AI at SANE/REBELS · 04
AI in Ops
The operating system behind the day: deterministic code for the safe part, AI for the judgment.
On this page
In our own house
What a BOS is
A business is made of hundreds of small decisions and chores a day. Most of them have a right answer. A Business Operating System is the layer of small programs that takes those on, so a person's hours go to the few that actually need judgment.
Not a chatbot
An operating system
Not an app you type into and hope. A handful of routines that frame the day and run in the background, whether anyone is watching or not.
Deterministic first
Code wherever it can
Anything with a right answer runs as tested code, the same every time. A model is added only where there's no clear right answer.
One source
Shared context
Every routine reads the same context, the Brand Hub. What's true in one place is true everywhere, because there's only one place.
Deterministic or judgment
The same dividing line we use everywhere, applied here to our own day. On the left, the chores that always run the same way. On the right, the calls that take experience.
Code does
The chores
Assemble the morning brief, move a to-do's state, file a note, post a digest, run a check. Repeatable, logged, reversible.
A model does
The judgment
Triage what matters today, draft a reply, summarize a call, propose an order. Always a proposal, never the final word.
What the model produces is a proposal or a summary. It sends no email, moves no money, changes no setting on its own. A human decides, every action is logged, and it can be undone.
The bookends of the day
Good morning, bye bye
Two routines bracket the working day. One starts it with a clear picture, the other closes it cleanly. Both gather with code and let a model only read and suggest.
Good morning
gather & surfaceCode gathers the state. Calendar, open to-dos, what came in overnight, the pipeline. From fixed sources, not from memory.
A model surfaces. It reads the state and says what needs a decision today. It decides nothing; it surfaces.
Bye bye
capture & tidyCapture what happened. The day gets summarized, and loose ends move back into the ledger as to-dos.
Leave the desk clean. Nothing lives in someone's head alone. Whatever's open has a place where it reappears tomorrow.
The To-do
One ledger, one place. The dividing line is clearest here: code keeps the ledger clean, the model helps sort, the human decides the order.
Code keeps it clean
One item, one truth
One place. No item lives in two systems.
An owner, a state. Every to-do knows whose it is and where it stands.
Idempotent. Reporting the same thing twice creates no duplicate.
The model proposes
Sorting, not deciding
What's urgent and what's gone stale
A suggested order for the day
Catching duplicates before they become two items
A human confirms the order in the end.
Shared context
The Brand Hub
The same context that grounds a built product and an ad account grounds operations too. One source every routine reads, reviewed and versioned.
Voice and banned words. Every draft, email or ad, sounds like us, not like generic AI.
Economics and goals. Margins, targets, and what counts as done, in one place.
Change once, effective everywhere. Change the hub and what every routine proposes changes with it.
One source for everything. This is the same principle as on the overview: an AI is only as good as its context, so it exists exactly once.
How we build this
Built so the safe thing happens by default, even on a busy day.
Architecture
Deterministic core, AI as the shell
The AI reads and proposes. It triggers no action on its own. What actually happens is tested code.
Governance
Every output is a proposal or an explanation
What the AI produces is display or a proposal, and needs an approval before it has any outward effect.
Engineering
Agentic engineering, not vibe coding
Our own developers review the routines. Most of the work is hardening them against the edge cases.
See it on a real example?
30 minutes, the whole loop live. You see what gets proposed, what a human approves, and what lands in the end.
Request a callRelated content
Last updated June 2026
Citation rule: if you use or build on these ideas, credit SANE/REBELS (KNUS GmbH) and link to sanerebels.com/about/ai.