# AI - Role by Role

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [AI - Role by Role](https://dylit.info/ch/ai-role-by-role/6ab35d7099638e119f592ceb)

Where Your Time Actually Goes Once AI Is in the Loop Here is the version of this that nobody puts on a slide. Before AI, a piece of knowledge work broke down roughly like this: eighty percent producing it, twenty percent reviewing it. You wrote the thing, then you gave it a once-over, then it went out. After AI, the shape changes completely. Something like twenty percent setting up the request, thirty percent iterating with the tool, and fifty percent reviewing and deciding. Notice what happened to the review share. It went from a fifth of the work to half of it. The uncomfortable arithmetic If you are not spending real time on that review, you have not saved anything. You have moved the risk downstream and told yourself it was efficiency. This is the single most common failure in how teams adopt these tools, and it rarely gets named because it does not look like a failure. It looks like speed. A task that took two hours now takes twenty minutes, everyone is pleased, and the fifty percent that should have gone into checking went into starting the next thing instead. It works for a while. Most AI output is broadly fine, so most of the time nothing bad happens. Then a fabricated figure reaches a client, or a contract summary misses a clause that was actually there, and the accumulated risk arrives all at once. The time saving is real. It is just smaller than the raw numbers suggest, because a portion of the hours you freed up have to go back into the work in a different form. What "setup" actually involves Twenty percent on setup sounds like a lot for typing a prompt. It is not just typing. Setup means deciding what you actually want. Who is reading this, what decision it supports, what length, what tone, what it must not say. It means gathering the context the model cannot have: the background, the figures, the previous version, the example of what good looks like. Most people do this badly because they never had to do it explicitly before. When you write something yourself, all of that lives in your head and gets applied without ever being articulated. Handing the task to a system that cannot read your mind forces you to state it. That is why setup takes real time. You are doing decision work that used to happen invisibly while you wrote. What "iterating" actually involves Thirty percent iterating is the part people underestimate most. The first output is raw material. It is almost never the deliverable. The useful pattern is a conversation: this is too formal, tighten the second section, you have assumed something I did not tell you, give me three versions of the opening. People who treat the tool as a vending machine, one request in and one answer out, get vending machine results and conclude the technology is oversold. People who treat it as a working session get substantially better output, and the difference is mostly rounds of back and forth. Two or three rounds is normal for anything that matters. If you are past four and it is still wrong, start a new conversation with a better opening prompt. What "reviewing and deciding" actually involves This is the half that carries the weight, and it is two distinct jobs. Reviewing is verification. Every specific detail treated as unconfirmed until checked: names, numbers, dates, citations, quotes. The overall reasoning and structure usually hold up well. Details are where it falls apart. Scrutinise hardest in the areas you know least, because that is exactly where you are least able to spot a confident mistake. Deciding is the part that was always yours. Is this the right recommendation. Is this the tone we want with this client. Is this a risk we are willing to take. No amount of drafting help touches that question, and no amount of AI polish makes a bad decision a good one. The bottleneck moved The genuinely interesting consequence is not about time at all. Your constraint used to be production speed. How fast you could write, format, restructure, get the thing into a shareable state. That constraint is largely gone. The new constraint is your judgment. How quickly and well you can evaluate work, spot what is wrong, decide what to do. That is the bottleneck now, and it is a much harder one to widen. You cannot buy a faster version of it. Which means the returns from these tools depend enormously on how good your judgment already is. Someone with deep expertise in their field gets enormous leverage, because they can evaluate output quickly and accurately. Someone working outside their competence gets a fast route to confident mistakes. That is also why the advice to "use AI to work in areas you know nothing about" is backwards. Those are the areas where you can least afford it. What this means practically If you are managing a team through this, the number to watch is not how much faster things get produced. It is whether review time actually went up. A team that halved its production time and kept review time flat has not adopted AI well. It has just started shipping less-checked work faster. If you are doing the work yourself, a simple discipline helps: when a task finishes early, spend some of the time you saved on the output before you move on. Read it properly. Check the specifics. Ask whether you would defend every sentence. The shift from producing to reviewing is not overhead. It is the entire point. The value of these tools is that they move your effort from making things to judging them, and judgment is where your actual expertise lives. Getting the time back is easy. Spending it well is the part that takes discipline.