How to Use AI to Build Agency, Not Bypass It
Essay 8 of the AI Contract Series
The distinction that determines whether AI makes you more capable or more dependent
There is a question underneath every AI interaction that almost nobody is asking.
Not “did this work?” Not “was the output good?” Not even “did this save me time?”
The question is: what did this do to me?
Not to the task. To you. To your capacity to understand, choose, act, and influence outcomes. To the part of you that does the thinking, makes the call, develops the judgment, builds the expertise.
That is the Agency question. The answer — across thousands of interactions, accumulated invisibly — is what determines whether AI is making you more capable or slowly substituting for the capabilities you thought you were keeping.
This essay is about how to use AI so the answer comes out right.
Why the distinction matters more than the tool
The THX framework is not anti-AI. It is a diagnostic — derived from twenty-five years of empirical observation about what interactions do to the humans inside them. When pointed at AI, it finds both genuine benefit and genuine risk. Understanding the distinction between them is not about using AI less. It is about using it in a way that leaves you more capable after the interaction than before.
The 12 Utilities map where AI delivers reliably: Speed, Accuracy, Ease of Use, Availability, and Access. These are real. The democratization of research, expertise, and analytical power that AI enables is one of the most significant expansions of access in human history. Someone without institutional resources can now work with the same quality of information and assistance as someone with all of them. That is worth saying plainly, without qualification.
The risk is not in the utilities. It is in the layer above them.
Agency — the felt capacity to understand, choose, act, and influence outcomes — is what transforms utility into development. A fast, accurate, accessible tool that does the work for you delivers five utilities and removes the process through which you become someone who could do that work yourself. The utility is real. The cost is invisible. The cost accumulates exactly as the benefit does: interaction by interaction, across time, below the threshold of notice.
The distinction that determines which way it goes is simpler than most frameworks suggest.
The one distinction
Use AI on tasks where output is the point.
Protect tasks where process is the point.
That is the whole framework. Everything else is application.
Output-is-the-point tasks are ones where the value lives entirely in the result: formatted, complete, delivered. A travel itinerary. A meeting summary. A boilerplate contract clause. A literature search. A first-pass transcription. Formatting a spreadsheet. Translating a document. Generating variants of a headline you already know the direction of. For these tasks, doing it yourself would not have built anything in you. The friction was purely administrative. Remove it freely. The Agency cost is zero because the process was never the point.
Process-is-the-point tasks are ones where the value lives in the experience of doing them — where the difficulty is not an obstacle to the output but the mechanism through which something is built in you. Strategic thinking you have not done before. Creative synthesis that requires you to find your own argument. Problem-solving at the edge of your current capability. Writing where finding the words is how you find the ideas. These are places where AI-generated output is not a shortcut to the destination. It is a bypass around the terrain that was the destination.
The confusion is that both kinds of tasks look the same from the outside. Both produce outputs. Both can be completed faster with AI. Both rate well on satisfaction measures. The difference is entirely in what was built — or not built — in the human doing them.
What this looks like in practice
The principle is clean. The application requires judgment, because the same task can fall on either side of the line depending on where you are in your development and what you are trying to build.
For a student writing their first research paper, the writing is process. The act of constructing an argument in prose, finding the places where the reasoning breaks down, discovering what you actually think through the pressure of having to say it: these are not obstacles to the paper. They are the education. AI-generated drafts bypass the mechanism entirely. The paper gets produced. The student does not develop.
For a senior executive writing their hundredth strategic memo, the drafting may be output. If the thinking is done and the draft is a formality, AI assistance on the draft may cost nothing in development. The strategic judgment that made the memo worth writing was not bypassed. It was applied.
For anyone encountering a genuinely novel problem — the kind where you do not yet know the shape of the question, let alone the answer — AI can be a thinking partner or a bypass, depending entirely on how it is used. Using it to generate an answer before you have sat with the problem is a bypass. Using it to pressure-test an answer you have worked toward yourself, to find the gaps in your reasoning, to surface what you have not considered: that is Agency-preserving use. The same tool. The same prompt, sometimes. A completely different relationship to your own development.
The practical test: after the AI-assisted task is complete, could you explain how you got there? Could you defend the reasoning, identify the assumptions, recognize the weaknesses? If yes, AI served your Agency. If no — if the output appeared and you accepted it — the process was bypassed, whether the output was good or not.
The five uses that reliably build Agency
Beyond the general principle, five specific modes of AI use reliably preserve or strengthen Agency rather than substituting for it.
The challenger. Use AI to argue against your position. After you have developed a view — genuinely worked toward it, not just thought about it for thirty seconds — ask AI to make the strongest possible case against it. This use forces you to defend your reasoning rather than outsource it. It builds the evaluative capacity that the False Helper archetype (Essay 6) most aggressively atrophies.
The expander. Use AI to surface what you have not considered. After you have mapped a problem as far as your own thinking can take it, ask AI what you missed, what adjacent questions you have not asked, what the people who most disagree with your framing would say. This use extends the range of what you examine without removing the judgment about what to do with what you find.
The stress-tester. Use AI to identify weaknesses in reasoning you already believe is sound. This is different from asking AI to evaluate your work generally — which often produces validation. It is specifically asking for the places the argument breaks down, the assumptions doing more weight than they should, the scenarios under which the conclusion fails. This builds evaluative rigor in a way that general feedback does not.
The research amplifier. Use AI to aggregate, summarize, and surface relevant material in domains where your own knowledge is thin. Then read the material. Do not let AI replace the reading. Aggregation is an output task — the value is in knowing that these sources exist. Synthesis is a process task — the value is in the understanding built by reading them. Use AI for the first and protect the second.
The articulation mirror. After writing or thinking through something, ask AI to reflect back what it understood you to be saying. Not to improve it: to mirror it. The gaps between what you meant and what it reflected are diagnostic. They show you where your articulation is unclear, which is usually where your thinking is unclear. This use makes AI a tool for developing your own clarity rather than substituting for it.
All five uses share a common structure: AI operates downstream of your thinking, not upstream of it. The human does the generative and evaluative work. AI amplifies, challenges, and extends it. That sequence — human thinking first, AI engagement second — is the practical architecture of Agency-preserving use.
The trap that looks like efficiency
There is a failure mode that is easy to fall into because it looks like exactly what it is not.
The trap: AI generates a good output, you refine it, you approve it, and you call the process a collaboration. It feels like you were in the driver’s seat. The output has your judgment applied to it. What more could Agency require?
What is missing is the generative phase. The part where you did not yet know what you thought, where you had to sit with the problem until a direction emerged, where the false starts and the confusion were doing the work of building the understanding that would eventually produce the answer. That phase is where expertise is built. That phase is what AI bypassed when it generated the first draft.
The refinement and approval phase develops a narrower and weaker capacity: the ability to evaluate outputs against an existing standard. That capacity is valuable. It is not the same as the capacity to generate, which requires having lived inside the problem long enough for your understanding of it to deepen.
The senior person in any field who has spent decades generating — developing judgment through the slow accumulation of experience, error, correction, and refinement — can evaluate AI outputs with significant accuracy because they have the deep model that evaluation requires. Someone early in their development who bypasses the generative phase to jump to evaluation is developing the capacity to check work that requires the work they are not doing to evaluate well.
This is the Agency trap. Not the absence of effort. Effort applied to the wrong phase of the process.
Where to draw the line in your own work
The framework is principle. Applying it requires knowing yourself — knowing which tasks are genuinely administrative for you, which are genuinely developmental, and which are somewhere in between.
The honest audit is simple. For any task where you are considering AI involvement, ask one question before starting: if I do this myself, will I be more capable afterward? Not more informed about the topic — more capable as a thinker, writer, analyst, or decision-maker?
If yes: do it yourself first. Then, if useful, engage AI to challenge, expand, or stress-test.
If no: use AI freely. The friction you would have experienced was not the mechanism of anything. Remove it and move on.
If you are not sure: default to yourself first. The cost of unnecessary work is time. The cost of bypassing necessary process is the capability you would have built and did not.
The accumulation of that second cost, across a career, is significant. It is also entirely preventable — not by avoiding AI, but by knowing which tasks are which and protecting the ones that matter.
Essay 9: What excellent AI interaction actually looks like — the three positive archetypes, and the named patterns for interactions that deliver all five reliable utilities while building rather than depleting the human inside them.