Why an AI teammate should ask before it ships

Quiet mistakes

A script that breaks shows an error. An agent that misreads a situation produces something that looks right: a reasonable message, a tidy PR, a confident summary. That is why the question of who approves what matters more for agents than for automation.

The NIST AI Risk Management Framework makes the same point in general terms: people should be able to oversee, understand and correct what an AI system does.

Three levels of autonomy

Suggest: the teammate says what it would do and why. Draft: it prepares the work, such as a message, an issue or a PR, and waits. Do it: it acts within limits you set and tells you afterwards.

Teams rarely want one level for everything. Posting a weekly brief can be “do it” from day one. Sending invites to users may stay at “draft” until you trust the wording. Code changes always wait for review.

A log, a reason and an undo

Every action should leave a line: what was done, when, why and on whose approval. If someone asks why Wirey emailed a user, the answer should take seconds.

Undo matters as much as approval: a draft that can be discarded, a message that can be recalled, a PR that can be closed. If an action cannot be undone, it should never run without a yes.

Limits, not just permissions

Permissions say what an agent may touch. Limits say how much: credits per month, invites per week, PRs open at once. Both belong on the same settings screen, owned by a human.

With levels, logs and limits in place, a team can hand the teammate more work without giving up control. That is the trade Wirey is built around: it does the work, you approve every change.