A hundred emails, one page, by seven.

The owner of a small business is the bottleneck in their own company, and the bottleneck is the first ninety minutes of the day. Somewhere between eighty and a hundred and fifty messages arrive overnight. Perhaps six of them need a decision only they can make. Finding those six is the job that happens before the actual job.

The second problem is quieter and more expensive. Commitments get made out loud in meetings, agreeing to send a quote by Friday, agreeing to call a supplier back, and they live in somebody’s notes rather than anywhere that will ever surface them again. Nobody drops them on purpose. They just never became a thing that could be dropped.

The usual answer is a productivity system, which fails for the same reason every time: it adds a fifth place to look for people who already cannot keep up with four.

Everything needed to work out what matters today already exists. The mail is in Google Workspace. The calendar is beside it. There is a task list of some description, and there are notes or recordings from yesterday’s meetings.

None of the four talks to the others, so the cross-referencing happens in one person’s head, every morning, at the point in the day when their judgment is worth the most.

There is also a trust problem waiting at the end of any solution to this. An assistant that starts by archiving things is an assistant that gets switched off in week two, right after the first time it archives the wrong thing.

The design constraint that shapes everything else: build a read-only layer that produces exactly one artifact a day. No new inbox, no new app, nothing to migrate into, and on day one nothing that sends, deletes or files without being asked.

The second constraint is that every ranking has to explain itself. A priority list that says an item is important without saying why gets skimmed by the second week and ignored by the fourth. The explanation is not a nice touch, it is the mechanism by which the thing stays in use.

The third is cost discipline. Sending every message to a model is both slow and needlessly expensive, and it is the reason most builds like this quietly get turned off when somebody looks at the bill.

At half past four each morning a scheduled job pulls the last twenty-four hours of mail, today’s and tomorrow’s calendar, the open task list and the notes from yesterday’s meetings. Each item is classified, ranked and assembled into a single page, delivered as an email and as a web page that works on a phone at the kitchen table.

Five decisions that separate a real assistant from a demo

1. Classification runs in two passes. A deterministic pass first: known senders, newsletters, receipts, calendar invites and automated notifications are sorted by rule with no model involved at all. Only what survives goes to the model, which is roughly fifteen percent of the volume and is what keeps the running cost in cents a day rather than dollars.
2. Every ranked item carries its reason, in one line, in plain language. “Top of the list because you committed to an answer by Friday in Tuesday’s call, and it is Thursday.” The reason is generated from the evidence the system actually used, so if the reason is wrong the owner can see immediately why it was wrong.
3. Commitments are extracted from yesterday’s notes and matched against the existing task list. Anything promised out loud that is not already captured becomes a proposed task, approved in one tap. It never becomes an actual task on its own, because a task list that fills itself with things you did not agree to is worse than no task list.
4. The brief is generated even when a source fails. If the calendar API is unavailable the page still builds, still ranks the mail, and says plainly that the calendar is missing. A brief that sometimes does not arrive teaches people to stop looking for it, and after that the system is dead no matter how good it is on the days it works.
5. Nothing is auto-archived for the first ninety days. The system proposes, the owner disposes, and every override is logged. Those ninety days of overrides are the evidence for what it can eventually be trusted to handle without asking, which is a decision made from data rather than optimism.

The end state is not an autonomous agent running the business. It is a very well-prepared briefing, on time, every day, that a person then acts on. That is a lower ceiling than the demos promise and a far higher floor than what most owners have now.