StudioConsultant shares the guardrails an AI agent used unprompted on a client's ERP
Original title: I gave an AI agent write access to a client's ERP. Here are the guardrails it used without being told.
A consultant gave his own AI agent write access to a construction client's Odoo ERP via its chat account, and found it had developed multiple safety practices unprompted. It asked the project manager before archiving 27 expired bid projects, refused to guess on a merge and marked it blocked, confirmed 148 projects unused before archiving 7 workflow stages, saved record IDs for reversal, reverted a side effect that auto-attached sales order lines on 3 projects, left a draft invoice at $0.00 with "rate still needed" when no hourly rate existed, and logged a server outage without advancing its message position. The author summarizes the pattern as plan, approve, act, and record the undo, and asks what guardrails others require for agents writing to production systems.
A firsthand ops log shows the safety practices an AI agent adopted on its own when writing to a production ERP, useful reference for permission design.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.