Agents as standard work
In lean operations, the best way to do a task is defined once and followed by everyone. AI agents benefit from the same discipline. When each person writes their own prompts, every shift and site gets a different answer to the same question, and nothing can be compared or improved. A well-designed operations agent has fixed instructions, a fixed output structure and a named purpose, such as "summarise this activity for the next tier meeting".
Typical tasks
| Task | Example output |
|---|---|
| Summarise | A status summary of an activity's history before the meeting |
| Highlight | Repeat findings across checklists, lines or sites |
| Suggest | A draft fishbone or countermeasure plan for the team to challenge |
| Act | Creating or updating an action when a person instructs it |
Where agents fit
Agents add the most value when their output lands inside the existing management system rather than in a separate chat window. A summary should appear at the tier board; a suggested countermeasure should become an action with an owner; a new checklist should become part of standard work. Humans stay accountable for decisions and for verifying that actions worked.
Data and context
An agent can only reason about what it can see. Incidents rarely have a single cause, so the most useful agents can look across safety, quality, production, maintenance and skills records, and into connected systems such as ERP, MES, CMMS and HR. Grounding answers in the organisation's own procedures and specifications, rather than generic internet knowledge, also makes them more relevant and easier to trust.
Governance checklist
- Who defines and approves each agent's instructions?
- Which actions can it take automatically, and which need human approval?
- What data can it access, and at which sites?
- How is output reviewed, and how are errors fed back to improve the agent?
Risks
- Plausible but incorrect summaries when records are incomplete
- People accepting suggestions without challenge
- Agents acting outside clear permissions
- Output that is never connected to an owner or action