When Should You Give AI Write Access?

The question is not whether to give AI write access. It is which writes, under what conditions, with what recovery. Three properties decide it.

Question one: is it reversible?

Can you put things back exactly as they were, and how long do you have?

Fully reversible (a status change, an assignment, a task creation) is where AI write access should start. The worst case costs an undo.

Reversible with effort (a bulk edit across hundreds of records) needs the ability to undo a whole run at once, not record by record.

Irreversible (a sent message, a submitted filing, a payment) does not get delegated. Not at level 4, not at level 8. External effects cannot be recalled.

Question two: what is the blast radius?

One record, or ten thousand? An agent with a bug that touches one task is a nuisance; the same bug across an entire pipeline is an incident.

Blast radius controls: rate limits per hour, staged rollout for bulk operations over a threshold, and approval requirements that scale with record count rather than being fixed.

Question three: would you notice?

The most dangerous category is an action that is wrong in a way nobody detects. A subtly miscategorized expense sits in your books until an audit; a deleted task is noticed immediately.

If an error would be invisible for weeks, that action needs review even when it is technically reversible, because reversibility you never invoke is not protection.

The order to grant in

Signals you moved too fast

• ›You are correcting the same kind of mistake repeatedly.
• ›Someone outside the team noticed an AI action before you did.
• ›You cannot explain, on request, why a specific change happened.
• ›The audit review has become too large to actually read.

Any one of these is a reason to move that scope down a level. Moving down is a healthy feedback loop, not a failure.

What stays human

Sending anything to a customer. Moving money. Hiring and termination decisions. Legal commitments. Not because AI cannot draft them well, but because the cost of being wrong is asymmetric and unrecoverable.

FAQ

How long before granting the first write permission?

Two to four weeks of good proposals in suggest mode is a reasonable bar for a low-risk verb.

Should different agents have different write access?

Yes. A proven agent doing one narrow job should be more trusted than a new one, and per-agent overrides exist for exactly this.

What if we need speed more than caution?

Then choose reversible verbs and rely on undo rather than approval. That is the honest trade: speed with recovery, not speed without a net.

Brainis scopes write access per verb, scales approval by risk, and undoes actions and whole runs. See how it works.