Agents on the factory side of the business

Walk into a plant with a pitch about autonomous decision-making and you will meet a room that has been automating things since before anyone said AI. They are not sceptical because the technology is new. They are sceptical because they have specific and hard-won opinions about what happens when a system acts on bad input.

Those opinions are usually correct, and they make the conversation better.

The line nobody should cross casually

Control systems are not a place for probabilistic software. Equipment that moves, heats or cuts operates under safety regimes that exist because the failure modes are physical.

Our position is straightforward: agents belong on the information side. They read, assemble, draft, flag and prepare. They do not adjust setpoints and they do not participate in a safety chain. This is not a limitation we expect to relax.

Stated plainly at the start, this usually makes the conversation easier rather than harder, because it is the concern the room walked in with.

Where the value actually is

Maintenance documentation. Decades of manuals, revisions and service records in inconsistent formats, where the relevant paragraph exists and finding it takes forty minutes. Retrieval with provenance turns that into a lookup.

Shift handover. The information exists across logs, notes and memory. A drafted handover assembled from the record, reviewed by the outgoing shift, is a small change that compounds daily.

Quality investigation. When a defect pattern appears, the work of assembling what changed across materials, settings, shifts and suppliers is slow and mechanical. An agent that assembles the timeline with citations lets an engineer start from analysis rather than gathering.

Supplier and compliance correspondence. High volume, structured, and largely drafting work.

None of this is glamorous. All of it removes hours from weeks.

What has to be true first

Data residency is usually decided before anything else. Production data frequently cannot leave the site or the organisation, which makes self-hosted or air-gapped deployment the default rather than an option.

Boundaries have to be enforced at the point of action. "The agent has read-only access to the historian" is a statement that must be true at the system level, not a configuration someone could change without review.

Every action needs an audit trail. In an environment where an incident may be examined by a regulator, an insurer and a customer, a record of what an agent did and what it drew on is not a nice-to-have.

The adoption pattern that works

Start with documentation retrieval, because it is useful immediately, has no path to physical consequence, and teaches the organisation how the system behaves before anything riskier is considered.

Give the agent a named owner on the operations side rather than in IT. Write evaluation criteria from real past cases. Set a budget that blocks. Widen only when the scores stop being interesting.

Next

The manufacturing and deployment pages cover self-hosted and air-gapped options. The Agent Orchestration System pages cover policy enforcement, audit and evaluation.