Notes · AI systems · 7 August 2026
Getting a model to write a decent client update is now the easy part of the job. The hard part is the two minutes a human spends approving it, and the thing that governs those two minutes is the clock.
A model can write a market summary, a weekly status note or a client update from structured data, and it will usually be good. That part is largely solved.
If the thing goes out under your name, somebody has to read it first. Everything difficult about these systems lives in that step.
Say the note goes at 7:00am. The reviewer has to open it, disagree with a sentence, fix it and approve it before then, and the whole promise of the system is that this costs less time than writing the note would have.
That one deadline sets every other one. The data has to have landed, the transforms have to have run, and the draft has to be sitting there before the reviewer sits down with a coffee.
It also changes what better means. A sharper draft that arrives at 6:55 is worse than a rougher one that arrived at 6:30, because the late one eats the time the system was supposed to give back. Once the deadline is fixed, most of the architecture follows from it: what runs on a schedule, what is computed in advance, and what a late source is allowed to do to the send.
Generation and sending are separate stages with a state boundary between them. The model writes to a pending table. The sender reads only rows marked approved. There is no route from the model to the client that skips a person, because the code that sends cannot see an unapproved row.
We prefer that to a policy saying a human reviews everything. Policies get skipped on a bad morning. A sender that can only read approved rows has nothing to skip.
The audit trail comes with it: what was drafted, what was changed, who approved it, when.
The easy metric is volume — drafts produced, notes sent. We have watched that metric look healthy while a system was making work rather than saving it, because a draft the reviewer rewrites from scratch still counts as a draft.
What matters is how much time the reviewer got back, and that cannot be known until real reviewers have used it on real mornings. Before that it is a target, and we say so. The distance between “the drafts are usually fine” and “reviewing is faster than writing” is where these projects fail to pay for themselves.
Once the drafts are consistently good there is pressure to remove the reviewer. Sometimes that is the right call.
The thing to weigh is that the reviewer is also the smoke alarm. They are how anyone finds out that a source has changed shape, or that the market did something the prompt never contemplated. Take out the gate and the detector goes with it. Build the replacement first.
More notes