Nobody Is in Charge of Thinking

A few months ago I sat down to build an automated trading system. I didn’t start by asking myself what problem I wanted to solve. I started by telling myself something much more fun: this could be an agent.

I asked a model to specify the architecture. Within minutes I had an impeccable document. Modules, responsibilities, data flows, error handling, all in order. It looked professional. It looked thought through. I built on that foundation for weeks, until the system became so rigid and so dependent on its own rules that it stopped being useful for the only thing that mattered. I had to take it apart completely and put back what I had skipped: clear roles, explicit data contracts and human judgment at the points where the system couldn’t decide on its own.

The mistake wasn’t in the code. It was in the first minute, when I went from “I have a problem” to “this is an agent” without stopping for the boring step of defining what I was trying to solve.

The question that changed in meetings

That sequence interests me because it isn’t mine. I see it in almost every conversation I have with teams starting out with AI. The question that opens the meeting is no longer what the problem is, but what we could automate. And the difference between those two questions explains a good part of the money being lost today on artificial intelligence projects.

Producing became free, understanding didn’t

There’s a structural reason. For decades, building software was expensive and slow, and that friction served a purpose nobody was grateful for: it forced you to think first. If a build took six months and three people, someone was going to ask what for. That question didn’t come from organizational wisdom, it came from cost. Today the friction of producing has disappeared and the friction of understanding stayed exactly where it was. Specifying an agent takes four minutes. Understanding the process that agent will intervene in costs the same as it did in 1995.

The danger isn’t that AI generates bad specifications. It’s that it generates plausible ones. A coherent, well-structured document with the right vocabulary looks a lot like a document that was thought through, and in a committee meeting nobody can tell one from the other. Plausibility became free. Understanding didn’t.

What Gartner says about agentic projects

The numbers are starting to show the bill. Gartner projected in June 2025 that more than 40% of agentic AI projects would be canceled before the end of 2027, and the reasons it lists aren’t technical: escalating costs, unclear business value and inadequate risk controls. In the same analysis it estimated that of the thousands of vendors calling themselves agentic, only about 130 really are. The rest is what the industry has already named agent washing: chatbots and the usual automations with new vocabulary. The line that stuck with me most is from analyst Anushree Verma: many use cases positioned as agentic today don’t require agentic implementations.

Read it again, because it isn’t saying agents fail. It’s saying the problem was never an agent problem.

Who is in charge of thinking?

This is where the conversation stops being technological and becomes organizational, which is where it always ends up. If the bottleneck is defining the problem well, the reasonable question is who in the company has that job. And the uncomfortable answer, if you look at any org chart, is nobody.

We have people paid to deliver. We have delivery metrics, delivery dashboards, delivery follow-up committees. What barely exists is someone whose job, evaluated and paid, is to say out loud that we still don’t understand the problem. That sentence isn’t seen as rigor. It’s seen as delay. And in an organization that measures speed, delay gets punished even when it’s the only thing that would have prevented six months of repairs.

The asymmetry became brutal with AI. Proposing two weeks to map a process, interview the people who run it and understand why it exists, when the team knows a model can deliver the full specification before coffee, no longer looks like prudence. It looks like obstruction.

Give the function back, don’t invent new roles

I don’t think the answer is inventing new roles. We’ve seen that movie with the data scientists who arrived at companies that had no clean data and no questions to ask them. The function I’m talking about doesn’t need to be created, it needs to be given back. It used to sit with process analysts and architects that many organizations flattened into execution roles, because thinking didn’t show up in any indicator and delivering did. The difference is that before, that loss was paid slowly, and today it’s paid fast, in the form of agents nobody uses and automations that have to be dismantled.

A simple test for your organization

There’s a simple test to know where your organization stands. The next time someone arrives at a meeting with a proposed AI solution, ask what problem it solves and stay quiet. If the answer describes how the tool works instead of the business pain, they’ve just shown you where your real bottleneck is, and it isn’t the model.

I learned this by taking apart my own system, which is the expensive way to learn it.

Who in your organization has explicit permission today to say we still don’t understand the problem well, and what happens in their performance review when they say it?