AI is most useful when it reduces friction around thinking without pretending to own the decision.
There is a temptation to frame every AI question as a contest between enthusiasm and fear. Either the technology will transform everything, or it will weaken the skills we value. Both positions contain legitimate concerns, but neither is especially useful when we sit down to do actual work.
A more practical question is narrower: which part of this work is repetitive, which part requires judgment, and where would another perspective help?
Let the tool carry repetition
AI can compare options, surface patterns, reorganize a messy draft, extract themes from notes, or create a first pass. Those are not small benefits. Repetitive work consumes attention, and attention is often the scarce resource behind better decisions.
Consider the early stages of a project. Notes may be scattered across documents, messages, and memory. A tool can help consolidate the material, identify repeated concerns, and propose a structure. That does not eliminate the need to understand the project. It makes the raw material easier to examine.
The same is true in writing. A blank page can carry unnecessary pressure. A deliberately imperfect first pass gives us something to question: Is this true? Is this my voice? What is missing? What is too smooth?
A first pass is valuable because it creates something to think with—not because it has earned the right to be final.
Fluency can disguise weakness
AI-generated language often sounds complete before the thinking is complete. This is one of its most useful qualities and one of its most serious risks.
A polished summary can miss what was implied. A confident recommendation can rely on a false premise. A list of options can hide the fact that the options were framed too narrowly. Fluency makes these weaknesses harder to notice because the answer arrives without the hesitation that normally signals uncertainty.
This means review cannot be a quick scan for grammar. We have to inspect the reasoning. What evidence supports the conclusion? Which perspective is missing? What assumption would change the answer? Is the output accurate, appropriate, and useful for the people affected?
Judgment is more than correctness
Good judgment includes context, responsibility, and taste. It asks not only whether an answer is technically correct, but whether it belongs in this situation.
A message can be accurate and still be insensitive. A workflow can be efficient and still make a customer feel trapped. A recommendation can optimize a metric while quietly creating cost somewhere else. These are not edge cases. They are where human responsibility becomes visible.
That is why using AI well requires more than knowing how to prompt. It requires knowing what a useful answer would look like before the tool produces one—and being willing to reject an impressive answer that does not meet that standard.
Build an intentional division of labour
I find it helpful to decide explicitly what the tool may do and what remains mine.
- The tool can organize; I decide what matters.
- The tool can generate options; I evaluate the tradeoffs.
- The tool can challenge a draft; I remain responsible for the claim.
- The tool can reduce repetition; I protect the parts of the work that build understanding.
This division is not fixed. It changes with the risk of the task, the sensitivity of the information, and the cost of being wrong. The more consequential the decision, the more visible the human review should become.
Use AI to increase agency
The best outcome is not maximum automation. It is greater agency: more capacity to understand, decide, create, and follow through.
If a tool makes us faster but less able to explain our own work, something important has been lost. If it helps us see a pattern, test an idea, or free attention for a better conversation, it has earned its place.
Protect the learning loop
Some friction is waste. Other friction is how understanding develops. Writing a first explanation, wrestling with a difficult concept, or tracing why a decision failed can be slow because the learning is happening inside the effort.
If AI removes that effort too early, we may receive an answer without building the mental model needed to evaluate it. The result is dependency disguised as productivity.
A useful practice is to decide when assistance enters. We might first outline what we believe, then ask the tool to challenge it. We might solve a representative example ourselves before automating the repeated pattern. We might use AI to explain a gap, then restate the idea in our own language.
Keep provenance visible
Trust becomes harder when generated material is mixed invisibly with observed facts. Notes, assumptions, sourced information, and AI suggestions should not all look the same.
In consequential work, we should be able to answer: Where did this claim come from? What was verified? What remains an inference? Who reviewed the final decision?
This discipline is not only about preventing error. It preserves the chain of responsibility that allows a team to correct an error when one appears.
Make the relationship deliberate
Tools shape habits. If we always ask for conclusions, we practice accepting conclusions. If we ask for counterarguments, missing questions, and alternative structures, we practice examination.
The quality of the relationship depends partly on what we repeatedly ask the tool to do. A thinking partner should make our reasoning more active, not more passive.
AI does not make judgment unnecessary. Used well, it makes the quality of our judgment more visible—and gives us more room to practice it.