Two weeks ago, an Intex swimming pool appeared in my garden. Eight thousand litres of water, and I was done — or so I thought. What I did not know: the moment you add the water, you enter a world much bigger than the pool itself. pH values. Chlorine. Alkalinity. Cyanuric acid. Words whose existence I had not suspected a month earlier.
A trip to the DIY shop made two things clear. First: there are more bottles, powders and testers than anyone can reasonably keep track of. Second: I was certainly not the only person struggling to keep the water clear and healthy. Search online and you find an endless stream of people with cloudy water, green edges and contradictory advice.
My conclusion came quickly: this is a perfect job for an AI agent.
And that is exactly what I built. Not with code, a team or weeks of work. I gave a handful of instructions, entered the maintenance products I had bought and the dimensions of my pool. Nothing more. Since then, I measure the values every other day with a simple tester, enter them and receive a neat analysis: what is happening, what I should do and, especially, what I should not.
The result is an excellent pH value and crystal-clear water. But that is not the interesting part.
The interesting part is what I actually have. I used AI to build something — Claude acted as co-designer, sounding board and engineer at once. At the same time, the thing I built is itself AI: an agent that reasons, advises and supports decisions. In my garden, the two blend seamlessly. I barely notice the distinction any more.
And that is where it starts to grate, especially from the perspective of my profession.
Because in life sciences, where I work daily with quality systems and validation, these are fundamentally different things. AI in the building process — used to develop, document and analyse faster — does not itself become part of the final product. AI in the product — the delivered thing contains AI, advises and helps decide — changes the rules completely. Who is responsible when the agent gets it wrong? How do you demonstrate that it does what it should? What does “validated” still mean when the system itself reasons?
With my pool, that distinction feels unimportant. If the agent makes a mistake, the water gets cloudy for a while and I correct it manually. In this example, I treat the stakes as negligible.
But scale the example to an environment where the stakes are not negligible — a production process, a quality decision, a patient — and the same blurring becomes far from innocent. Then “am I using AI to build, or is AI inside what I deliver?” is no longer semantics. It is the core question.
My swimming pool gave me that question in its most innocent form. In a future article, I want to ask it in its least innocent form.
AI in the building process, or AI in the product? More on that soon.