When not to automate

APD/M is aggressive about automation and strict about where it stops.

Always human

• Production deploy. Ship gate stays with a person who will answer when it breaks.

• Live user research. No agent between you and the human whose problem you claim to solve.

• Public voice as the brand. Drafts can be agentic. Publish is approve/deny.

• Ethics and privacy calls. The model said it was fine is not a policy.

• Irreversible money and authority. Anything that moves funds, grants access, or bulk-rewrites the board of record needs a human or dual control you designed on purpose.

• Prioritization when the map is wrong. Fix ground truth before you automate harder.

Automate hard here

• Status assembly from real systems

• Health collection and anomaly flags

• First-pass research and competitive scans

• Boilerplate implementation under review

• PR prep, test scaffolding, checklist generation

• Draft documentation that tracks decisions as they happen

The rule of thumb

If the cost of being wrong is a messy PR, automate more. If the cost of being wrong is customers, money, reputation, or a board you cannot reconstruct, keep a human on the gate.

Restraint is part of the method

The goal is not maximum automation. The goal is maximum trustworthy throughput. Untrusted throughput is just waste with a GPU bill.

Build agents that earn scope. Do not gift scope to impress a demo. That is how APD/M stays production-grade.