---
date: 2026-09-05 19:57
tags: [work]
---
Not a month passes without hearing the news that a peer has started a paid Claude or OpenAI subscription and begun integrating AI deeper into their workflow. From them, I have heard much praise about how “everything is within reach now” or how this allows them to spend time on more interesting tasks without dealing with practical details. 

And me? So far, my personal LLM usage has been relatively limited. Most of my interactions were unstructured exchanges with chat assistants about conceptual questions or small code snippets, but the integration was limited to that. I have always had a few concerns regarding careless use of LLMs. Some of them are conjectural, such as the high energy footprint and the rebound effects on the economy of consumer electronic goods. These are bound to be solved as the hype fades and technology improves. Others are structural: the risk of deskilling and the ethical problems of untraceability of training data.

In the meantime, two facts have become clear to me: AI is here to stay, and AI is becoming undeniably good at tasks that were previously typically human trades. As people adopt AI into their workflow, the productivity baseline that people expect from others increases. As a manager, it will become less and less viable to hire someone that guarantees no AI intervention in their work. The other perspective is that it will become harder for a non-AI user to compete with mindful AI users, assuming equal expertise. The implication of these observations is that clinging to the old-school way is probably a bad economic personal decision in the long run.

How to reconcile this logical conclusion with my doubts about the technology? Is there such a thing as mindful AI usage in work? These questions have been at the center of my mind for a few weeks now. So much so that I have decided to write this post. Leaving time to solve the conjectural problems, and society to act on the ethical issues, the question that remains mine to answer is my personal relationship with the tool: how to mindfully leverage LLMs while maintaining a sharp understanding of my work? How to prevent laziness and skill decay for the tasks I am good at? At the time of writing, I do not know what to expect in the long run, but I know that history has shown time and again that the introduction of a new tool decreases the lower-level know-how but allows users to gain higher-level understanding and control. I think I must approach my delegation of tasks to AI in a way that makes a skillset organically shift from a lower-level oriented expertise to a higher-level expertise. This is what I would try:

1. Maintain a rigorous mental model of my work.
1. Any AI output is to be confronted with the mental model: then, reject or actively adopt into the understanding. This includes actively auditing the AI's output with my knowledge.
1. Delegate well-bounded subproblems whose purpose, constraints, and role in the larger task I understand.

These principles are likely to hinder some of the theoretical productivity speedups reachable with AI use, but they ensure that AI is used as a *tool* and that the user *owns* (in liability, intellectual property is another issue) the output of the AI. By instrumenting the AI this way, a higher-level task expertise is actively curated, maintained and extended. A consequence of these three principles above is that the user has no choice but to gain an understanding of any technical decision they wish to defer to an AI if the subtask is not explicitly a black-box within the mental model.

I will leave these ideas to mature by themselves and maybe one day, I will be ready.