SHEET B04 — BUILD

Business process automation and AI agents

The automation worth building is not the one that saves ten minutes. It is the one that notices, at three in the morning, that you have been spending for three days and getting nothing.

SPEC

We build two things on this sheet. Automations remove a step a person repeats: a document that generates itself, a record that updates itself, a report that arrives without being asked for. Watchers are different — a process that lives inside something already running 24/7, checks the state of the business on an interval, and messages you directly when something is wrong.

The rule that makes a watcher safe enough to trust is asymmetric autonomy, and it is enforced by the shape of the code rather than by an instruction. A watcher may pause, stop and protect on its own. It may never spend, never create, and never send a message to a real customer. That asymmetry is the whole design.

We do not build a watcher for a problem nobody has felt yet. Both of the ones running today came out of an incident: one business spent three days on ads with zero leads before anyone noticed; another had a payment bug and a redesign that passed an automated QA while being wrong.

SCOPE

What it watches

The specific thing that hurt last time — spend, silence, a queue, a number that should never move.

Cadence

An interval that matches how fast the thing actually changes, not the fastest one possible.

Alerting

A direct channel to one person. One bot per business, never a group.

Permitted actions

Pause, stop, protect. Written into the code as what it *can* do, not as what it *should*.

Forbidden by design

Spending, creating, and messaging your customers. Not configurable.

State

Kept in the database, never on disk — a redeploy must not make it forget what it already did.

Audit

Every action it takes is recorded, and every one is reversible by you.

FIELD CASE

Two watchers, two businesses, both in production

One checks ad spend and leads every four hours: it can pause a starving ad set and move daily budget inside a monthly ceiling the owner sets, and it cannot step outside that cage. The other reads live lead conversations every thirty minutes and pauses ads with no return. Both report to their owner on a private Telegram channel; neither can create a campaign or message a customer.

4h / 30mincheck intervals in production

0actions that spend outside the ceiling

1:1alert channel, never a group

FAQ

Is this n8n or Zapier?

Those are good at joining two services that already have buttons for each other. We write the ones that need judgement, that must not break silently, and that have to keep state across a redeploy.

Would it act without asking us?

Only in the direction of stopping something, and only inside limits you set. It cannot spend, create or contact anyone. That is a property of the code, not a setting.

Where does it run?

Inside a service that is already running for your business around the clock. If there is not one yet, building that comes first — an automation that depends on somebody opening a laptop does not watch anything.

What if it makes a mistake?

Every action is logged and reversible, and the alert reaches you before you would have noticed on your own. The failure mode of a watcher is a false alarm; the failure mode of not having one is three days of spend.

Can it use AI to judge?

Where judgement is actually needed, yes — reading whether a conversation went wrong, for instance. The rest is deterministic code, because most of what a watcher checks does not need a model and is more reliable without one.

Start a project

Have something worth building well?

Tell us what you are working on. We reply in English or Spanish, usually within a day.

hello@axisworks.studio ALL SHEETS

EMAILhello@axisworks.studio

BASEDSpain ✕ Bali

LANGUAGESEnglish / Español