The art of delegation in the age of AI
The psychology of handing over real work and how to actually do it.
You open Claude, type a prompt, and wait. The output arrives. You refine it, ask for another pass, go back and forth a few more times, then copy what you've got into a doc, edit it down, and move on to the next thing.
I’ve had afternoons evaporate this way.
You’ve technically been working with AI the entire time. But the work itself was still entirely on you.
Most people who say they “use AI every day” are doing exactly this.
Last week we covered Dispatch, and how to send work to your desktop from your phone while you’re out doing something else. But underneath the practical question of how was a much bigger one that we didn’t cover.
Are most of us even capable of letting work go?
Real delegation has always been a skill. The age of AI has only made the cost of being bad at it spectacularly higher than it used to be.
Most of us have never been taught how. INSEAD’s research on delegation found that it was rated the second most important leadership skill, behind empathy, for mitigating employee burnout, yet only 28% of leaders reported receiving any training in it.
The skill was already in short supply before AI showed up. Now it’s the skill the whole AI era runs on.
Today I’ll show you:
Why most of us are still doing the work ourselves while pretending we’re delegating
Three habits that keep work stuck on your plate
Three things you can hand off to AI this week, with the specific setups I use
How to turn each one into a permanent system that compounds
Let’s get into it.
The delegation gap
The most widely cited piece of research on AI in knowledge work is the BCG study from Harvard, MIT, Wharton, and Warwick. It involved 758 BCG consultants across 18 realistic consulting tasks. Half were given GPT-4, half weren’t.
Consultants with AI completed 12.2% more tasks, 25.1% faster, and produced more than 40% higher-quality work on tasks within AI’s “jagged frontier” of capability.
But there’s a second number people quote less often. On tasks outside that frontier, where AI was confidently wrong, consultants using GPT-4 were 19% less likely to produce correct answers than consultants without it.
They used the same tool and instructions, but they had wildly different outcomes depending on how it was used.
In a follow-up paper published in 2026, the same research team went deeper. They analysed 4,975 human–AI interactions across 244 BCG consultants and identified three distinct ways professionals were actually working with AI.
Cyborgs (60%) fused with AI throughout the workflow. Treating it as a constant collaborator at every stage of the problem.
Centaurs (14%) directed AI selectively. They kept the analytical work themselves and used AI for targeted help on specific subtasks such as looking up methods, mapping a domain, or polishing a draft.
Self-Automators (26%) handed the entire problem to AI in one prompt, accepted what came back, and made superficial edits at most. Almost half of them accepted the output without any modification at all.
Cyborgs newskilled, they got measurably better at working with AI itself, building a capability they didn’t have before. Centaurs upskilled, they deepened their domain expertise by using AI to learn faster. But Self-Automators built neither. They produced output, yet didn’t learn anything in the process.
This places a magnifying glass on the emerging chasm in AI use. Across the same models and abilities, the gap is psychological. And it shows up as three habits we don’t do.
1. We don’t brief the AI
Imagine handing a piece of work to a stranger you’d never met. They know nothing about your business, your voice, your customers, your goals, or your taste. You’d expect the result to be generic. How could it be anything else?
That’s how most people use AI.
You wouldn’t get a useful email out of a new hire by saying, “follow up with that client.” They’d ask which client, what’s it about, what tone, and what’s already been said. You’d give them a brief, point them at the relevant docs, and tell them what good looks like.
With AI, it’s not immediately obvious to do that, because the chat box doesn’t make you do it. So you don’t.
INSEAD’s delegation framework calls this process-based delegation: rather than handing off a single task with detailed instructions every time, you teach the system how you make decisions, then let it apply that judgment across many tasks. The recommendation in their piece is to “record a video or a voice note that explains how you approach inbox management.” That turns the action into a process that can be repeated many times over.
The equivalents for AI already exist. Claude Projects and Skills. None of these are technical features. Yet they’re essential to set up correctly—think of them as onboarding documents for an infinitely patient employee who remembers (most) things.
This is the unglamorous infrastructure of delegation. If your AI doesn’t know you, all you can do is direct it and re-explain yourself time and time again. And direction is exhausting at scale.
2. We don’t actually let it work
A surprising number of capable, intelligent professionals would rather spend ten hours doing something themselves than two hours having AI do it and one hour reviewing the result.
There’s a name for this in management literature: failure to delegate. It’s one of the most well-documented blockers in new managers. People struggle to let go because they think they’ll do it better than anyone else, because reviewing someone else’s work feels like they’re not really working, or because the act of handing over makes them uncomfortable in a way they can’t quite articulate.
Every one of those instincts now applies to AI. And almost nobody admits to it.
The voice in your head says: I’d rather just do it. I’ll do it better.
You probably will. You’ll also do it ten hours later, when the moment to ship and move fast has already gone.
The research shows this preference permeating through to real work. A 2025 Microsoft study surveyed 885 of their own product managers about how they use generative AI. PMs delegated freely on tasks they didn’t strongly identify with, such as formatting meeting notes, scheduling, and summarising. But on tasks central to their professional identity, such as writing and strategy work, they resisted, even when the AI would clearly have been faster. The work was their identity. Handing it off felt like handing off a piece of themselves.
The hedonic treadmill makes all of this worse. The first time AI generated a thoughtful email for you, it felt like magic. By the hundredth time, you barely notice. You start nitpicking the output. You think I could just do this myself. You forget that three years ago, this same task would have taken you thirty minutes, scratching your head inside a Google Doc. The bar moves with the tools.
Pride is expensive in the age of AI. Everyone’s focused on the “AI replacing you” headlines, when in fact, the real cost is much quieter.
While you keep your hands on the steering wheel for tasks that don’t need a human at all, your competitors are moving ten times faster. That gap continues to compound for every day of inaction.
This is the part that requires real psychological work. A willingness to actually let go of something.
3. We don’t review what comes back
The third habit is the mirror image of the second. If under-delegating is I’d rather do it myself, over-trusting is I’m too busy to check it.
A 2025 study by Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers and analysed 936 real examples of AI use at work. The headline finding was that the higher the user’s confidence in the AI, the less critical thinking they applied to its output. The higher their confidence in themselves, the more critical thinking they applied.
In other words, as we get used to AI being good, we stop checking. The trust that should be carefully calibrated quietly becomes blanket trust. Then something slips through.
The model is structurally incentivised to produce output that looks correct. That’s what it was rewarded for during training. Whether the output is actually correct, either factually accurate, on-brand or well-judged, is a separate question that only you can answer.
If a junior on your team handed in a draft and said, “I’ve finished,” you wouldn’t sign it off without reading. You’d read it. You’d actually push back and ask clarifying questions and send it back if it wasn’t right. But you’d never just submit it without oversight.
That principle doesn’t change because the work was generated by a machine instead of a person. If anything, it gets more important.
Delegating without reviewing is abdicating.
Verifying and validating outputs with a critical eye, sense-checking the claims and pushing back where something feels off is the cheapest insurance you’ll ever buy.
It’s also the line between a professional who uses AI as a partner and one who’s quietly handing over their steering wheel—and agency—without noticing.
The shift
The unlock most knowledge workers haven’t felt yet is what happens next. Claude outputs can take anywhere from a few seconds to multiple minutes. Once your first task starts generating, you can parallelise your workflows. I like to batch context-relevant activities to avoid focus drift. One Claude conversation tab running research, another summarising a specific report, and another drafting the slide.
This is orchestration. And the team happens to be made of models.
Below are three things you can hand off to AI this week.