Everyone is capable

In February 2025 Andrej Karpathy gave a new way of working a name: vibe coding. You describe what you want, an AI model writes the code, and you hardly read it. Since then a marketing manager builds a lead tool over a weekend, a sales rep scrapes competitor prices into a spreadsheet, and a controller automates the monthly report without a ticket to IT. Two years ago this work needed a developer or an agency. Today it takes an afternoon and a subscription.

The opportunity

The people closest to a problem can now build the solution themselves. They no longer write a requirements document, wait for a slot on the IT roadmap and explain the problem a second time. A first version exists before the meeting about it. For a small team the time from an idea to a working prototype has dropped from weeks to hours.

The other side

A tool built in an afternoon can also move company data out in an afternoon. Customer lists go into a chat window, source code goes into a prompt, and a small app stores client data in a database nobody at the company knows about. In 2023 engineers at Samsung pasted source code and meeting notes into ChatGPT, and Samsung then banned generative AI tools for its staff. In the Cost of a Data Breach Report 2025 by IBM, one in five organizations had a breach caused by AI tools that nobody had approved, and such a breach cost USD 670,000 more than the average.

The data goes first. What an employee types into a public model is outside the company’s control, and in finance or insurance that alone can break a regulation.

The knowledge goes next. A pricing logic, a scoring model or the way a team checks a property valuation belongs to the company. Once it lives in a prompt on a private account, the company can neither audit it nor take it back.

Last, the workflow becomes invisible. The self-built tool works until its builder changes jobs. After that nobody knows what it does, which data it reads or who has access to it.

Domain knowledge and experience decide

An AI model delivers whatever you ask for, even when the request is wrong. It does not know your business. It does not know that a price without a date is useless, or that customers leave when the checkout takes too long. The person who writes the prompt has to know that. This knowledge comes from experience: from years in an industry and from mistakes you have already made once. People with that experience can use AI to build solutions that fit their company.

About the author

For more than twenty years I have worked on the questions before the prompt: which data a business needs, who may use it and how long it stays true.

Today I work a lot with AI agents, virtual organizations and workflows. I set up AI agents for clients and write the rules they work under. I have worked with data from banks, life insurers, online shops and real estate. A company that works with me gets industry knowledge and deep IT knowledge in one person.

In my research I work on data aggregation, workflows and virtual organizations. All three deal with the same problem: how data from many sources stays correct when it evolves or it gets changed.

If you want to work together, get in touch.