Shadow AI Is not only a risk. It's might be your best pilot programme.

Something interesting has been happening at our clients, and since long at Redfield too. People finally got access to Claude, Copilot; sometimes officially, sometimes through a private account because waiting for IT wasn't working and the enthusiasm is hard to miss. They come to us with results. Real ones. Stuff that would normally mean a ticket to IT, or a consultant, or waiting two weeks for someone who's "swamped right now."

I get why security teams are nervous and it is a real problem to be solved. But watching it up close, it doesn't feel like a risk to me so much as a pilot programme that ran itself. Nobody had to design it. Nobody had to sell it. The domain experts, the ones that actually knows anything within their area, build something useful for them selfs. Without having to wait for the IT-department to get ready for it. There is one thing that concerns at me though. When someone shows me what they built, I sometimes wonder: do they actually understand the code the AI wrote for them? Could they read it, reason about it, hand it to a colleague? And if the answer is no, how can you know it is actually working and what happens the first time it breaks? That's really where this post starts.

The tool I keep reaching for and where it stops

We've used KNIME at Redfield for years. I still think visual tools like it are underrated; for a lot of business users and domain experts, the visual paradigm is the right mental model. You can see your flow. You can hand it to someone. You don't have to explain what a virtual environment is. It is great for sharing your results with your team. But we're also running into the ceiling more and more often. At some point you need code for the flexibility, for the libraries, for version control, for things visual tools quietly struggle with once the workflow gets big. And for a while I've been looking, mostly without luck, for code's answer to KNIME. Something that lets business people and domain experts actually build and maintain their own applications, not just prototype them. AI closes part of the gap. It writes the Python. That's the easy part now. The harder part is whether the person who asked for the Python can still see what they have.

Stumbling into Kedro

I found Kedro more or less by accident, half-scrolling one evening. Open-source, Python, and the second I looked at it the KNIME parallels jumped out: reusable nodes, a visualisable pipeline, configuration kept separate from logic. The same mental model I like about visual tools, but the artefact underneath is proper code, versionable, testable, reviewable. That was the moment it clicked. Not "replace the visual tools." Complement them. Give the people who've outgrown drag-and-drop a place to land that doesn't require becoming a software engineer first. And give the AI-generated code somewhere structured to live, instead of drifting around as loose scripts nobody quite trusts. We have to try this, I thought. And probably offer it to our clients too.

What we're doing next

Right now we're experimenting with Kedro alongside AI-assisted coding especially with people who are Python-curious but not Python-fluent, and who've been leaning on Shadow AI to get things done. The goal isn't to take the AI away. It's to make what comes out of it something the person who asked for it can actually own. We'll share what we find.