| “Return a fraud decision inside the request’s tail-latency budget.” | A conventional service with a deterministic model and an explicit SLO. | Crewship runs are asynchronous; model, container, and approval latency are variable. |
| “Reformat every row in this ten-million-row table exactly the same way.” | A SQL/data pipeline or a one-off program with a checkpointed batch. | An agent adds token cost and variance where there is no judgment to supply. |
| “Push this production migration automatically and roll it back if health drops.” | A deployment system with tested rollout and rollback semantics. | Crewship can prepare a change and pause at a waitpoint, but it is not your production controller. |
| “Review this repository, but the crew cannot read the repository or its tests.” | Grant the workspace or attach the repository first, then reassess. | A better prompt cannot create missing files, permissions, acceptance criteria, or domain context. |
| “Give the agent our regulated customer export because it might help.” | Keep the data in the approved system and use a purpose-built, audited integration. | Credential injection governs delivery; it does not make an unapproved data boundary acceptable. |
| “Keep an autonomous process alive forever and let it decide what to do next.” | An event-driven service or scheduler with explicit state transitions and owners. | Crewship is queue-and-trigger based; routines finish, fail, or pause at durable waitpoints. |
| “Have the agent change a security chokepoint without a test or reviewer.” | A controlled engineering change with red-test-first review and protected deployment. | Crewship can coordinate the work, but an agent run is not evidence that a security-sensitive change is correct. |
