Let an AI agent propose a work order before granting it authority to create, change, release, or close one. For each write action, decide who approves it, what evidence and action are recorded, how an error is corrected, and what happens if the underlying information changes before execution. A generated recommendation is not an authorization to commit labor, order a part, or alter equipment status.
The OWASP guidance on excessive agency recommends least privilege, downstream authorization, and human approval for high-impact actions. The permission ladder below is a practical local design proposal, not a published risk standard or a claim about any particular connector’s default settings.
Start by distinguishing a proposal from a record change
A planner may accept an assistant’s summary of a pump alarm while rejecting its proposed priority. That requires a visible draft and source references. Creating a CMMS record changes the system of record even when the order is labeled draft; some configurations can trigger notifications, workflows, or downstream integrations. Releasing an order can schedule labor or reserve parts. Closing one can alter equipment history or compliance reports.
Map actual consequences in your environment. “Read-only” access can expose sensitive maintenance records and still needs appropriate access controls. “Draft” is not necessarily consequence-free. Equipment control and safety interlocks require their own engineering and authorization controls; a work-order interface should not be assumed to operate physical equipment.
Use a permission ladder
| Level | Agent capability | Suggested control |
|---|---|---|
| 0 | Read approved data and return a sourced summary. | Limit assets and data fields; log access as required. |
| 1 | Prepare a proposed work order outside the CMMS. | Planner edits and enters it manually. |
| 2 | Create a clearly identified draft record. | Permit only where draft side effects are understood; validate fields and route to a named reviewer. |
| 3 | Update selected low-impact fields. | Explicitly authorize fields and state transitions; require approval appropriate to the consequences. |
| 4 | Release, commit resources, or close an order. | Treat as high-impact; require authorized human approval and system-enforced controls. |
For an initial deployment, stopping at a proposal or a controlled draft is a reasonable starting point, depending on local risk. It is not a universal maturity requirement. A system may require approval even for draft creation, and human approval does not replace authorization, input validation, segregation of duties, or safety procedures.
Answer four questions before enabling a write
Who approves? Name the role and backup for each action. Make the reviewer see the asset, proposed change, priority, supporting evidence, and likely consequences. An approval must correspond to the specific action and data reviewed, not a generic permission given at the beginning of a conversation.
What gets recorded? Keep the requester, agent or service identity, target system and record, timestamp, proposed and actual field values, source references, approval decision, and transaction result. Record failures and retries as well as successful writes. Respect the site’s retention and access requirements for logs.
How is a mistake corrected? Identify the authorized correction, reversal or compensating workflow before launch. Some actions cannot simply be undone: a notification may already have been sent or a purchase committed. Preserve the original record and link the correction according to the CMMS audit rules; do not assume deleting an order reverses its effects.
What if the input is stale? A work order can be changed after the assistant reads it. At execution, recheck the target record’s version or other concurrency condition and the relevant evidence. If it changed, stop and ask for renewed review rather than overwriting the newer record. Confirm asset identity and whether a similar open order already exists.
Test a realistic failure path
Imagine a vibration alarm creates a proposal for inspection of pump P-17. The planner sees that the source alarm refers to P-17, but the pump was replaced and the asset mapping changed that morning. The proposal should return to review; matching the tag name alone is not enough to identify the current equipment. If a draft was already created, use the documented correction process and link the corrected action to the original.
Before production, test invalid asset IDs, duplicate requests, delayed approvals, expired source data, authorization failures, retry behavior, and a failed reversal in a safe environment. Check the actual system audit trail and downstream effects rather than relying on the assistant’s confirmation message.
A successful pilot lets the planner see why an action was suggested and decide whether to make the commitment. Expand permissions only after the relevant owners have reviewed observed errors and the system can enforce the next boundary.


