Automation and AI

Every step knows how to undo itself

An automated work cycle is a distributed transaction: every step has its own way of being undone, a retry ceiling and saved state. If it falls over halfway, the system knows where it was and what to undo — it doesn't start again hoping nobody notices.

Every step is written to be repeatable: sooner or later it will actually be repeated.

A real queue, not a reminder

Jobs live in a queue with exclusive claims, so two workers never do the same thing twice. The queue is drained on a schedule and right after an insert, so nothing waits for the next tick without reason.

AI where it adds, people where it counts

Language models are useful for understanding and proposing: draft replies, classification, extracting data from documents and invoices. But while an action is public or irreversible, the proposal goes through an approval — in the app or as a two-button message — and stays on the record.

The value is not "AI does everything": it is removing the mechanical preparation and leaving the decision to a person, with the context already laid out.

What we automate first

Not what annoys people most: what costs most. They are two different things, and the most hated task is almost never the dearest. You measure first — minutes per occurrence, occurrences per day, people involved, errors in the last three months — and start from the top line.

A good candidate has three properties: clear rules, high volume, low ambiguity. If one is missing, the automation costs more than it returns, and it is better said before starting.

Where AI helps and where it doesn't

It helps where understanding is needed: reading an invoice in a format never seen, classifying a free-text request, summarising a long case, drafting a reply. Tasks where an occasional error is tolerable because someone checks afterwards.

It doesn't help where guaranteed exactness is required: calculating a total, deciding a discount, applying a contractual rule. Those stay written, deterministic, verifiable rules — and the fact that today you can ask a model anything doesn't mean you should.

The practical line we use is irreversibility: if the action sends something outside the company or deletes, it needs a human yes, and it stays on record who gave it and when.

What must be decided before, not after

Which data may go to an external service and which may not: customer records, health data, documents under confidentiality. A decision to take once, write down and make known — not one to leave to the habits of whoever uses the tool.

Who answers for what goes out: accountability stays with whoever signs, always. No supplier takes it, and anyone claiming otherwise hasn't read the contract they are handing you.

How it is measured: hours saved, errors avoided, response times. If after three months you can't say whether it helped, you don't have a project, you have a subscription.