Episode 11: The AI will figure it out

“We just tell the AI what we want, and it’ll figure it out.”

I’ve heard that sentence often. And I’ll admit it: at one point, I told myself the same thing.

It’s an easy illusion to fall for. AI really can do a great deal. Its technical skills are out of the ordinary, and highly specialised. It writes a connector in a few minutes. It builds a complete workflow while you’re drinking your coffee.

There’s a first guardrail, though. You can’t let the AI write a connector or a workflow and then tell yourself, “There, it’s done everything, it all works.” The code it writes has to be checked: security holes can hide in it, and so can inconsistencies. And the whole thing has to stay consistent with the system’s architecture.

The second point goes further. The strength of our agents is that they connect all the systems we work with, as well as the external systems we pull information from, and sometimes write to. Keeping that architecture consistent takes a full understanding of what is possible. It also means being able to sign off every step the AI puts to you, fully aware of what you are approving.

A real-life example. I want to connect to the OpenAI API to access a model that generates diagrams. The agent guides me: send it the content, supply an API key for the connection. Step one is to understand what it is telling me. Step two, to check that this is the right method. Step three, to generate the key in the right place and hand it over to the agent. And then, to test. The agent tests the flows itself, of course. But we have to be able to steer it, and to ask it specifically to cover the edge cases, those tricky situations where systems misbehave.

In the end, the agent does back us up. It produces very good code, very quickly. But that takes supervision, and a perfect grasp of the overall architecture.

And this is where I have to say how pleased I am. This way of working makes the most of the expertise of the developers I work with, and shifts their role towards higher-value tasks. It’s exciting for them, and it’s exciting for me. It lets us turn out very sophisticated integrated systems and workflow automations, sometimes very complex ones, and do it faultlessly. Their behaviour has been thoroughly tested, and their architecture designed as well as it can be.

So yes, you tell the AI what you want, and it figures it out. As long as someone beside it knows exactly what it is doing.

Laurent Cuénoud, founder of Djtal, August 2026