AI in manufacturing

AI Agents with Human Approval: Design Patterns for the Shop Floor

Agents should do the legwork and stop before anything irreversible. Patterns for approvals, limits and audit.

Trikaan Research & Engineering10 Nov 20264 min read

KEY TAKEAWAYS

Agents should prepare work and stop before anything irreversible.
Define triggers, tools, limits and approvers explicitly.
Every agent action needs an audit trail.
Start with agents that save preparation time, not ones that decide.

AI agents can carry out multi-step tasks: gather data, compare options and draft actions. In manufacturing, where actions move material and money, the design question is not ‘can the agent do it?’ but ‘where must a person decide?’

Here are the patterns we use.

Anatomy of an agent

Element	Example: procurement agent
Trigger	Projected stock below cover
Tools	Stock, plan, vendor history, price lists
Output	Draft purchase request with vendor comparison
Limits	Approved vendors only, value cap
Approver	Buyer for the category
Log	Inputs, reasoning summary, approval

Pattern 1: Draft and approve

The agent prepares a complete draft — schedule, purchase request, work order — with its reasoning, and a named person approves, edits or rejects. This is the default for anything that changes plans, orders or money.

Pattern 2: Notify and suggest

The agent watches for conditions and alerts the right person with context and a suggested action. Useful for quality drift or maintenance risks.

Pattern 3: Bounded automation

Low-risk, reversible actions within strict limits can run automatically — for example, sending a scheduled report or creating a reminder task — with logging and an easy way to undo.

Guardrails

Agents act with the permissions of a named user or role.
Value and scope limits enforced by the system, not the prompt.
Clear display of what the agent will do before approval.
Kill switch per agent.
Regular review of approval and rejection rates.

Where to start

Choose tasks with heavy preparation and clear decisions: re-planning after a breakdown, shortage-driven purchase requests, daily MIS preparation. Measure time saved and approval rates, then expand.