Business automation

How I understand business automation.

I did not automate my company because it was fashionable. I did it because I had to cut costs.

What I learned: AI is volatile, and much harder to put to good use than it looks in the first minute. A dashboard takes an afternoon. Taking work off people's shoulders, reliably and without supervision, takes a lot of scaffolding.

Code is land. AI is open water.

An island is code: the same input takes the same road every time, and if the road is wrong, it is visibly wrong.

A crossing is one AI call. It reaches places no road goes, but it can land somewhere other than where it was aimed.

A harbour is a backtestable point: the result is written down with its evidence, so it can be replayed against cases whose outcome is already known.

One order, end to end

CodeOrder arrives, PDF parsed

AI reads the order

Recorded in units

CodeStock, shelf life, fallback

Stock status, replayable

AI proposes the strategy

Strategy written down

CodeWarehouse, invoice, close

Island: code, fixed road Crossing: one AI call Harbour: a backtestable point

So we never sail far. Every crossing starts and ends in a harbour, and a new rule only steers after it has been replayed over the voyages already made. That is how I build reliability fast.

The rules that follow

Six rules I learned the hard way.

01One desk owns an item at a time. Everyone else adds facts, nobody else moves it.
02Hand over ownership, not a conversation. The handover is written with the change of state.
03Never two AI crossings in a row without a harbour in between.
04Every harbour carries its evidence: input, rule version, output, who acted.
05A new rule runs pro forma first, replayed over history, and is released only once it matches.
06Gate where an error gets expensive and hard to undo. Log everywhere else.

Desks, not agents. Mostly code.

At ADONiS every area of the business is a desk with its own dashboard and its own agent. Underneath, the work is overwhelmingly code.

266 scheduled processes

Order intake · invoice on arrival · bank match · marketplace restock · newsletter draft …

209 pure code

57 call a model

EmailOrdersCustomer serviceProblemsAmazonWebshopNewslettersReceivablesPayablesSupply chainLogisticsAccountingForecastKey accounts

Where AI shines

The machine builds itself.

AI is excellent at writing code. A request becomes a written task, one AI builds it, a second AI judges it, and it goes live on the next refresh. Reviewed, tested and logged.

AskA request, in words

OrchestratorQueues and numbers it

BuilderAI writes the code

JudgeA second AI checks

MergeInto the main code

DeployBuild and restart

LiveOn the next refresh

Visibility

No dark corners.

Dashboards do not calculate when you open them. The numbers run around the clock in the background, down to the margin of every shipment. At ADONiS that is more than 170 live pages, from cash to stock cover to every open order.

Cash

Receivables

Stock cover

Open orders

Illustration, no real data

Safety is designed in, not added on.

Money is prepared, never moved

Payments are drafted in your bank. A person releases them.

Exact approval

A yes covers the exact text or amount. Change one word and it asks again.

Full audit log

Every action and every message sent is recorded.

Switches per area

Any part can be paused on its own, and AI spend has a hard monthly cap.

Confidential stays confidential

Bank, cash and HR data visible to named people only.

Monitored around the clock

Every connection health checked every 15 minutes, with offsite backups.

Lines of code

Integrations

170+

Live pages

Book a 30 minute call