Everyone’s telling you to hand off more to AI. That’s backwards.
The question isn't what AI can do for you. By now you've figured that part out. The real problem is quieter: will you notice when it gets it wrong?
Because AI doesn't fail loudly. It fails fluently. It hands you a paragraph that sounds right, a number that looks right, a recommendation that reads like it thought it through. And you ship it without noticing. A mistake you'd have caught in thirty seconds.
A Harvard study tested 758 consultants on this. On most tasks, the AI users did better. But on a task where AI was subtly wrong, the AI users did worse than consultants who didn’t use AI at all. By 19 percentage points.
The confident wrong answer was more dangerous than no answer.
That's the real problem with handing work to AI. Not what it can do. Whether you can catch it when it’s wrong. So before you assign more work to AI, ask "if AI does this badly, will I notice it?”
How to hand off a task.
Two questions decide what to do with any task:
Can AI do it? Can it produce a first pass that's 70% right?
Can you check it? If AI is wrong, will you catch it in 90 seconds or less?
Answer those two questions for any task and you land in one of four boxes:
Hand it off. This is where you want to live. Meeting and sales-call summaries, competitor pricing pulled into a table, reformatting data you already have, routine customer replies where the facts are easy to check. AI does the work, you scan the output in a minute, it leaves your desk. If you can eyeball it and catch errors fast, hand it over.
Hand it off, but check it. This is the trap, and it's where most people spend their AI time. Client proposals, first-draft strategy memos, weekly revenue updates, refund emails to unhappy customers where one wrong detail turns a save into a legal problem. AI will get you to a first draft faster than you could write it, but checking AI’s work isn't about whether it sounds right. It's about facts, tone, and commitments. If you skip the check, you're not saving time. You're stacking up problems you'll pay for later.
Describe what "checked" looks like. "I'll check every number against the source doc, read the scope section twice, and make sure the tone matches how I'd talk to this client." If you can't write that sentence, the task isn't ready to hand off yet. It's still a “keep it yourself” task.
Skip AI. Tasks where AI gives you something that sounds fine but is wrong in ways you can't easily spot. The "what's my revenue this month" question where AI confidently makes up a number that looks reasonable. Quick one-offs where prompting takes longer than just doing it. If correcting AI takes longer than doing it yourself, AI isn't helping. Do it yourself, or hand it to a human.
There's a whole group of these tasks, the ones that quietly eat your week, that an actual person handles better than AI. Booking appointments. Negotiating bills. Sitting on hold with the airline. This is where a human assistant still wins, and I've found it doesn't have to cost what it used to. I use Duckbill* for this. It starts at $50/month (versus $2,500+ for a traditional assistant), and me+machine readers get three months free.
*not a paid promotion, but I do have a friend who works there.Keep it yourself. Pricing. Positioning. Whether to take on a client. Whether to fire one. Hiring decisions where the signal is in what the resume doesn't say. These are hard to check because judgment calls don't have a right answer to check against. That's also why you shouldn't hand them off, even when AI gets good enough. Doing them is how you stay close to your business. Give them up and you're not running the business anymore. You're watching it.
But the boxes aren't fixed. The same task moves between them depending on how you set it up. A weekly revenue view is hard to check if AI gives you a raw number. It's easy to check if AI gives you a trend line next to last month's numbers, because you'd spot a 10x error instantly.
The format is the check. So is a checklist written before the draft. That's the leverage. You don't just sort tasks. You upgrade them.
What you lose when you hand off too well.
The four boxes sort this week's tasks, but they don't tell you what happens to YOU over the long run if you sort them wrong. The thing you lose is the one thing your business actually runs on: your judgment.
The same research team did a follow-up study with 244 of those consultants. They found three patterns. Some used AI as a clean hand-off. They did the parts they were best at, let AI handle the rest, and came back to finish. These people got better at their own job. Others used AI in every step of their work. They got better at AI. And a third group handed off entire tasks and walked away. They got the output, but didn't engage with the work. These people didn't get better at their job or at AI. They were on autopilot.
Going on autopilot looks like productivity. The drafts go out. The proposals ship. The numbers get pulled. It feels great for a few months. Then you'll sit down to make a real decision (a pricing call, a hard client conversation, a pivot) and realize you haven't thought about your business in a while. You've been managing outputs, not making decisions.
The fix isn't to hand off less. It's to stay in the work while you hand off. Checking, done right, forces you to.
Checking is the work you keep.
Most people treat checking as the cost of using AI. The tax you pay to get leverage. That's why so many people end up on autopilot.
But let's flip that. Checking isn't the cost of leverage, it's where your leverage actually lives. Google DeepMind made the same argument in a February 2026 paper: no checking, no real hand-off.
If you hand off the draft and also hand off the check, you're not working faster. You're working blind. The leverage comes from AI doing the volume and you doing the judgment. Skip the judgment and the whole trade falls apart.
So how do you check without burning the time you were trying to save? Three habits. Each one is a piece of judgment you refuse to hand over:
A definition of "done" before the draft. Before you ask AI to write anything that matters, write down what "finished" looks like. Not a prompt — what good looks like. "Tone sounds like me talking to a longtime client. Scope stays under $15k. Must mention the usage cap. Must not commit to a timeline." Three to five lines. Takes ninety seconds. Do this before AI writes, not after. Once AI produces a draft, you'll latch onto what it gave you. You start editing its judgment instead of asserting yours. Writing it down first keeps your taste in front of the machine, not behind it.
Check the one thing that has to be right. You don't have to check everything in an AI output. You have to check the one thing that, if wrong, hurts most. The number going in the proposal. The date you're committing to. The customer detail that signals whether you were paying attention. Pick the claim that has to be right, check only that, ship the rest. This keeps your risk judgment without drowning you in line-by-line review. People who try to check everything give up and check nothing. The one-thing check is the middle path.
The 90-second test. Before you hand off a task, ask: can I check the output meaningfully in 90 seconds? If yes, hand it off with confidence. If no, you have two choices. Build a faster check (better format, clearer "done," smaller scope), or don't hand it off yet. What you can't do is hand it off anyway and hope. That's how you end up on autopilot. The test isn't about speed. It's about whether you've designed the hand-off or just thrown work over the wall.
None of these are prompt tricks. They're what stays yours when everything else gets faster. AI will keep getting better. The habits don't get less valuable. They get more valuable, because the gap between "sounds right" and "is right" keeps growing.
Checking is the work you keep. Everything else, you can negotiate. Pick one to try this week. Hint: the right one is usually the one you keep almost doing.
AI will keep getting faster. Will you still be in the work while it does?
This is Issue #5 of me+machine. Every week: one way to make AI work harder for you. Nothing you need a tech background for.
Sort your tasks. Name your checks. Keep your judgment.
me+machine
Your turn: What's one task you shouldn't be handing off to AI, even though you could?