Built AI-first · Chapter 8 of 9
"Should the AI do it automatically?" is the wrong shape of question, because it asks about an agent when the answer differs per act. The same role can reasonably be trusted to move a calendar entry unattended and not trusted to send a one-line email to a client.
Five rungs, granted one act at a time. Where to start, what earns a promotion, and why the order is structural rather than cautious.
The spectrum
Read the rungs as grants of authority over a specific act — "draft and send the weekly status note", not "the delivery agent".
The rules
Rule one
An agent is not "trusted". A specific act by a specific agent, with a specific scope, is at a specific rung. Ours sit at different rungs for different work, and the grant is written down beside the routine rather than held as a general impression of how well it has been doing.
Rule two
Two questions, both required. How expensive is the undo — minutes, an apology, a client, a licence? And how long is the verified record for this exact act? A long record on a similar act does not transfer, which is the mistake that produces the confident surprise.
Rule three
You cannot audit trust you never measured. If nothing logged the outcomes at the lower rung, there is no case for the higher one — which is why the verdict log from chapter five is a precondition for this chapter rather than a nicety.
Rule four
An envelope breach, a failed check twice running, or a changed input drops the act a rung by rule. If demotion requires somebody to convene a meeting, it will not happen at the moment it is most needed.
Autonomy is not a reward for good behaviour. It is a bet on reversibility, priced with a record.
The sequence
The argument against skipping rungs is usually made on safety grounds. The stronger argument is political, and it is the one that actually decides whether an AI programme survives its first year.
An organisation gives a new capability roughly one serious mistake. Make it at rung one, and it is a bad draft that somebody caught — the system worked. Make the same mistake at rung four and it is an invoice sent to the wrong client, a records response released early, a message that went out under the organisation's name. The cost is not the error; it is that every subsequent proposal now arrives into a room that has seen this go wrong.
The sequence also produces the evidence the later rungs need. Six months at draft is not six months of delay — it is six months of exactly the record required to make the case for acting unattended, assuming you logged it.
In your world
Assembly and noticing are good early rungs — the certification about to expire, the grant window opening, the export that did not arrive. Anything that becomes a public record or leaves the agency stays at approval permanently.
Packet assembly and first-pass drafting are genuine relief. Release of anything disclosable is a decision with legal consequence and an owner with a name, and no record length changes that.
Quote assembly, follow-up drafting and job-cost noticing move up quickly because they are cheap to reverse. The number itself is the business and it stays with the owner.
The value is in the artefact being ready, not in it being sent automatically. Internal assembly and monitoring can climb; anything a client reads keeps a person on the last step.
What we will not claim
Where this sits
Eight chapters of structure become useful when they are put in an order with dates and gates. That is the last one, and it is the one to read if you only read two.