Does this picture look familiar? The CRM knows the customers, the ERP knows the orders, the decisions are buried in the chat archive, and a long-serving colleague knows where the rest is. Each of these solutions made sense on its own. Together they form a system that nobody designed.
Last week was about the unwritten code behind the org chart. Now AI tools are joining in. And many secretly hope that AI will sort out the mess. It does the opposite.
🐛 Amplifying
An AI agent works with whatever it finds. In seconds it combines what is spread across five systems, including the version that has been outdated for years. A messy filing system turns into a hundred answers that all sound plausible.
That is worse than no answer. An empty search result at least shows you that something is missing. A fluently worded paragraph does not.
The Swiss newspaper NZZ puts it plainly: the threat here and now is not a malicious, uncontrollable superintelligence, but human automation bias. (Barnaby Skinner, NZZ, 19 September 2026, translated)
🐛 Cementing
The second effect is quieter and more expensive. Every prompt, every workflow and every automation builds on today's silos. With every month, the mess becomes the foundation on which everything new stands.
Anyone who wants to clean up later no longer just tidies the old filing system. They also have to tear down everything that was built on top of it in the meantime.
In software development this is known as technical debt. AI acts on it like a loan: it delivers speed right away, and you pay the interest with every agent you put on top of the mess. The unnecessary tokens too, by the way.
Why "Clean it up!" doesn't work
Anyone who calls for a big clean-up in this situation gets polite agreement and then nothing. That has little to do with unwillingness. Psychology calls it reactance: people who sense that a choice is being taken away from them defend the status quo. The call to clean up quietly declares the decisions of the past ten years to be mistakes, and the people who made them are sitting at the table.
I am not writing this as someone who always got it right. Order did not come from foresight in our own company either. Some of the rules our own AI agent works by today exist only because exactly what they now prevent went wrong before.
The reverse order works better. First acknowledge that every purchase made sense. Then make visible what the current state costs as soon as an agent works on it. The people involved will draw the conclusion themselves.
🦋 Now is the cheapest moment
Before the first agent, order and structure cost a week. After the tenth, they cost a project. That is why it pays to start small now:
🦋 Choose one area where the first agent is meant to work, and create order only there: one valid version per document, clear filing, one responsible person.
🦋 Make the status visible: is this a draft, has it been reviewed, has it been approved? What the agent cannot tell apart, it treats the same.
🦋 Subtract before you add: for every new tool, retire an old one or a duplicate filing system.
🦋 Document workflows before they become dependencies: what does the agent access, and what happens when the source changes?
One price list instead of three
What this could look like at a construction supplier
A construction supplier with 120 employees wants to use an agent for quote requests. Before starting, the team takes two weeks for a single area: the price lists. Three versions become one. The old lists are kept but marked as archived. One person from sales is responsible for the valid list. Only then does the agent get access.
The quotes are correct. And when prices change in spring, a single file is updated instead of three workflows that would each have been built on a different list. That is the difference between a week and a project.
Fewer copies, more knowledge
What order brings in everyday work
🦋 Sales, production and service work from the same basis instead of copies.
🦋 Knowledge stays in the company, even when people leave.
🦋 Agents work by your rules instead of their own assumptions.
Once you know the interim goal, you can see what needs tidying first. Small steps beat no steps.
But every month you wait is another month of concrete.
👉 Which area would you sort out first, before the first agent arrives? We are happy to look at it together.
Part 3 of the series «The Company Code»: seven theses on how companies make AI effective without automating their chaos.
(AI modified)