Most people are using AI back to front

I see it all the time. People use AI back to front.

They ask it to write the strategy, come up with the ideas and draft the opinion piece. Then they spend the afternoon on the things AI is good at: searching, reading 10 articles to find 2 useful paragraphs, pulling numbers into a table, summarising a report nobody will read in full.

That gathering adds up. In 2012 the McKinsey Global Institute estimated that managers and professionals spend nearly 20% of their working week looking for internal information or tracking down colleagues who can help. That’s roughly a day a week (and it’s an old number, from before anyone had a chat assistant to hand the searching to).

I understand why people hold on to that work. Grunt work feels like work. You’re busy, you’re producing something, and nobody asks awkward questions about a spreadsheet you’re filling in.

Thinking is harder to see. From the outside it doesn’t look like much, and at the end of it you have to commit to something.

It’s the wrong way round, and I think it’s the biggest mistake people make with AI right now.

What to hand over

Research is the obvious one. How 5 competitors price a product, what the current guidance says on a topic, what’s changed in a framework since you last looked. That’s information gathering, and AI is fast at it.

How fast is harder to pin down. Anthropic sampled 100,000 conversations on Claude.ai and estimated that the tasks in them would have taken about 90 minutes without AI, and that Claude cut that time by about 80%. Those are Claude’s own estimates, and Anthropic says plainly that they leave out the time people spend checking the output, so the real saving is lower.

The same goes for first drafts of things you already know how to write, summaries, comparisons, meeting notes, and turning a messy document into something structured. Each of these needs someone to check it, which is a much smaller job than doing it.

Here’s what handing over research looks like in practice:

Find how Asana, Monday.com, ClickUp, Notion and Wrike price their team plans
this month. Put the per-seat monthly price, the minimum number of seats and
what the cheapest plan leaves out in a table. Link the vendor's pricing page
for every figure. Flag anything you couldn't confirm on the vendor's own site.

The last 2 sentences are what make the table checkable. Without the links and the flags you’re trusting a confident table, and you’ll end up redoing the research to find out whether you can.

What to keep

The part that needs you starts when the information arrives.

Deciding which findings matter for your situation. Knowing that the best practice everyone recommends won’t work in your organisation, because you know how it actually runs. Spotting the link between something in the research and a conversation you had last month. Having a point of view, then owning the decision and living with it.

AI can’t do that. It doesn’t have your context and it doesn’t carry the consequences. And it will happily give you a confident answer to the wrong question.

This is the creative side of work. Making something that didn’t exist until you thought about it.

Your expertise matters more now

There’s a fear that AI makes expertise less valuable. I think the opposite is true.

When AI does the gathering, the bottleneck moves to judgement. The best evidence I know for this is a 2023 field experiment by researchers from Harvard Business School and others, run with 758 consultants at Boston Consulting Group.

On tasks that suited AI, consultants using GPT-4 completed 12.2% more tasks, finished 25.1% faster and produced work rated more than 40% higher in quality. On a task designed to sit outside what AI could do well, consultants using AI were 19 percentage points less likely to get the right answer.

The researchers call this a jagged frontier. AI improves performance on some tasks and makes it worse on others that look just as hard, even inside the same workflow. You can’t tell which side of the line you’re on without knowing the subject.

Delegation only works if you can judge the output. The more you know, the more you can safely hand over.

One caveat

Some research is how expertise gets built. If you’re new to a field, working through the material yourself is often the point.

An Anthropic randomised trial showed this clearly. Anthropic recruited 52 mostly junior developers to learn Trio, a Python library none of them had used. The ones with an AI assistant averaged 50% on a quiz about the concepts they’d just used, against 67% for the ones who coded by hand. The biggest gap was on debugging questions.

How they used the assistant made the difference. The participants who scored well asked follow-up questions and requested explanations, and used AI to understand the code as well as to write it.

So the rule depends on where you are. When you know the field, delegate and check. When you’re learning it, ask AI to explain the material before you ask it to produce anything.

A question for your week

Look at what you did last week and split it in two: the work that needed your knowledge and judgement, and the work that just needed doing.

I suspect for most of us the second pile is bigger. That’s the time AI should be giving back. What you do with it is the interesting part.

If you look after a team where most people still use AI as a faster way to look things up, this is what I teach on the Power User Gap, on the team’s own work.