Delegate or Relegate? What Should AI Handle for You?

We are beginning to see a clear difference in how people use AI. Many people use AI as an advisor. They ask for personal advice and follow it. Others use it to build tools that allow them to make their own decisions. Because the former is more common, expectations between AI consultants and their clients can easily diverge. To avoid this, we need to differentiate two concepts we often conflate and use interchangeably.

First, we need to ask ourselves what a “decision” is. If the question is “What is 12 x 32?” we would not call “384” a “decision.” In other words, if the answer can be derived by logic and known facts, we are not actually making a decision. Say you cook a plate of pasta but something is missing. So you ask, “What does this dish need?” You notice that it is completely lacking acidity, so you add a splash of lemon. This is not a “decision” either, because you derived the answer through reasoning.

Likewise, many machine learning models, like those that recognize human speech and handwriting or classify flowers and spam email, are not making “decisions.” They are deriving answers based on past patterns.

Life is full of uncertainties and unknowns, yet we have to choose a path in order to move forward, to achieve something. Let’s say two companies offer you a job and you have to choose one. You can collect a lot of data about both companies to help you make a decision, but such an effort is ultimately only a drop in the bucket of what you would need to derive an answer purely from reason. More data isn’t necessarily better either, because it is today’s data, which may become irrelevant or even contradictory if something fundamentally shifts in the economy (like the introduction of LLMs). Against these insurmountable uncertainties, we still have to make a choice. That is what a “decision” ultimately is.

So, the first difference we need to keep in mind is decision versus derivation.

The second cut we need to make is between delegation and relegation. Let’s say you own a business and you need a new website. You pick a person for the project and tell her, “Build a website for us. You know what to do. I trust your judgment.” That is delegation: you are asking her not only to design, code, and write the content of the website, but to be responsible for the outcome.

Relegation, on the other hand, is where you tell someone exactly what to do to achieve the goal you seek. Instead of handing off responsibility, you are handing off only the execution of it. If you were doing this to a human, it is where you would be accused of “micromanaging.”

But in the new era of agentic AI, we need to rethink what these terms mean.

AI agents can execute many tasks much faster than humans can, but one thing they cannot do is take responsibility. So even though AI can fly an airplane, we are unlikely to see a self-flying one any time soon. We as passengers will demand someone on board who can take responsibility for our lives.

What this means is that we cannot truly delegate anything to AI agents. We can only relegate. Technically you could delegate, but if an AI agent makes a wrong decision, you will be on the hook anyway. So you still need to know and understand what decisions it is making. “Sorry, ChatGPT wrote it, not me,” doesn’t hold up in a court of law. This needs to be clarified when AI consultants design systems for their clients; otherwise the consultants themselves may be on the legal hook.

In this new era of agentic AI, we humans will relegate as many derivations as possible to AI while taking on as many decisions as we can. Here is how that will look in practice.

Traditional software applications built on logic or algorithms were derivation tools. They are not going away. What changes is who builds them: AI agents, not humans. Derivations must be automated as efficiently as possible, which means building proprietary systems optimized for your own business. No more sharing the same application with millions of others, sending in feature requests and hoping someone will read them.

Many new possibilities have emerged with the advent of machine learning and the fine-tuning of LLMs. These are still tools for derivation. Even though the latter can be used to make decisions, you cannot delegate those decisions, because you remain on the hook for any consequences. You still have to understand the decisions the LLMs are making, so technically, you are still making all of them yourself, whether you know it or not.

Blurring these concepts can potentially cause catastrophic failures. If millions of people start using OpenClaw without delineating them, the economy could crash with nobody to be held accountable for it.