What This Looks Like
A tool, connector, function, MCP surface, or integration is called for one target, but its effect reaches another object, file, record, workspace, branch, state surface, or workflow area. The user may see the requested action happen, but another area changes too.
Why Users Blame AI
The effect appears after an AI or agent action, so the user may experience it as “the AI changed the wrong thing.” The underlying cause may be tool behavior, broad permissions, shared state, default targets, hidden side effects, connector configuration, or an integration whose effect boundary is wider than the prompt described.
What to Check First
- Whether the tool has a broader default target than the user intended
- Whether the agent had access to more files, records, or workspaces than needed
- Whether the tool call included the correct target identifier
- Whether the integration changes shared state as a side effect
- Whether read scope and write scope are separated
- Whether the workflow compares intended change surface to actual change surface
When This Is AI-Adjacent
Use this AI-Adjacent Issue when a tool effect crosses the intended target boundary. If the agent modified unrelated state, use the related Workbench Issue for agent-modified unrelated state. If the requested action changed something else too, use the related side-effect Issue.