Becoming an AI Power User

Before any of the names, here is the idea under all three. The division of thinking between you and AI is a setting you can turn, and where you turn it changes what you get. Turn it one way and you tend to get accuracy and sharper judgment about the task. Turn it another way and you get fluency with the model itself, plus room to play and try things you could not try alone. Turn it a third way and you get speed with little learning. No setting is correct in the abstract. The move is to look at the task, decide what it most needs, and turn the setting to match before you start.

Researchers at Harvard Business School and Boston Consulting Group watched people split work with AI on real consulting tasks and saw two of these patterns show up on their own, without anyone teaching them: dividing the work at a clear seam, and interweaving it in tight back-and-forth. This book adds the third, full delegation, to complete the set. As of 2026 the work in that study looked like writing prompts into a chat model, and the interface will keep changing under it, toward voice, then glasses and earbuds, then a brain-computer interface, and toward far more capable systems after that. The three splits survive the change, because they describe how two intelligences share a job, not which screen they share it through. Even when the intelligence on the other side is far stronger than today's, you are still choosing how tightly to interleave your thinking with it. The point worth carrying is that no single split wins across the board, so naming what each one is good for lets you choose with intent instead of settling into whichever feels easiest.^[1]

Direct the work when accuracy matters most and you can split the task cleanly

Start with the split where you stay the planner. You break the task into pieces and hand each piece to whichever side is stronger on it. You might keep the strategic call about what the email should accomplish, and hand AI the formatting, the research, and the arithmetic. Or you let AI write a first draft and keep the editing and the fact-checking for yourself. The line between your pieces and the model's is not fixed in advance. Where the model's judgment on a piece is better calibrated than yours, you can let it lead there; where yours is sharper, you keep that piece. You are the one deciding who does what, piece by piece.

Make it concrete with the email. You decide the structure, AI drafts inside it, you edit for accuracy and tone. You write the three points the email has to land and the order they go in, because that is the strategic call and you know the reader. You hand the drafting of each point to AI. Then you read the draft against what you wanted and fix what is off. The reason this works is that the pieces come apart cleanly: you can judge the model's draft of point two without it being tangled into your decision about the order. This split is at its best when the pieces separate cleanly and you can check the model's contribution on its own. When the pieces bleed into each other and you cannot tell whose call produced what, the clean handoff breaks down and a different split fits better.

Fuse with the model when you want to get more fluent with AI itself

The second split removes the seam. You and AI work in tight back-and-forth, trading at the level of single sentences, so closely that neither side could have produced the result alone. You write a sentence, ask AI to strengthen it. It offers an argument, you push back on it. You revise a number, it recalculates around the change. There is no moment where you hand off a whole piece and wait. The thinking weaves together as you go, and the work that comes out carries both of your fingerprints on every line.

Here is the email again, fused this time. You type the opening line. You ask AI to sharpen it, and it gives you a version that is crisper but slightly too formal. You take half of its change and reject the other half in the same breath, then write the next line yourself, then ask it to pressure-test the logic of the two lines together. The output and the conversation are the same thing, because you are never out of the loop long enough to receive a finished piece and review it. You are inside the drafting the whole way through.

Delegate the whole task when speed is the priority and the stakes are low

The third split hands over everything. You give AI a whole task, write the report, run the analysis, build the summary, and your job becomes reviewing what comes back to approve it, reject it, or send it back. You are not breaking the task into pieces and you are not trading sentence by sentence. You describe the result you want, the model produces a complete version, and you decide whether it is good enough to use.

Delegated this way, the email is one move: you tell AI who it is going to, what it needs to accomplish, the tone, and any constraints, and it writes the whole thing while you wait. This split is the fastest of the three, and on the right tasks that speed is exactly what you want. It fits work that is low-stakes, well-defined, and repeatable: formatting a document, scheduling, routine summaries, standard messages where the cost of a small miss is small. For that kind of work, spending your attention on a clean seam or a sentence-by-sentence trade would be using more system than the work needs.

Pick the split by which axis the task most needs

Lay the three splits side by side and the tradeoffs come into focus. Each split is strongest on a different axis, and choosing well means asking which axis this task most needs. The table below holds the directional tendencies, not measured scores. Rather than reading it as one long run of numbers, hold it as three families of difference: what you get out (accuracy and persuasiveness), what grows in you (task judgment and AI fluency), and what each one costs you (speed and the mental effort it takes). Read each column as one split's profile across those three families.

Read the table back as three plain defaults. Reach for directing when accuracy matters most and you can see how the task should be built. Reach for fusing when you are growing your fluency with AI, or working a problem that is new enough that going back and forth will surface a better answer than either side would reach alone. Reach for delegating when speed is the point and the stakes are low enough that a small drop in quality is fine. These are starting points you adjust with experience, not fixed rules.

Under all three splits sits one principle that does not change. You spread the work out and route each judgment call to whichever side is better on it, while you stay accountable for what you finally ship. That is how AI work scales: by distributing pieces and letting the better-calibrated side decide each call. Neither side's judgment is fixed, so the routing is not fixed either, since both your judgment and the model's keep improving. The split you choose just sets how tightly the two of you interleave. Directing keeps the routing deliberate and piece by piece. Fusing blends it in real time. Delegating sends the calls to AI and keeps your judgment for the review. The choice stays deliberate, made on the stakes, what you want to learn, and the shape of the task.

Once you have run the comparison, do not leave the lesson in your head. Encode it as a short rule that names the split you reach for by task type, then keep evolving that rule as you learn. For each kind of task you do often, record which split you choose and why. The rule does not stay frozen. As you run more tasks, you sharpen it, and you can hand the sharpening back to AI: ask the model to read your notes across several tasks and propose where your split choice tends to misfire, then fold its read back into the rule. This is the same move the chapter on turning each interaction into a compounding improvement teaches: a decision you made once, encoded where your next task can reach it and revised every time the work teaches you something new.

One more place these splits show up is when you run several AI conversations at once on one project, which the Parallel Work chapter develops. The three splits can coexist across the different threads. A research thread might run fully delegated, because the task is well-defined and low-risk. A strategy thread might run fused, because the reasoning needs your continuous judgment. Choosing the right split per thread is the same skill as choosing it per task, applied across a few tasks at once.

Run one task three ways to feel the difference