Prompts Expire. Knowledge Compounds.

A year ago, everyone was told to learn prompt engineering. Then came custom GPTs. Then agents. Now business owners ask me whether they should bet on ChatGPT, Claude or Gemini, and whether that answer will flip again next month.

Probably, I tell them.

Waiting won’t help, though. The better question is where your time and money still pay off after the next model takes the lead.

What expires

When you invest in AI, you build up two kinds of knowledge.

The first is about the tools. Which model is ahead right now, which prompt trick works, where the button sits in which interface. That knowledge goes stale fast. A lot of what I learned about prompting eighteen months ago I barely use today, because the models have become much better at understanding what I mean.

The second is about your own company. How customers actually use your product. Why a big deal fell through three years ago. Which questions every new hire asks in their first weeks. That knowledge doesn’t expire with the next model release. If anything, every better model can do more with it.

In April I wrote about why AI is only as good as the knowledge behind it (The Technology Isn’t the Bottleneck. The Knowledge Is.). This time I want to talk about time. Tool knowledge has a half-life. Your own knowledge, if you keep it up, earns compound interest.

AI is an instrument

I recently heard a picture I keep coming back to. AI is less like a screwdriver and more like an instrument. Anyone can blow into a flute and get some kind of sound out of it. Making music takes practice.

A lot of owners treat AI like a screwdriver. Buy it, use it, done. If the result disappoints, the tool gets the blame, and the next one gets bought.

I’m no exception. In May 2025 I wrote: “AI isn’t the goal. It’s a tool.” (KI ist kein Schnellboot, in German) I still stand by the first sentence. I’m more careful with the second one now.

With an instrument, every bit of practice shows you where you’re stuck. And with AI, you’re almost always stuck in the same place: the answers are generic because the AI doesn’t know your business. If you play regularly, you see those gaps. If you blow into it once, you never will.

In June I asked whether a managing director’s duty might soon be less about using AI and more about seriously engaging with it (Vielleicht besteht die Pflicht nicht darin, KI einzusetzen, in German). Practice is what that engagement looks like.

The loop

Knowledge management has a bad reputation, and it earned it. Plenty of companies set up a wiki twenty years ago that nobody touched after six months. The order was wrong. Write everything down first, use it someday. The payoff came too late, so the upkeep stopped.

I’m not writing this from the sidelines either. In May 2025 I still told business owners to start a wiki as step three (Knowledge Is Your Superpower). That knowledge needs upkeep and an owner still holds. The order is what I’d change today.

AI flips that order. A handful of documents is enough, and the AI works with them on day one. Every question it answers badly points to a gap. Someone fills it, and from then on everyone who asks the same question benefits.

Build knowledge, use it, look at the questions to see what’s missing, fix it. Then go around again.

That’s the compounding. Each round makes the next one worth more. After a year you have a knowledge base that answers exactly the questions people in your company actually ask. Nobody can sell you that, and no competitor can copy it.

One thing matters here. The loop doesn’t run by itself. The AI shows you where the gaps are. A person has to close them. Same as with the flute: the instrument doesn’t get better, the player does.

The snowball

The objection I hear most: we don’t have time for this, and our documents are a mess anyway.

Neither should stop you. A snowball starts as a handful of snow. A few PDFs, a manual, the answers to the ten questions your sales team gets every week. Then you start rolling, and every turn picks up a bit more.

I’d start even smaller, with yourself. Before you roll anything out to the company, build the loop for your own work. I’ve been working with a so-called second brain for a while: plain text files with my meeting notes, projects and decisions that my AI tools can read. After a call, the transcript goes in, and a few minutes later I have a finished note with the context from earlier conversations already pulled in. I described the setup here: Every Company Wants AI. Almost None Are Ready for It..

Starting with yourself has two upsides. You learn the instrument before you hand it to anyone else. And you know from your own experience where the effort goes and where it pays off.

From one person to the company

We tried the loop at a larger scale with ScormIQ, in a pilot with a large corporation. There, the knowledge base was the existing e-learning. Employees could ask questions right inside the course. When questions pile up at one spot, the course is probably unclear there, and that’s where the next round starts.

If you want to do this for yourself and your team, that’s what our KI-Praxisprogramm is for (in German). Managing directors work in a small group on a real case from their own company instead of stockpiling tool knowledge.

Three questions before your next AI investment

1. Where does the knowledge live that your business can’t run without? In documents, inboxes, people’s heads?
2. Who makes sure it gets maintained? A name, not a department.
3. How will you notice that the answers are getting better?

In two years, nobody will care which model you picked today. The knowledge you start maintaining today will be your head start by then. You can’t buy that.

Start small. But start.