means real integration with PLM, ERP and email from the first use case, not a side tool someone copies answers out of.
This is also why process definition matters. Agents follow processes engineered to L4, where every step, decision and system touchpoint is written down. Where that detail does not exist yet, defining it is part of the work.
Keep people in control
Risk and quality teams will not allow an agent near production without three things:
- A named approval step before anything changes in a system of record
- A full audit trail of what the agent read, decided and proposed
- Monitoring of accuracy, volume and exceptions, reviewed like any other production system
Designing these in from the start is what turns a pilot into a platform you can scale.
Measure against a baseline
Before the agent goes live, measure the process as it runs today: hours spent, cycle time, error and rework rates. Measure the same things in production. Without a baseline, value is a matter of opinion; with one, the next use case funds itself.
Where to begin
Pick one process where the volume is high, the rules are clear and the data is reachable. Put one agent into production on real work, with a human checkpoint and a measured baseline. Then scale the pattern to the next function.
