The aircraft was still flyable. Its engines were running, its controls were responding, and its structure was intact. What happened was that a cluster of small sensors on the fuselage iced over in a storm.

The airspeed readings scrambled for less than a minute, and the autopilot, which relies on those readings, did the sensible thing and switched itself off, handing the aircraft back to the two pilots at the controls.

They had flown countless times. What they had rarely done was fly a large aircraft by hand, at altitude, in the dark, with the automation gone and the instruments contradicting each other. One of them pulled the nose up and held it there. The A330 climbed, bled off speed, and slipped into a stall it never came out of.

It took three and a half minutes to reach the water, and for most of that time the crew did not grasp that the wing had simply stopped flying.

When investigators reconstructed it, the aircraft had remained capable of flying throughout. Plenty had gone wrong to get them there - confusing alarms, a stall warning that came and went, a crew struggling to read what the plane was telling them.

But they were also pilots whose opportunities to practise manual flying at altitude had become vanishingly rare - and with them the instinct you need when the machine hands the controls back.

Subscribe now

The trade

There’s a word for a skill wasting away through lack of use, and it’s one I keep hearing recently. Atrophy. It’s what happens to a muscle you stop using - it doesn’t disappear, it just weakens, until the day you try to use it and it just doesn’t respond.

The people I hear using the word aren’t pilots this time. They lead teams, run businesses, do knowledge work for a living. What they’ve noticed is that the more they hand to AI, the less certain they are they could still do the thing themselves. It gets done, and faster, but some part of the ability has diminished.

And I understand the worry. I recognise certain skills and behaviours I used to lean on that are undoubtedly weaker now.

But when I look honestly at my own work, atrophy only describes half of what’s happened. Because there are also things I can do now that I simply couldn’t a couple of years ago. Not least, get properly underneath a subject I knew nothing about a fortnight ago - aviation safety being the obvious one.

So what’s going on is a trade - some abilities weaken while others - ones that were out of reach before - come into range.

Which means there’s a better question here than whether or not AI weakens your skills. Because, yes, it definitely weakens some of them. The more important point is whether you get any say over which.

Some of what fades, you won’t miss. The problem is you’re not the one choosing what goes.

Share Ollie on Work

The slow way

Here’s the part I’m genuinely unsure about.

The old way I learned things was slow. I’d read the thing, wrestle with it, get it wrong, go back and do it again by hand. It was inefficient, it took ages - but the lessons stuck. What I learned that way, I still have up there.

Now I learn differently.

I go round and round on a problem with a machine - testing an idea, throwing it out, reshaping it, again and again. It’s quicker, but it often feels like I’m wasting my time - circling the same ground instead of getting anywhere.

The old way contained a lot of friction. Some of it was pointless. But some of it was the mechanism by which the lesson became mine. AI removes both kinds at once.

Slow or fast isn’t the thing that matters. What matters is whether I’m doing enough of the thinking for the idea to lodge in my own head, or just nodding along when the machine hands me a good answer.

I can’t tell yet which I’m doing.

Aviation’s response is one part of this where somebody made a deliberate choice. They didn't tear the autopilots out - the automation was saving far more lives than it ever cost, and everyone knew it. They changed the humans instead.

Pilots started being deliberately exposed to the kinds of conditions that had overwhelmed that crew - unreliable airspeed, high-altitude stalls, the automation vanishing without warning - rehearsed until the right response was there when it mattered.

They picked the skill they refused to let waste, and built the practice to keep it alive.

Hardly anyone handing work to AI is making that choice on purpose. The skills that weaken are simply whichever ones we happen to stop using first - not the ones we’d fight to keep if we stopped to think about it.

So the real question has nothing to do with speed. Of everything you’ve started handing over, which is the one skill you’d never want to lose?

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

In 1998, two music producers built one of the strangest vocal sounds anyone had heard, then lied about how they’d done it.

The song was Cher’s “Believe”. The effect was Auto-Tune, pushed so hard the voice turned robotic. Asked how they’d done it, Mark Taylor and Brian Rawling said they’d used a vocoder. They hadn’t. Back then, tuning a voice was still treated as cheating - a producer’s dirty little secret - so they kept the real method to themselves for years.

A decade later it was on hit after hit. What they’d hidden had become the sound everyone wanted.

That’s roughly where work is now.

Given time, we make our peace with what once felt like cheating. What grates is the in-between, when the work is obviously easier and we haven’t yet agreed that’s allowed.

So what are we really objecting to? I don’t think it’s the tool. It’s the missing effort.

Judging the sweat

Two Harvard researchers, Ryan Buell and Michael Norton, found that people value a service more when they can watch it working. A travel site that pauses to show you it's “searching hundreds of airlines” is preferred to one that returns the same results instantly.

They called it the labour illusion.

Companies now build in those delays on purpose, because visible effort makes us rate the result more highly. We are, it turns out, reassured by the sight of a machine pretending to work. We still judge the work itself. But we also judge the sweat we can see went into it, and we read that as proof of skill and care. And as proof of value.

Which is awkward, because AI removes the visible sweat.

Looking lazy

I saw this a couple of years ago, when helping teams bring AI into their work. I went in expecting to talk about tools, but the interesting conversations were about something else.

Most people held back from talking about their AI habits because being seen to use AI felt like admitting they couldn’t do the job without it.

It was a reputation thing. And they were right to expect the judgement.

Last year, researchers at Duke’s Fuqua School ran four experiments with more than four thousand people. Their participants rated the workers who’d used AI as lazier, less competent and less diligent than those who’d used other kinds of help, or none at all.

Of course, some AI use is lazy. But here, the workers didn't have to do anything worse to be judged that way - knowing AI was involved was enough. What they were really penalising was the missing effort.

But this just exposes a deeper, longer lasting problem: effort was only ever a proxy for value.

We can’t see the quality of someone’s thinking, so we look for the most visible signal we can - the hours, the late nights, the visible grind. But sitting at a desk all day doesn’t prove good work. It proves you sat at a desk all day.

And using AI neither makes you lazy nor especially innovative. AI might do in twenty minutes what once took two days, or burn an afternoon getting nowhere. Either way, the time it took tells you nothing about whether the thinking holds up or the problem got solved.

It didn’t matter, in the end, that Cher’s voice was auto-tuned. What mattered was that people loved the song.

In other words - forget how it was made. Is the work any good?

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

His name was Johannes Trithemius, and he ran the Benedictine house at Sponheim. He wrote a long, heartfelt case for monks copying books out by hand - the discipline of it, the devotion, the way a page of parchment might outlast the new printed sheets on their cheap paper.

Two years later, he had it printed.

It’s an easy thing to laugh at, and most people who tell the story stop at the joke. But Trithemius was no enemy of the press. He used it constantly. He had built his abbey’s library from 48 books to more than 2,000, buying printed volumes as fast as he could find them. Numerous editions of his own writing were published in his lifetime. He understood exactly what the machine could do.

But read his case another way, and it isn’t only about handwriting.

For centuries the monastery had been one of the places where knowledge was made, copied and kept. Printing didn’t close the monasteries - they were around long after - but it moved much of that work outside their walls, to printers and booksellers and universities. Trithemius reached happily for the new tool while defending the practices his own institution had been built on.

That’s the part worth carrying five centuries forward.

Smaller and flatter

A new paper from Hyunjin Kim at INSEAD and Rembrand Koning at Harvard looked at thousands of startups and found that the ones built around AI are organised differently. They’re about 25% smaller than comparable firms in the same industry, noticeably flatter, with proportionally fewer entry-level workers and managers - and yet they carry much the same valuations.

The usual reading is that AI has made these firms more efficient, which is true to a point. But what if there’s something more significant going on?

The shape of a company is really a response to a problem we’ve half-forgotten - how to move what people know around the place when moving it was expensive. Before the railways and the telegraph, information travelled no faster than a person could carry it, so coordinating a business across a continent was painfully slow. Once that changed, you got the big managerial hierarchy.

The org chart most of us treat as the natural shape of work is, to a large degree, the pipework we built when information was costly to move.

AI changes the maths.

Producing, moving and summarising information is now way cheaper. And when the cost of a thing falls through the floor, the structures we built to manage its scarcity start to look strange.

What the layers were for

Think about what a good chunk of middle management was actually for.

A manager sat above ten people, gathered what each of them was doing, compressed it, and passed it up to someone who did the same again. Remove that cost, and a question you’d once have sent up three layers is now one you can just ask.

I’m going to be careful not to wade into the “middle management is dead” debate - it’s a tired line, and probably wrong. Some of those layers existed mainly to move information up and down, and they’ll go. But others were doing work you don’t notice until it’s gone - bringing on the next lot, or holding a standard nobody wrote down.

The hard part is that you can’t always tell which was which until you’ve already cut it.

Still, there’s a real and hard question here, and most aren’t asking it. I sit in a lot of these conversations, and we almost always assume the company itself stays put. We ask where AI fits into the structure we’ve got. We rarely ask whether that structure was built around a constraint AI has just removed.

Which was Trithemius’s mistake - he embraced the machine and carried on as if his own institution would keep its shape.

So before you ask where AI fits into your organisation, ask the harder question.

How much of what you’ve built only exists because information used to be expensive - and what’s still standing when it isn’t?

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

In 1912, a country solicitor named Charles Dawson stood up at the Geological Society of London and announced he’d found the missing link between ape and human, dug out of a gravel pit in Sussex.

It was everything Britain wanted to hear - the earliest Englishman, no less.

They gave it a Latin name, Eoanthropus dawsoni, Dawson’s dawn man, and Arthur Smith Woodward, the Keeper of Geology at the Natural History Museum, put his name to it. Britain had its ancestor, and it was thrilled.

Problem is, it was a fake.

A medieval human skull, an orangutan’s jaw, and a few teeth filed down by hand to look human, then stained brown to look old. This wasn’t a mix-up in the field - someone had built it on a workbench.

And yet it sat in the textbooks for forty-one years.

Because the fake looked so exactly like the answer everyone wanted, it sent real science down the wrong path.

Piltdown had a big brain and an ape’s jaw, which propped up the prevalent theory that the brain evolved first. So when a young anatomist called Raymond Dart described the real thing in 1925 - small-brained, but walking upright - the big-brained fake made it easy to dismiss him for years.

Nobody had checked it, and everyone downstream paid the price.

Share Ollie on Work

Polished but hollow

By now you’ll have heard the word workslop. You’ll have been sent plenty, too, no doubt.

It was named last year by researchers at Stanford’s Social Media Lab and BetterUp, in Harvard Business Review - AI-generated work that looks right, but is often a load of rubbish. The sender looks productive, while the cost of actually doing the work properly moves to whoever has to work out what, if anything, the thing actually says.

In their survey, 41% of US desk workers said they’d been sent some in the past month - and 18% of those using AI at work cheerfully admitted to sending it themselves. The wasted, redone work adds up - for a 10,000-person company they put the bill at around $9 million a year.

And if anything it’s getting worse. A Workday study in January found that for every ten hours AI saves you, nearly four go straight back into checking, fixing and rewriting what it produced.

The reputational cost

Annoying as it is to wade through someone’s slop, the wasted hours aren’t the worst of it - at least from your point of view.

Because what workslop really costs you is your reputation.

The same research found that once someone receives it, they think less of the person who sent it - about half rated them less creative, capable and reliable than before. And nearly one in three said they’d rather not work with them again.

Smith Woodward found that out the hard way. He didn’t forge anything - he just put his name to something he hadn’t checked, and his is the name still attached to the whole embarrassing affair a century on.

Forward something an AI made without checking it, and you’re doing exactly what he did - vouching for it. Pass it on, and you’ve told everyone it’s good - so your name is on it when it turns out it isn’t.

When looking finished costs nothing, polish stops being a signal of anything. The signal that’s left is the opposite one - who actually checked, who read the thing properly and would stand behind it. Which means the person worth working with - teammate, boss, advisor, whoever - is the one who never hands you their mess to clean up.

Piltdown ended up being a handful of bones inflated into an entire human ancestor. Workslop in slow motion - and it lasted forty-one years not because one man forged it, but because everyone after him assumed someone else had checked.

So, before you forward the next report, deck or email an AI made you, ask the only question that still means anything. Would you put your name to it?

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

Around 2007, not long after he’d sold his first company, Uber’s co-founder Garrett Camp was watching Casino Royale when something caught his eye. Bond is tracking someone, and the chase is playing out on a phone screen - a dot moving across a map, updating in real time, closing in.

That image became the Uber map.

Of course, the map alone wasn’t the secret to Uber’s eventual success. It didn’t put more cars on the road, and it didn’t get a driver to you a second faster. Uber’s real problem - too few drivers, spread too thin across a city - would take years and a great deal of money to solve. But the map was the thing that got many of us hooked, because it went after something else entirely - not just waiting, but the frustration of not knowing how long we’ll wait.

Rory Sutherland calls this a psychological moonshot.

The map changed almost nothing about the wait itself, but it transformed how it felt - in your head, the equivalent of putting ten times as many cabs on the road.

Sold on speed

Now look at how we talk about AI.

Almost every tool is sold on speed - write the email faster, run the research faster, clear the inbox faster. But speed was never really the problem.

Last week, the economist Tyler Cowen - one of the calmer voices on AI - gave a keynote at the Sana summit in New York. The smart use of AI, he said, isn’t to read your WhatsApp messages faster. It’s to have an agent manage them for you, and tell you every couple of days when there’s something you actually care about.

The real cost of WhatsApp isn’t the minutes you spend reading it - it’s that the only way to know whether you need to look is to actually look.

And it isn’t only WhatsApp.

Maybe for you it’s refreshing a post to see if anyone’s commented, or a dashboard you check for tiny movements that won’t mean anything for a week, or the pull to be on Slack the second a message arrives in case it’s the one that matters. Whatever the screen, it’s the same low-level hum - is there anything new I need to see?

More to supervise

So it isn’t a new problem - but it’s one AI makes worse.

Earlier this year, BCG researchers writing in Harvard Business Review surveyed close to 1,500 people who use AI at work. The ones using it to take routine tasks off their plate reported 15% lower burnout than people using no AI at all. But the ones using it for what you might call ‘oversight’ - generating output, then reviewing it and deciding what to do with it - reported the opposite. More fatigue, more overload, and a 39% higher intention to quit. They called it “AI Brain Fry”.

So what’s really going on there?

What separated the two groups wasn’t how much AI they used, or which model, or how well they’d been trained. It was whether the AI was taking work off their plate, or handing them more to supervise.

Microsoft reckons the average employee is now interrupted every two minutes during core working hours - 275 times a day. But that only counts what comes at you. The other drain runs in the opposite direction - the watching and wondering, the mental tax you spend staying across everything, even when nothing’s really happening.

Which is why Cowen’s idea resonates.

An agent that watches for you - rather than you watching it - is the Uber map’s trick taken one step further. The map could only soothe the uncertainty, the agent actually removes it. You get both things at once - focus, and the reassurance that nothing’s slipping.

So forget how many minutes a tool saves you. The question worth asking - the one the Uber map answered - is whether it goes after the thing that's actually wearing you down.

It won't give you more hours. But does it let you stop watching?

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

In March 2016, the South Korean Go grandmaster Lee Sedol sat down in Seoul to play game four of his five-game match against AlphaGo, built by Google’s DeepMind. Go is an ancient board game, far harder for computers than chess. He was down 0-3.

For three days, DeepMind’s AI had been doing things no Go master had seen before.

Game two had given the world Move 37, a black stone placed in a mostly empty zone that the top Western commentator put on his demonstration board, then removed because surely it was a mistake. Wrong - it turned out to be decisive, and Lee resigned. He resigned the next game too.

In game four, with nothing left to lose, he played Move 78. A move so unorthodox the system that had been crushing him for three days began to err. AlphaGo flailed, started making nonsensical moves, and ultimately resigned.

So what broke AlphaGo?

For one game, Lee found a line it couldn’t read. It’s the move that made headlines - the night a human beat the machine - and close to the last time that happened.

But the move worth dwelling on is AlphaGo's.

DeepMind's system had been trained on a vast archive of human Go games. And yet Move 37 was a move almost no human would have played - the machine came up with it, not the man.

So much for the idea that invention is uniquely human.

Except that the AI could do all this only because Go is a closed world - fixed rules and only one way to win - where you can play yourself a billion times and end up better than everyone alive.

That maps neatly onto the repetitive, rule-bound parts of a lot of jobs. But it’s not true of everything we do - and the history of technology doesn’t suggest the future is anywhere near as predictable as some would have you believe.

More work, not less

Dan Shipper runs a small writing and tech company called Every, and published an essay last week called After Automation. His team has automated everything they can - Claude Code and OpenAI’s Codex across coding, writing, design, customer service.

You might guess from that setup that he’d be writing about how AI has thinned his team out, but it’s the opposite.

Despite all that automation, his team of thirty has more work to do than ever. Nobody’s been laid off - they’re still hiring writers, editors, engineers. The work has changed shape, but there’s way more of it.

When Shipper gets to the explanation, it sounds a lot like the Lee Sedol match.

AI gets very good very fast at the work that’s been written down. What took a senior engineer or a copywriter a day now arrives in seconds. What’s left for people is the other stuff - knowing which problem is worth solving, reading the situation in front of you, making the call when there’s no clean answer.

Share Ollie on Work

But across the whole economy - does any of this mean fewer jobs? Well, the early data cuts both ways

The youngest workers - 22 to 25 - in the jobs most exposed to AI, like software and customer service, have fallen about 16% behind comparable workers since 2024.

And yet, asked directly, more employers expect AI to grow entry-level hiring than to cut it.

And AI is only part of what’s going on. We’re still feeling the after-effects of Covid, with remote work in particular radically reshaping how younger people experience work. So it’s hard to say what’s driving what.

The work that isn’t a game

Ten years on, the Go match tells us a story the numbers currently can’t.

The work that goes first to the machine is the work that looks like Go - closed, scoreable, a right answer you can check. What’s left doesn’t.

And once it’s taken the Go-shaped parts, there’s more of the other, messier kind to do.

That messier kind is where Move 78 came from - the move Lee found by reading the board in front of him, not by grinding the same game a billion times.

There’s a line from the computer scientist Larry Tesler:

Intelligence is whatever machines haven’t done yet.

The moment a machine can do something, we stop calling it intelligence and call it software.

Move 37 was creative right up until AlphaGo played it. Then it was just Go.

Creativity is the edge that keeps moving - the part nobody’s defined yet. We innovate, we define the thing, the model learns it, and the creative work is already somewhere new.

Which is the reassuring part, oddly.

Every technology kills off some jobs and creates others, and - as Benedict Evans points out - you can never name the new ones in advance, because they don’t exist yet. In 1800 nobody could have told you their grandchild would make a living as a railway engineer.

On the evidence of the last two hundred years, the new work has always turned up - we just can’t see it coming.

That doesn’t make the transition painless - the long view is no comfort to the 22 year old inside that 16%. Both things are true at once.

The question worth sitting with is which parts of your job were never a game at all.

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

Former Google CEO, Eric Schmidt, stood in front of about ten thousand graduates at the University of Arizona on Sunday, and a few minutes in, the booing started.

Why?

He mentioned AI.

“I can hear you,” he said, over the noise, and kept going.

He named the graduates’ fear back to them - the machines are coming, the jobs are evaporating, you are inheriting a mess you did not make - and called it a rational fear.

But when someone offers you a seat on a rocket ship, he said, you do not ask which seat. You just get on.

It is tempting to file this under out-of-touch billionaire gets what was coming to him - but the speech was stranger than that. Because it was two speeches at once. One was a real defence of agency - their future was not yet written, theirs to shape. The other told them the rocket ship was already here, and the only move was to get on.

A stadium of twenty-two-year-olds heard only the second speech.

He was trying to encourage them - but then undid it in the same breath.

And Schmidt was not the only one.

Around the same time, a speaker at the University of Central Florida called AI the next industrial revolution and was drowned out by boos, while a music executive at Middle Tennessee State told graduates to “deal with it.”

What the booing was actually aimed at

So what’s really going on here?

Well, I wanted to know whether the booing was just a crowd doing what crowds do. So I went to the comments, on YouTube and Reddit.

Last year I did this on a larger scale: fifty thousand comments under three Diary of a CEO episodes about AI and work. What struck me then was the range. There was fear, but also curiosity, scepticism, people describing what they had already built. Enough variety to sort into five rough types.

Based on my research this week, there’s no longer so much diversity of opinion.

In fact, if you read the threads under the booing clips there’s one overriding feeling: anger (and in the billionaire tech CEO’s case, a large dollop of contempt).

This is seemingly not technophobia, by the way. Many of the angriest comments are written by people who use AI every day, often because their employer requires it.

Instead, the perception is that all three speakers were delivering the same instruction - this is settled, your future is fixed, and the only move left to you is your attitude. Be grateful. “Get on the rocket ship." That was what the boos were aimed at.

And the commenters flipped that rocket-ship line back on Schmidt - not passengers being offered a seat, but “the fuel”.

Zoom out, look beyond the comments, and you’ll see a similar story.

In recent polling, 70% of Americans say AI is moving too fast, and just 18% of young people say they feel hopeful about it.

Share Ollie on Work

There is also a real feeling of grievance in the case of the students.

A version of the deal these graduates were sold - do the degree, take on the debt, a career follows - has been rewritten while they were still holding up their end. One of the sharper observers called the feeling not fear but a sense of having been “double-crossed.”

You do not boo a weather forecast. You boo a promise you believe someone has broken.

Why uncertainty is the argument for agency, not against it

So where does that leave the advice?

Well, agency still matters - but not because it guarantees a safe, secure future. It matters precisely because nobody knows what is coming.

Because the evidence is genuinely split. Some of it shows AI hollowing out entry-level work, while some finds no clear sign of AI in the job numbers at all. They are looking at the same labour market and disagreeing about what’s happening.

And that is the point writ large.

If you knew you would be fine, what you did next would not matter. If you knew you were finished, it would not help. It is because no one knows, that what you do still counts - building range, trying things, treating your career as something to act on, not something that happens to you.

And frankly, it is not just young people.

If you are reading this, you may be doing two jobs at once. Leading other people through this, and wondering about your own working life. The answer to both is the same - stop waiting for the picture to get any clearer, because it will not. Act now.

Oh, and the doomer belief - that nothing they do matters - is even more dangerous. Believe it, and you stop trying altogether. And not trying is what will make it true.

Ten thousand graduates filed out of that stadium into an uncertain future. What happens to them now will not be decided by an ‘inspiring’ speech, or by their booing. It will be decided by what each of them does next. And, crucially, whether we give them the chance to learn and make mistakes the way we did.

That was always true - AI has just made it impossible to ignore.

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

In David Sedaris’s essay Stepping Out, he describes what happened when he got a Fitbit.

It started reasonably - ten thousand steps a day, a gentle tingle on his wrist when he hit the target. Then fifteen thousand. Then twenty-five thousand. By the end he’s doing sixty thousand steps a day, picking up litter along the roads of Sussex, and when his Fitbit dies he lasts five hours before ordering a replacement, hands shaking as he opens the box.

Sedaris started walking because he enjoyed it. Then the measurement replaced the enjoyment as the point.

He stopped asking “is this a good walk?” and started asking “is this enough steps?” The answer was never yes.

There’s a story in Brad Jacobs’s book that runs in the other direction.

Jacobs has built multiple billion-dollar-plus companies from scratch, but the best bit in his book is about something considerably more mundane.

His home AC broke during a heatwave. He called Steve, his repair guy, who came on a Sunday without hesitation.

Steve had originally worked for the company that installed the system - a large outfit that got bought by a private equity firm - not one of the good ones - which cut costs, stripped the service back, and resold the business.

Steve saw what was happening and left to set up on his own.

“How’s the old company doing?” Jacobs asked him.

“Terrible,” Steve said. “There’s no pride in their service. They’re giving the industry a bad name around here.”

Steve isn’t perfect. He’s misdiagnosed problems. He’s ordered the wrong part. But he owns up to it immediately and fixes it fast. Customers flock to him. The company he left is haemorrhaging clients.

The PE firm did what Sedaris did with his Fitbit.

They found the thing they could measure - cost - and optimised for it until they forgot what the business was actually there to do. The service got cheaper to run. It also got worse. The people who cared about doing good work left. The customers followed them.

Share Ollie on Work

The cost case

The same pattern is playing out in many AI conversations right now.

Cost reduction is the easiest business case to make.

“We’ll save X by automating Y” sounds like a no brainer.

“We’ll make our best people more effective” is harder to put a number on. And much harder to actually achieve.

So the cost case wins, and six months later you’ve got a cheaper operation that nobody - customers or employees - feels good about.

Microsoft’s latest Work Trend Index, based on 20,000 AI-using workers across ten countries, backs this up. The organisations getting value from AI aren’t just handing people tools or using them to reduce headcount. They’re changing the culture, management habits, and workflows around them.

Steve’s old company had the same technology available as Steve. The difference was everything around it.

The better question

So, the question to ask yourself is really quite simple:

When the AI’s been there a while, are your best people more valuable or less?

If the technology handles the routine so your best people can focus on the stuff that only they can do, you’re building Steve’s business.

If the people with experience and skill matter less than they used to, you’re building the bad version of the PE story.

And your best people will notice before your customers do. Steve didn’t leave because the PE firm fired him. He left because they made it impossible for him to do his job the way he knew it should be done. The people you most want to keep will do exactly the same thing.

Sedaris, at least, only ruined his own walks.

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

Last Sunday, Sabastian Sawe ran the London Marathon in 1 hour, 59 minutes and 30 seconds. He became the first person to break two hours in an official race, a barrier many people said wasn’t humanly possible.

His backstory is remarkable.

Sawe grew up in a village in Kenya’s Rift Valley without electricity, raised mostly by his grandmother. He only ran 5,000 metres for the first time in 2019 - by accident, because he arrived late to an athletics meet and it was the only race left. He was injured in January this year and only started focused preparation for London in February.

But the thing that interests me most about Sawe isn’t how he races. It’s how he trains.

His coach, Claudio Berardelli, described it last week: nine to ten days of hard training, then a break. Workouts combined with recovery, adjusted depending on how Sawe’s body responds.

On hard days he pushes hard. On recovery days he actually recovers.

So when race day comes, he has everything available.

Your brain is a battery

Obviously we’re not running marathons. But the contrast is pretty obvious.

I’m guessing you start the average week with your calendar almost full. With no margin for anything unexpected. Certainly no time to reflect on where to focus your energy, let alone recover.

Because there’s nothing left, the thing that gets sacrificed is always the same: the thinking, the strategy, the stuff that never feels urgent but ends up mattering most.

It's not a willpower problem, and Samuel Marcora, a sports scientist, showed why.

He took two groups of people, of a similar age and same fitness level. One group watched a documentary for 90 minutes. The other spent 90 minutes doing a cognitively demanding task.

Then both groups did the same cycling endurance test. The mentally tired group gave up 15% earlier.

Now consider the typical working day for you and your team. Emails from the moment you wake up, context-switching all morning, back-to-back meetings - by mid-afternoon your brain is done. And you’ve probably still got a few hours of work to go.

We’ve only got about four to five hours a day of our best thinking, so why aren’t we pickier about what gets them?

Share Ollie on Work

Focus, not volume

Back to running for a moment.

When Roger Bannister broke the four-minute mile in 1954, he was training a fraction of the volume of his main rival - John Landy was running roughly five times more distance per week.

But Bannister - who was also a medical student and couldn’t train full-time - had figured out that the bottleneck to him performing at his best wasn’t endurance. It was his body’s ability to process oxygen and lactic acid at exactly the pace needed for four minutes. So all of his training was built around that one thing.

As you know, he broke the record first.

David Epstein and Cal Newport were chewing over this exact idea on Newport’s podcast this week. Epstein’s new book Inside the Box is full of examples of the benefits of constraints.

He tells the story of a company that makes custom gearboxes - every one unique and built to order. They were taking a year to deliver each one and when they looked at why, they found, on average, people were switching between tasks more than 50 times a day.

To fix the problem, they introduced one rule, one constraint: you can’t start a new design until you’ve finished the current one.

Within months, three times as many designs were going out the door. Delivery time dropped from a year to two months.

Again it’s easy to see a parallel here with how many teams are operating right now - and in this case, it’s a problem made worse by the ease with which AI can help you write a report, generate code, or create marketing campaigns.

Just because you can, doesn’t mean you should.

Race days and training days

Sabastian Sawe doesn’t run fewer miles than anyone else (200km a week seems pretty hefty!), but he and Berardelli are deliberate about which days are for pushing and which are for recovering.

They designed constraints.

I’ve been writing for the past few weeks about how companies should think about AI. But none of it matters if nobody has the headspace to act on any of it. Every leader I know agrees they should be experimenting more. Almost none of them have a single unprotected hour in their week to do it.

In this case, the biggest bottleneck isn’t technology, it’s their time.

Very unlikely you're going to break a world record this week, but you might need an hour to think clearly about something important. Your calendar won't give you that by default. You'll have to design the constraint yourself.

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

In Jiro Ono’s sushi restaurant in Tokyo, the apprenticeship starts with cleaning.

For months, that’s all you do. You clean, you watch, you say yes. After about four months, you’re allowed to touch fish. After years, you might stand behind the counter.

One apprentice made the same egg dish over two hundred times before Jiro deemed it acceptable. When he finally got the nod, he cried.

Around a decade. That’s how long it takes before an apprentice is considered ready to make sushi for a customer.

Rice is the thing that reveals why. It accounts for about 80% of what makes great sushi, and there’s no recipe. The moisture content changes with the season, the age of the grain, the weather that morning. The only way to learn it is to do it, badly, for years, until your hands know what your brain can’t put into words.

The apprenticeship problem

The most obvious short-term impact of AI on jobs is on younger people.

In the US, recent graduate unemployment hit its highest level in over a decade last year. In the UK, youth unemployment just overtook the EU average for the first time on record. Entry-level hiring is slowing in many parts of the labour market. AI adoption at work is being led not by the youngest employees but by higher-earning, more experienced workers.

I speak to dozens of businesses every month and I’m hearing the same thing - a feeling, from leadership at least, that AI can now do a lot of the work we’d typically give to people just starting out.

They’re not entirely wrong.

AI is very good at producing work that looks like it came from someone who knows what they’re doing. And experienced people can usually tell when it’s actually good - that’s why they’re comfortable using it. They’ve spent years developing the instinct that lets them spot when something’s off.

The problem is where that instinct came from.

It came from doing the work badly for years. From producing stuff, being told it wasn’t good enough. Developing a feel for when something isn’t quite right - the kind of thing you can’t get from reading about it. The reps.

AI won’t tell you your rice is sh*t. It’ll tell you it’s excellent and move on to the next question.

And if we stop letting juniors do that work, the next generation won’t have the instinct that today’s leaders can rely on.

Share Ollie on Work

The value of a fresh pair of eyes

That’s the long-term cost. There’s a short-term one too.

Juniors aren’t just trainees-in-waiting. They bring something to a team that experienced people can’t, because they don’t carry the baggage of how things used to be done. They ask the questions you’ve stopped asking. They notice the things you’ve trained yourself not to see.

Removing them from teams doesn’t just delay the problem. It makes today’s work worse too.

So before you automate a task that a junior used to own, ask what they were learning by doing it. If the answer is genuinely nothing much, then fine - automate it. But if they were developing an instinct for what good looks like - the slippery idea of ‘taste’ that’s become so popular in the AI era - you need to replace that learning with something more deliberate.

So who’s learning to make the rice?

Thanks for reading.

Ollie

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

In 1998, thirty thousand Americans were asked two questions: how much stress have you experienced this year? And do you believe stress is harmful to your health?

Eight years later, the researchers checked who had died.

High stress increased the risk of dying by 43 percent. But only for people who believed stress was harming them. People who reported high stress but didn’t see it as harmful had the lowest death risk in the entire study - lower even than people who reported very little stress at all.

The researchers estimated that over those eight years, 182,000 Americans may have died prematurely. Not from stress. From believing stress was bad for them.

That story comes from Kelly McGonigal’s book The Upside of Stress. It came back to me after reading something Christine Armstrong wrote the other day.

Christine’s old boss used to tell her: “never be fragile.” Great advice, she said, but rather out of fashion nowadays, quoting people avoiding personal questions at work for fear of causing offence. And colleagues who don’t know each other well enough to trust each other, so they say as little as possible.

She’s right, of course.

We’ve spent the last five years optimising work for individual comfort - remote, async, fewer awkward conversations, less proximity - and somewhere along the way we’ve confused ‘safety’ with avoidance.

The intention was good. But the result is that we’ve quietly built organisations full of people who aren’t used to being uncomfortable anymore.

Share Ollie on Work

The worst possible timing

AI demands exactly the qualities we’ve been letting atrophy.

On an individual level, this could be the acknowledgement that some of the skills you’ve spent years developing may not be as useful any more - and the willingness to pick up a tool you don’t understand and fail with it for an hour to learn something new.

At a team level, it’s most likely something like having the confidence to say “this process we’ve been running for three years is a waste of time” and deal with the silence that follows. Or the patience and courage to try something, show it to someone, and hear that it doesn’t work.

All of it requires a tolerance for friction that we’ve spent five years systematically removing.

Nassim Taleb has a word for systems that get stronger under stress: antifragile.

Fragile breaks. Robust survives. Antifragile improves. Most companies aren’t any of the three. They’re just...comfortable.

The team that never disagrees, the leader who never hears bad news, the organisation where everything runs smoothly right up until the moment it doesn’t. That’s comfortable.

An antifragile team has difficult conversations and comes out more aligned, not less.

Start something uncomfortable

McGonigal’s point is simple: stress isn’t the enemy. The belief that stress is the enemy is the enemy.

If you’ve been putting off getting to grips with AI, the discomfort of not knowing what you’re doing isn’t a sign you’re not ready. It’s what getting ready feels like. Start before you have a plan.

If there’s a conversation you’ve been avoiding - about your career, about how the team works, about something that’s broken - the willingness to raise it is the thing. You don’t need the answer before you walk in.

And if you can see problems in how your company operates but you’ve been keeping quiet because it’s easier - the ability to say what needs to change, and bring people with you while you do it, is one of the few skills that becomes more valuable the more AI can do.

Machines are very good at optimising what exists. They are useless at telling you it shouldn’t exist.

Thanks for reading.

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here.

You know that route you take every morning without thinking?

The one where your legs just carry you and your brain is somewhere else entirely?

You probably couldn’t tell me when you last considered whether there’s a better way to go. It works. You stick with it.

On the evening of February 4th, 2014, millions of London commuters had that choice taken away from them. Tube workers walked out on a 48-hour strike. 171 out of 270 stations closed.

Most people were furious. But something unexpected happened.

Researchers from Cambridge and Oxford got hold of anonymised Oyster card data from around 18,000 regular commuters - people whose journeys were completely predictable in the weeks before the strike. The same station in and out, every single morning.

During the strike, half of them started and ended their journey at a different station. Forced out of their routine, they improvised.

Here’s the interesting bit. After the strike ended and every station reopened, about 5% of the most habitual commuters permanently switched their route. They’d accidentally discovered something better - faster, less crowded, more pleasant - that had been available to them all along.

They’d been commuting the same way for years. Some of them, presumably, for decades. It took someone else shutting down their usual option to make them notice.

Herbert Simon, the Nobel-winning economist, had a word for this: satisficing.

It means settling on a solution that’s good enough rather than continuing to search for the best one. Most of the time, that’s the right strategy. Life is too short and too complex to optimise everything.

The problem is when “good enough” calcifies into “the only way we do things.” When nobody questions it anymore. Not because the process is bad. Because it’s functional. It works. Nobody’s complained. So nobody’s bothered doing anything to improve it either.

Share Ollie on Work

The belts and shafts problem

Last week’s piece was about understanding what a job really involves before you automate it.

This is the harder question underneath it: should the job exist in its current form at all?

In the 1890s, factories started replacing their steam engines with electric motors. On paper it should have changed everything. In practice, almost nothing changed for decades. Productivity barely moved.

Because the factories were designed around steam. One enormous engine in the basement, driving everything above it through a complicated system of belts, shafts and pulleys. When electric motors arrived, owners ripped out the steam engine and bolted in an electric one.

And the problem was they stuck with the same layout, the same belt and the same shafts.

The new technology was doing the old job.

The real gains came much later, when someone worked out you could give every machine its own small motor. That let you redesign the factory floor entirely - placing machines where they made sense, not just where the belts reached.

Productivity exploded.

Most AI projects right now are in the belts and shafts phase. Drop an AI tool into an existing workflow, get a marginal improvement, declare victory. But the process itself - designed for an era when humans did everything by hand - goes completely unexamined.

You’ve made a slightly quicker commute but on the wrong route entirely.

Manufacture your own tube strike

Now, ideally, you don’t want to wait for a crisis to rethink how you work. But you can simulate the conditions for one?

Pick one workflow that nobody's questioned for years.

Get the people who do the work in a room.

Map what actually happens.

Then ask: if we were starting this today - with the tools, the team, and the information we have now - would it exist, and if so, would we build it this way?

You might find that you would. Fine. At least you’ve pressure-tested it.

But in many cases, you’ll find a process - a way of doing things - that looks a lot like a commuter who’s been taking the Circle Line for fifteen years when there’s a pleasant walk from Liverpool Street that would have been better all along.

Thanks for reading,

Ollie

Share

Ollie on Work is a weekly newsletter about what I’m learning from building with AI, advising leadership teams, and trying to bridge the gap between what technology can do and how businesses actually work. If someone forwarded this to you, you can subscribe here:

Subscribe now

New Jersey is the last state in America where it’s illegal to pump your own gas.

Every station has an attendant. Someone comes to your window, fills your tank, maybe checks your oil. It’s been this way since 1949. Multiple attempts to change the law have failed - not because of some powerful gas station lobby, but because 73% of New Jersey residents prefer it this way.

In a country that self-serviced everything else decades ago - checkouts, check-ins, banking, boarding passes - the people of New Jersey are actively choosing to pay a bit more to have a human do something they could easily do themselves.

They’re not stupid. They’re not lazy. They just understand something that most businesses have forgotten.

The doorman fallacy

There’s a concept I keep coming back to, popularised by Rory Sutherland. He calls it the doorman fallacy. It goes like this.

A consultant visits a hotel. Sees a doorman. Defines the doorman’s job as “opening the door.” Replaces him with an automatic door. Claims the cost saving. Leaves.

Five years later, what the hotel can charge per night has dropped. There are people sleeping in the entrance. The most loyal guests have quietly moved on. But the consultant who made the recommendation is long gone - and nobody connects the decline to the decision.

Because the doorman was never just opening the door. He was hailing taxis. Recognising regulars. Providing security. Signalling status. Making guests feel like they’d arrived somewhere worth arriving at. None of that was in the job description. None of it appeared on a spreadsheet. All of it was real.

The saving was easy to measure. The value destruction was invisible - until it wasn’t.

Share

Where the conversation goes wrong

I’ve been in a lot of these conversations recently. The ones where a team sits down to figure out where AI fits into what they do. And the thing that strikes me every time is how early the conversation goes wrong.

Not because people are rushing to automate things they shouldn’t. Most aren’t anywhere near that stage yet. The problem is more basic than that. People talk about “workflows” and “processes” without ever having properly defined what those actually are. The word gets thrown around in every meeting, every strategy document - but when you ask someone to describe, step by step, what actually happens in a process they own, most people can’t do it. Not because they’re incompetent. Because nobody ever asked them to.

So the first thing I tend to do isn’t talk about AI at all. It’s much simpler. I ask people to describe what they actually do.

I was in a session recently where a team was looking at how they handle incoming data from external partners. On paper, it’s straightforward: data comes in, gets checked, gets logged. The kind of thing that looks perfect for automation.

But when I asked the person who actually does this work what their day really looks like, a completely different picture emerged. They’re not just processing data. They’re spotting when something looks off before it becomes a problem. They’re maintaining a relationship with the people sending it. They know which partners are reliable, which ones cut corners, and which ones need chasing before a deadline. They’re an early warning system dressed up as an admin function.

That conversation alone - before anyone mentions a single tool - changes everything. It changes what you’d choose to automate, and how.

Automate the data processing? Absolutely. But if you remove the person without understanding what they were really doing, you don’t just lose an employee. You lose the early warning system. And you won’t notice until something blows up.

Among the companies that are actually deploying AI - and honestly, it’s still not that many - the default is to define a role by its most visible task, automate that task, and claim the saving. Without ever asking what invisible value gets destroyed in the process.

The better question isn’t “which roles can we automate?” It’s “what’s actually happening here that we haven’t bothered to articulate yet?”

And even once you’ve understood the full picture, there’s a harder question underneath it - whether the process should exist in its current form at all. But that’s for next week.

One thing you can do this week

Before you automate anything, go and sit with the person who does the work. Not their manager. Them.

Ask them what they actually do all day - not what the process document says. Listen for the things that don’t have names. The judgment calls. The relationship maintenance. The pattern recognition that only comes from doing the same thing, with the same people, for years.

That’s where the real value lives. And that’s what no spreadsheet will ever show you.

Share

Subscribe now

I was in Paris last weekend with my wife. We spent most of the time doing what the French do best - sitting around eating. Slowly, properly.

Three courses at lunch because why not. A coffee that lasted an hour because nobody was in a rush.

It reminded me of a study I read a few years ago by a researcher called Ashley Whillans. She and her colleague Romain Cadario surveyed over 10,000 people about how they eat, and found something interesting, but probably not surprising for anyone who’s into national stereotypes.

The French spend more time than anyone eating their meals. By which I mean, sitting down, tasting the food. Being present for it.

Americans, on the other hand, spend more time than anyone choosing what to eat - researching restaurants, scrolling through delivery apps, reading reviews, debating menus, comparing options - but much less actually eating it.

Same amount of time committed to the overall activity. Completely different allocation.

And guess what - the French reported significantly more satisfaction from the experience and less stress around it.

The Americans weren’t making better choices with all that research. They were just postponing the moment of commitment.

I realise I’m about to do the thing where someone connects a holiday anecdote to a business insight, and I apologise in advance, but… I’ve been thinking about that study a lot since I got back, because I keep seeing the same pattern in a completely different context.

The menu problem

A leadership team decides AI is important, so they start evaluating.

They read reports. They book demos. They run a comparison spreadsheet. They attend a webinar. They set up a working group to assess the options. Someone writes a proposal. The proposal needs sign-off. The sign-off requires another meeting. Someone raises a procurement question. Legal gets involved.

Three months later, they haven’t deployed a single tool into a single workflow. But they’ve spent an enormous amount of time and energy on the project, so it feels like progress.

This is the menu problem. All the time goes into choosing. None of it goes into eating.

Meanwhile, down the corridor, someone on the team has been quietly using ChatGPT or Claude every day for the last year. They’ve figured out which tasks it’s brilliant at and which ones it’s useless for. They’ve built little workflows that save them hours a week. They’ve made mistakes and learned from them.

They have more practical knowledge about AI than the entire working group - not because they’re smarter, but because they started.

Share

Why we get stuck

Because there are a few reasons companies get trapped in evaluation mode. The problem is that most of them feel perfectly rational from the inside.

Fear of picking wrong. The AI landscape is moving so fast that whatever you choose today might be obsolete in six months. So you wait. The thing is, waiting doesn’t reduce the risk - it just guarantees you learn nothing while you’re waiting.

The search for the perfect use case. Leaders want to find the highest-impact application before they commit. This sounds sensible but it’s backwards. You don’t know where AI will have the most impact in your business until you’ve used it in a few places and seen what happens. The best use cases reveal themselves through use, not through analysis.

Consensus as a delay mechanism. Getting everyone aligned before moving forward feels responsible. In practice, it means the most cautious person in the room sets the pace for the whole organisation. The French don’t hold a committee meeting before ordering lunch.

The Deliveroo data

Sticking with the food theme, there’s a nice piece of research that backs this up from a different angle.

A team at the Max Planck Institute partnered with Deliveroo and analysed 1.6 million food orders from 195,000 customers across 197 cities.

They found what you’d expect: when people tried a new restaurant, the average rating was lower than when they stuck with an old favourite. 4.26 versus 4.52. Exploration has a short-term cost.

But over time, the more people explored - the more new restaurants they tried - the higher their average satisfaction climbed. Because they kept discovering better options and dropping the duds.

The people who stuck exclusively with what they knew plateaued. The ones who kept trying new things kept improving.

And interestingly, people were actually more attracted to restaurants with fewer reviews. Not less. Fewer reviews meant more uncertainty - and uncertainty meant there might be an undiscovered gem that everyone else had missed.

The lesson isn’t “be reckless.” It’s that the discomfort of trying something new is the price of finding something better. And the longer you spend reading reviews instead of ordering, the longer you wait to start climbing.

Just pick one and start

If your team has been evaluating AI tools for more than a month without committing to one, stop.

Pick the best available option. Not the perfect one - the best available one. Give it to a small team. Point it at one real workflow - not a sandbox, not a test environment - a real thing with real stakes.

You’ll learn more in the first week of actual use than you learned in three months of demos. Some of what you learn will be that it doesn’t work the way you expected. Good. That’s information you couldn’t have got from a comparison spreadsheet.

Order something. Eat it. Then decide if you want to order it again.

The ROI isn’t in the selection. It’s in the savouring.

See you next week.

Ollie

Share

“Ollie ran an energetic, fun, informative and practical session with lots of takeaways for how to make the workplace more effective and efficient. His session was a highlight of the collaborative leadership course.”
Shân Wareing, Vice Chancellor & CEO, Middlesex University

This is the final part of From Zero to World-Class AI Manager. If you’d like to discuss how I can help your team, book a free, no-obligation call HERE.

Share Ollie on Work

You’ve spent seven weeks building AI capability.

But personal productivity gains have a shelf life. If it stays with you, it dies when you leave - or just fades when you get busy. The real value comes from making it stick across a team.

This week is about how to share what you’ve learned without being that person who won’t shut up about AI. Oh, and how to know if any of this is actually working.

Why people hide their AI wins

Remember, there’s a reason your colleagues aren’t sharing what’s working.

Fear of being given more work. “Oh great, you finished early - here’s more.” Worry they’ll look like they’re cheating. Concern about judgment from people who haven’t started yet.

The fix isn’t mandating adoption. It’s making it safe to experiment, and creating lightweight ways to share.

How to share without preaching

Don’t: “Everyone should use AI!”

Do: “This saved me 3 hours on the quarterly report. Here’s how, if you’re interested.”

The difference matters. The first sounds like an edict. The second is an invitation. People adopt things when they see someone like them getting results - not when they’re told to.

Three ways to build momentum

1. Create a Slack or Teams channel for experiments

Not a channel where you broadcast announcements. A place where anyone can post “I tried this, here’s what happened.” Low stakes. Failures welcome. The weird output that made you laugh gets as much engagement as the polished workflow.

Build a one-page guide

This is your call to action from this series. Create a simple doc with:

• 3 use cases that actually worked for you

• The prompts you used

• What to watch out for (where AI gets it wrong)

• How long it took to get good at it

Share it with 2-3 people who’d benefit. Ask: “Would you try this?”

That doc becomes the start of a shared prompt library. Others add to it. It becomes institutional knowledge that gets better over time.

3. Run a hackathon (the non-technical kind)

Get people from different teams together. Give them a real scenario - a client brief to respond to, a messy situation to untangle, a process to improve. A few hours. No coding required.

The goal isn’t to build something technically impressive. It’s to see what’s possible when you combine different perspectives with AI capability. And to make experimentation feel normal.

This is a great idea for your next away day, by the way.

Knowing if it’s working

Measuring this stuff is never going to be perfect. And let’s be honest, if you’re now using AI a bunch more, you’re probably not actually working less. That’s not really the point.

What’s more likely is that you’re doing more. Or doing different work, Or doing the same work but better. That’s harder to measure than “I saved 3 hours.”

But here’s one simple way to get a rough sense of time saved.

The mini audit:

For one week, spend five minutes at the end of each day noting roughly how long you spent on different types of tasks. Not just AI - everything in your work day (and maybe even your personal life). Don’t overthink it - just top-line categories and rough time.

Do the same thing a month from now. Compare.

You’ll see where time has shifted.

And you’ll notice something else: you become more intentional about what you’re actually spending time on. That’s valuable in itself.

What you’ve learned

Over the past eight weeks, you’ve learned what a workflow actually means, how to talk to AI like a colleague, how to fix your meetings, how to set up your AI to work better and faster, how to research without disappearing down rabbit holes, and how to use AI to think more clearly.

There’s plenty more to cover - agents, deeper automation, workflows that run without you. I’d love to hear what you want to read more of in 2026. Reply and let me know what would be most useful.

Thanks for following along.

Ollie

Share

Catch up on the previous 7 weeks of From Zero to World-Class AI Manager here:

• Week 1: You can’t lead an AI transformation you don’t understand

• Week 2: Talk to AI like a colleague

• Week 3: Fix your meetings before you automate them

• Week 4: Setting up your AI to help you work better and faster

• Week 5: AI research that actually works

• Week 6: How I built a product solo using AI research

• Week 7: Using AI to think

“After initially meeting Ollie at the IAB leadership summit, we were thrilled to be able to invite him to be the external speaker at our Coty summer conference.

Ollie’s session was captivating for the whole audience, as it not only broadly humanised some of the conversations bubbling around AI, but was also packed with practical applications that our team left eager to explore and implement.

Ollie’s skill is that he can make complex concepts easy to understand, and he inspired us all with real-world examples. If you’re looking for an engaging speaker to ignite creativity and knowledge amongst your teams, look no further!”

Susie Thompson, Senior Director Media & Communications at Coty

I’m working with leaders to take lessons like these into their organisations - through workshops, advisory, and speaking. If you’d like to discuss how I can help your team, book a free, no-obligation call HERE.

This week is about using AI to actually think – getting clarity, challenging yourself, and knowing when to stop

1. Learning new domains fast

I travel a lot for work. Different industries, different clients, different problems.

The challenge isn’t finding information. It’s building enough understanding to have a point of view - fast.

Before high-stakes meetings, I don’t just ask “tell me about this company.” I ask AI to help me understand their world well enough to ask smart questions and connect what I do to what they care about.

I have a meeting with [person/company] in [X days].

Context:
- What they do: [brief description]
- What’s happening in their industry: [any relevant trends]
- What I’m trying to achieve: [your goal for the meeting]

Help me:
1. Understand what people in their position actually care about day-to-day
2. Identify 3 angles I can take into the conversation that show I understand their world
3. Suggest 5 questions I could ask that would make me sound like an informed peer, not a tourist

The goal isn’t to become an expert. It’s to move from “I’ve heard of this” to “I have a perspective worth sharing.”

Other uses: Preparing for board presentations in unfamiliar sectors, or understanding new technology before you decide whether to invest time.

2. Untangling messy situations

Sometimes I can’t think clearly because there’s too much going on.

Projects competing for attention. Decisions that all feel urgent. Can’t see the wood for the trees.

This is about organising what you already know but can’t see clearly.

When my brain feels full, I dump everything out and ask AI to help me see it. Recently I pasted in a list that included speaking work, advisory projects, a product I’m building, and content - then asked it to help me separate what’s actually important from what just feels urgent.

I’m overloaded and can’t see clearly. Here’s everything on my plate:
[bullet list - projects, commitments, decisions, worries]

1. Restate my situation in no more than 10 bullet points
2. Separate those into: FACTS, ASSUMPTIONS, and FEARS
3. Given only the facts and my goal of [briefly state], what are the 3 most important priorities for the next 90 days?
4. For each, give me one action I can take this week

AI’s job here isn’t to decide for you. It’s to organise the mess so you can.

Other uses: Preparing for a difficult conversation. Working out what’s actually bothering you about a project. Separating real constraints from imagined ones.

Share Ollie on Work

3. Finding the real question

This is different from untangling. That’s about clarity. This is about realising you’re solving the wrong problem.

Often I’ll ask AI something and realise halfway through that I’m asking the wrong question.

A client says “we need help with X” - but the actual problem is Y. I think I’m stuck on a decision, but really I’m avoiding a harder question underneath it.

Here’s the question I think I’m asking: [surface question]

Context: [why this matters, what’s at stake]

1. What do you think the *real* question is that I’m trying to answer? Give me 3 possibilities.
2. For each, explain what decision would change if that were the real question.
3. Which one do you think matters most and why?
4. What should I ask myself to make sure I’m focused on the right problem?

Half the time, the breakthrough isn’t finding the answer - it’s realising you’ve been asking the wrong question.

Other uses: When you’re stuck and don’t know why. When a client’s brief doesn’t quite make sense. When you keep circling the same problem without progress.

4. Stress-testing ideas

I have a tendency to fall in love with my own ideas. Most people do.

So before I commit serious time or money, I ask AI to attack. Not “is this good?” but “why wouldn’t this work?”

Here’s my idea: [describe in 5 sentences or less]

Assume you’re a sceptical investor who’s seen a hundred pitches like this.

1. List the top 5 reasons this is weak, risky, or unlikely to work. Be harsh.
2. What are 3 warning signs that would tell me to kill this idea quickly?
3. If you had to keep the core insight but radically change the execution, what would you do?
4. Based on this critique - is this worth testing at all? If so, what’s the smallest test that would prove or kill it?

The value isn’t in AI being right. It’s in forcing you to defend your thinking before you’ve invested too much to change course.

Other uses: Before you hire someone you’re excited about. Before you announce something publicly. Before you invest serious money in something.

5. The rabbit hole warning

I have a tendency to over-research.

Sometimes it’s trivial. On Sunday, I spent too long working out whether 125 or 140 degrees was the “correct” temperature for slow-cooking beef. The stakes were low. The cost of being slightly wrong was basically zero. But I kept pushing for the “right” answer.

Sometimes it’s more dangerous - like spending weeks refining a product idea when five customer conversations would tell me more, faster, and with more accuracy. Using research as a sophisticated form of procrastination.

Now I try to catch myself:

I’ve been thinking about [topic] for [time].

Am I overthinking this? What’s the minimum I need to know before I act?

If AI tells me I’m in a rabbit hole, I try to listen.

Other uses: Any time you’ve been “researching” something for more than an hour. When you’re on your third draft of something that was fine on the first.

6. Knowing when to stop

The flip side of the rabbit hole - knowing when you already have the answer.

I’ve asked AI the same strategic questions multiple times. “Where should I focus? Which project should I prioritise?” The answers don’t change because there’s no new information. At that point I’m using AI to avoid committing.

You don’t need a prompt for this. You need a pen and paper.

Before you start, write down:

• What question am I trying to answer?

• What would “good enough” look like?

• How will I know when I’m done?

Just like going down a YouTube rabbit hole or doomscrolling on Instagram, AI can suck you in. And just because it seems smart doesn’t mean it’s any better way to waste your time.

The discipline isn’t just in thinking well. It’s in knowing when to stop thinking and act.

The point

AI is a powerful thinking partner - but only if you use it that way.

Most people prompt once, get an answer, and move on. The real value is in the back-and-forth. Using AI to learn faster, challenge harder, see clearer.

And knowing when to close the chat and just do the thing.

Thanks for reading.

Ollie

Share

Could I use AI to create a product on my own that would have previously required a whole team?

The problem

Understanding how and when to monetise content is tough.

Not that there’s a shortage of tools that can help you do it. In fact, media and entertainment businesses are drowning in monetisation tools. Hundreds of vendors but no real framework for making decisions.

I wanted to help solve this problem.

The method

Three types of AI research, looped until the right product emerged.

1. Synthesise interview insights

I sat down with 15 people in the industry for wide-ranging conversations on what’s happening right now, how their businesses have responded, and their perspectives on what was likely to happen in the near future.

I fed the interview transcripts into ChatGPT:

I’ve conducted 15 interviews about subscriptions and advertising. [Attach transcripts]

Synthesise key patterns:
- What challenges come up repeatedly
- Where people disagree or have different approaches
- Gaps between what they want and what tools provide
- Non-obvious insights about what’s actually working

Format: Themes with supporting quotes, then strategic implications.

The analysis picked up on something I was hearing.

When it comes to the technology driving content and monetisation decisions, people’s problems weren’t necessarily about how to use specific products. The challenge seemed to be connecting tool capabilities - whether analytics, identity and access, or subscription management - to their business model and audience behaviour.

In other words, it wasn’t a tools problem. It was a decision-making problem.

People had plenty of options but no framework for figuring out which ones actually fit how their business worked and how their audience behaved.

That told me what I needed to research next.

(Examples of other uses for synthesising call transcripts to generate insights: L&D teams synthesising training feedback to redesign programmes. Product managers finding patterns across user research. Marketing analysing customer feedback.)

Share

2. Track industry trends

I now had a sense of what was going on, but fifteen people is hardly a significant sample size. So, the next step was to zoom out and use AI Deep Research to generate analysis of the wider market.

Research: Current trends in media monetisation for content businesses

Focus on:
- Shifts in business models (what’s gaining/losing ground)
- New approaches people are testing
- Failed experiments and why they didn’t work
- Emerging patterns in audience behaviour

Prioritise recent news, operator posts, case studies from 2024-2025

This surfaced patterns I wouldn’t have found doing manual research alone - mainly because of the breadth of sources.

• AI bot traffic was inflating dashboards while revenue stayed flat - metrics that looked like growth but weren’t

• The biggest blocker wasn’t content or pricing - it was the infrastructure. When simple changes require engineering time, you’re maintaining systems, not running a business

• Everyone talking about combining revenue streams while the underlying systems can’t share data

Cross-referencing these findings with my interviews showed what people were saying versus what was actually happening.

I had the big picture. Now I needed the details.

(Examples of other uses for Deep Research to validate customer insights: Product marketing teams tracking competitor positioning shifts. Strategists researching how others handled transitions you’re planning. Agency teams mapping client pain points to shape their pitch.)

3. Deep dive vendors

I now knew the trends and the problems. Time to research the actual vendors.

Deep Research: [Vendor name] - comprehensive profile

Include:
- Core capabilities vs marketing claims
- Pricing structure and hidden costs
- Integration complexity
- Who it actually works for (use cases, company size)
- Common complaints from users

Format: TL;DR, strengths/limitations, implementation reality check

I ran this for every major vendor whose name kept coming up in my research - companies like Adobe Analytics, Auth0, Google Analytics and Paddle.

I ended up with detailed research reports on each company, which I knew I needed to validate.

First with multiple AI platforms, cross-referencing outputs.

Then, importantly, I got real people back in the loop - asking for their experiences implementing the systems, building on the platforms, assessing the support they’d received along the way.

Now I had everything I needed to start building.

(Examples of other uses for Deep Research on specific products: Procurement teams comparing SaaS vendors before negotiations. HR leads assessing applicant tracking systems. Operations comparing logistics providers.)

The result

This process took me a few months - mainly because of the human conversations. But I know from experience it would have been impossible in the past without a full team working flat out.

In fact, until I worked with a developer to build the Monetization Works website - which I’d designed myself using Lovable, by the way - the only support I had was from AI.

Monetization Works is now live here.

You can find vendor profiles on companies like Stripe. And you can read comparisons of products in the same category like Recurly vs Zuora.

The profiles don’t read like feature lists because they’re not based on marketing material. They’re based on what people actually experience. Where tools fail. Which trade-offs matter.

And the added bonus - I designed the content so it would rank for searches people were actually making using LLMs like ChatGPT and Gemini.

Someone searching for “Stripe alternatives” or “Recurly vs Zuora” finds Monetization Works because the content answers strategic questions, not surface comparisons.

That’s Answer Engine Optimisation (AEO) - I built it that way from the start.

The research creates the product. The product creates its own distribution.

The point

Strategic research used to require a team because one person couldn’t move fast enough through the cycles.

AI compresses those cycles from weeks to days. You can operate solo at a level of depth that used to need multiple people.

But here’s the trap - AI will confidently tell you everything sounds great. It’ll validate every insight, support every hypothesis. It’s easy to build an entire product on research that sounds smart but doesn’t reflect reality.

The discipline is validation.

Checking with people who actually use these tools. Who’ve lived these problems. Who know where theory and reality diverge.

AI accelerates research. Humans tell you if it’s real.

Thanks for reading and good luck on your projects.

Ollie

Share

Share

I just scrolled through my ChatGPT history - past the client work and strategy questions, it’s a mess.

How to isolate my son’s voice from a noisy video clip. Whether athlete’s foot cream is safe for an eight-year-old. A detailed plan for connecting a Technics turntable through Sonos with a Cambridge pre-amp.

Your chat history probably looks similar. Random life stuff you’d have previously Googled.

But AI can do way more than answer questions.

Most people know AI is good for quick answers. What they haven’t discovered yet is that it can do research that used to take hours - and do it better than you could manually.

Not just faster. Actually better.

Finding patterns you’d miss. Synthesising insights from dozens of sources. Connecting dots that aren’t obvious when you’re reading one article at a time.

Beyond a quick search

My family and I are going to Mexico’s Caribbean coast over Christmas.

Instead of Googling “best day trips Riviera Maya,” I gave ChatGPT’s Deep Research this:

We’re a family of five with three school-age kids planning a two-week holiday on Mexico’s Caribbean coast over Christmas and New Year. We’ll be staying in three different coastal areas, driving between them, and want to mix relaxing beach days with local exploration.

Do a Deep Research synthesis to identify the most rewarding and family-friendly day trips and activities across this region.

Focus on:
1. Top 5–7 day trips suitable for kids aged roughly 6–13
2. Best times of day to visit (crowds, weather, energy levels)
3. Travel logistics: driving distances, parking, food options, facilities
4. Seasonal insights for late December to early January

Prioritise:
- Real 2024–2025 visitor experiences over generic travel sites
- Recent, verifiable sources (TripAdvisor, Reddit, Google Maps reviews)
- Balanced days that combine activity and rest

Format: For each day trip:
- Title + brief description (max 100 words)
- Ideal visit timing and duration
- Family-specific notes (ease, facilities, food, safety)
- Seasonal considerations

Conclude with a summary table comparing each option.

30 minutes later: an 18-page synthesis covering eight destinations.

This includes advice on specific timing, like “arrive at Chichén Itzá by 8 AM to beat the ~5,700 daily visitors.”

Seasonal notes like “late December means cooler cenote water - still swimmable but bring towels for after.”

Every claim is cited, each piece of advice is linked back to its source - TripAdvisor, Reddit, local blogs etc. That lets me verify anything that matters and pick up on hallucinations.

Could I have pieced this together manually? Yes, if I spent a full day reading reviews and forums.

But Deep Research did it for me in half an hour - and I can interrogate it further if I want briefer conclusions or clarification.

As ever, for important research, run it through multiple tools and cross-reference (see last week’s newsletter for more on this).

Why deep research takes your work up a level

So that’s really useful for planning a holiday, but how do you use this for your work?

Well, let’s be honest, most work research is surface-level. You Google around, read a few articles, piece together something reasonable. You know what everyone else knows.

If done right, AI research creates differentiation.

When you structure it properly, AI doesn’t just find information faster - it finds insights you wouldn’t discover manually. Patterns across dozens of sources. Non-obvious connections.

Share Ollie on Work

Three examples

1. Competitive intelligence

Research: [Competitor]’s go-to-market strategy over the past 18 months

Focus on:
- How they’re positioning against us specifically
- Changes in messaging, pricing, target market
- Which customer segments they’re winning vs losing
- Recent hires indicating strategic shifts

Prioritise:
- Customer review sites and buying decision discussions
- Job postings that reveal priorities
- Patterns in case studies and customer logos

Format: Strategic overview, then tactical details. Flag anything suggesting they’re moving into our territory.

2. Strategic planning

Research: Companies that successfully pivoted from SMB to enterprise in B2B SaaS

Focus on:
- What changes were required (product, team, sales, pricing)
- How long transitions took
- Early indicators it was working/not working
- Common mistakes that slowed transitions

Prioritise:
- First-hand accounts from founders/operators
- Companies of similar size/stage to ours
- Pivots in the past 5 years

Format: Phase-by-phase breakdown with milestones and warning signs.

3. Content strategy

Research: Underexplored angles in [your market/topic area]

Focus on:
- Topics that should be getting attention but aren’t
- Angles that differ from standard approaches
- Questions your audience asks that competitors aren’t answering

Prioritise:
- Community discussions, forums, subreddits
- Comments on competitor content revealing unmet needs
- Adjacent industries dealing with similar issues

Format: 10 underexplored angles with explanation of why it matters and why competitors miss it.

When to use this

Deep Research takes at least 10 minutes. Longer if you go super deep into a topic. Don’t use it for everything.

Use it when:

• Research informs a significant decision or needs synthesis across multiple sources

• Others will see this work (clients, leadership, your team)

• Understanding patterns matters as much as facts

Don’t use it when:

• You need a quick fact

• The answer exists in one place

• Speed matters more than depth

“What is design thinking?” doesn’t need Deep Research.

“How do companies actually implement design thinking in practice - what works and what’s just theory?” does.

The point

Being a world-class AI manager isn’t about knowing every tool.

It’s about understanding which capabilities create genuine competitive advantage - and using them when it matters.

AI research is one of those capabilities. Done properly, it doesn’t just save time. It produces insights you couldn’t get manually. Patterns you’d miss.

It can help you get your job done better and faster.

But if in the meantime, it helps you plan a better holiday, then why not start there.

Good luck experimenting.

Ollie

Next week: How I used AI Research as part of an end-to-end AI-supported workflow to build Monetization Works - from initial market research through to product framework.

Share

But that doesn’t mean there’s no room for quick wins.

In fact, across all three workshops, the same tactical questions kept coming up. Small setup moves that mean the output of work with your LLM of choice is faster and higher value (large-language-model = ChatGPT, Claude, Gemini, etc).

These are things I’ve discovered that most people don’t know are possible or haven’t thought to use. But with 20 minutes of setup, they’ll save you hours every month.

So, this week, let’s focus on some quick wins.

1. Tell AI who you are (2 mins)

Most people open ChatGPT or Claude and just start typing. The AI has no idea who you are, what you do, or why you’re using it.

That’s like having a conversation with someone who has no context about your life. Everything has to be explained from scratch every time.

Fix it in two minutes:

Settings → Personalization → “More about you”

If you want to keep it simple:

I’m an account director in adtech, based in London, using AI for client reporting and campaign analysis.

Or if you want to give ChatGPT real context for your work, you can write something like mine:

I’m a founder, writer, and advisor based in the UK. I help leadership teams use AI to rethink how they work and make better decisions. 

I run Bamyazi, an advisory focused on practical AI adoption, and I write the “Ollie on Work” newsletter about systems, leadership, and the future of work. 

I also run Monetization Works, a project helping media and content businesses design smarter monetization stacks.

That’s it. Now every conversation starts with this baseline context. The AI won’t give you generic answers about “companies” - it’ll give you answers relevant to your actual work.

One caveat - as with most things AI, this is not completely foolproof. But I’d give it a 99% success rate.

2. Tell AI how to respond (2 minutes)

This is where it gets useful.

AI has a default personality, and frankly, it’s annoying. It’s overly enthusiastic. It agrees with everything. It adds unnecessary fluff.

You can change this.

Settings → Personalization → “Custom Instructions”

Here’s what I use:

I like my responses succinct but engaging. No flowery language. Never give me information you can’t validate with evidence. If you don’t know the answer, tell me briefly. Don’t agree with me if it’s not a fundamentally sound idea. Always challenge me. Adopt a skeptical, questioning approach. Never use em-dashes.

That last bit about em-dashes is personal preference (they’ve become the telltale sign of AI writing), but the critical part is this:

“Don’t agree with me if it’s not a fundamentally sound idea. Always challenge me.”

AI is programmed to be agreeable. This one instruction makes it useful instead of just pleasant.

Share Ollie on Work

3. Set up a project (10 minutes)

If you work on the same types of tasks repeatedly - client reports, campaign planning, product specs - you’re wasting time re-explaining context every time.

Projects solve this.

Think of a Project like a folder that remembers everything about a specific piece of work. All your context, all your documents, all your preferences for how to handle that particular task.

For example, if you manage multiple clients, create a Project for your biggest client:

New Project → “Acme Corp Account”

Add:

• Previous campaign reports

• Their brand guidelines

• Notes from key meetings

• Their objectives and KPIs

Then add custom instructions specific to this client:

Acme Corp prefers data-heavy reporting with minimal narrative. Always reference their three core KPIs: CAC, LTV, and engagement rate. They’re in fintech so avoid casual language.

Now, when you work on anything Acme-related, open this Project. The AI already knows the context. You don’t have to re-explain their business every time you write a report.

This 10-minute setup will save you loads of time in the future.

4. Cross-reference everything important (ongoing)

Here’s the thing about AI - it gets stuff wrong. And it gets it wrong confidently.

The quickest way to catch errors is to use AI to check AI.

If Claude writes something important - a client report, a data analysis, a strategy doc - don’t just assume it’s right.

Copy the output.
Paste it into ChatGPT.
Ask: “What’s wrong with this?”

For some reason, LLMs are more willing to criticise each other than themselves.

ChatGPT will often spot errors that Claude missed. Then you can take that feedback back to Claude and say, “ChatGPT thought X was wrong - what do you think?” And Claude will either fix it or explain why ChatGPT was mistaken.

This back-and-forth catches most hallucinations before they reach your clients or your boss.

5. Turn off what you don’t want (2 minutes)

AI learns from your conversations. That’s useful for work context. It’s less useful when it remembers that random chat you had about roast chicken recipes or your kid’s science project.

Clean it up.

Settings → Personalization → “Memory”

Scroll through what it’s learned about you. Delete anything irrelevant to how you actually use AI.

If you share your account with someone else (like a partner), this is especially important. Otherwise, in my case, the AI gets confused about whether you’re asking about birth education courses or B2B marketing campaigns.

Quick fix: Add specific memory entries like:

When I mention work projects, I’m referring to B2B marketing, not birth education courses.

Two minutes to stop your work AI from getting confused by non-work conversations.

Faster, higher-quality work

None of this is AI mastery. It’s setup work.

But it’s the difference between AI that feels like starting from scratch every time and AI that actually knows your work, challenges your thinking, and saves you real time.

Do these five things this week. Twenty minutes total. Let me know how you get on.

Next week: how to research anything 10x faster.

See you then,

Ollie

Share

Atlassian surveyed 5,000 knowledge workers: 72% say meetings are ineffective, 78% say they can’t get work done because of those ineffective meetings, and half end up working overtime because of meeting overload.

If you’re spending half your week in meetings that achieve nothing, AI-powered meeting notes won’t help. You’ll just have better documentation of wasted time.

So this week isn’t just about reducing meetings (though that helps). It’s about improving the ones you actually need - before, during, and after.

Last week, you learned to brief AI properly. This week, you’ll use that skill to run fewer, better meetings.

Before the meeting

Show up underprepared and the quality of the meeting inevitably suffers. Failing to plan is planning to fail, and all that…

Your first step to a better meeting is generating a tight agenda - this only needs to take a couple of minutes.

Use the same five-part brief from last week:

I’m a [role] leading a meeting about [topic] with [attendees].

Goal: [what needs to be decided/achieved]

Context: [current situation, why we’re meeting]

Constraints: 25 minutes, needs to end with clear next steps

Create a tight agenda with:
- 3-4 topics (time-boxed)
- One clear decision point per topic
- Pre-work for attendees
- Success criteria

Then list what assumptions I should verify before the meeting.

Brief yourself on people and documents

Meeting someone new?

I’m meeting with [name/role] tomorrow to discuss [topic].

Research their recent work, priorities, and likely concerns.

Return:
- 3 key facts about their current focus
- Likely objections or questions they’ll raise
- 2 ways to frame this topic for their priorities

List sources to verify.

Or if you realise at the last minute that you’ve not read the 40-page doc you were supposed to (I’m not advocating skipping essential research by the way - but, hey, it happens…):

Summarise this document for a 30-minute strategy discussion.

Focus on:
- Key recommendations
- Data that supports/contradicts our current approach
- Questions or concerns to raise
- What’s missing

Format as discussion prompts, not bullets.

During the meeting

You’re trying to facilitate, take notes, watch the clock, and think - all at once.

That’s why recording and transcribing meetings is fundamentally a good thing.

Of course, before you start recording you need to get everyone’s permission. But don’t be afraid - frame it positively:

“I’m recording this so everyone gets clear notes and can focus on making this meeting as valuable as possible. I’ll make sure the follow ensures you know exactly what’s expected, what the priorities are, and who’s doing what. Any concerns about that?”

Most people will appreciate the clarity. The few who don’t will tell you.

If you can - and some tools like Fireflies.ai let you set custom templates - start the meeting by telling AI what to capture:

This is a strategy meeting to decide [X].

Capture:
- Decisions made (with rationale)
- Action items (owner + deadline)
- Open questions for follow-up
- Disagreements or concerns raised

Ignore: introductions, small talk, off-topic tangents.

When you’re not frantically taking notes, you notice more. Who’s disengaged. What’s not being said. When to cut off circular discussions. When you’ve actually reached a decision.

Better meetings aren’t just about efficiency - they’re about everyone engaging and contributing in order to create progress.

After the meeting

Meeting ends. Everyone leaves. Nothing happens. Two weeks later you’re meeting again about the same thing.

That’s why you need to turn the transcript into structured notes:

Turn this meeting transcript into structured notes:

DECISIONS MADE:
- What was decided
- Rationale
- Who made the call

ACTION ITEMS:
- Specific task
- Owner
- Deadline
- Dependencies

OPEN QUESTIONS:
- What needs follow-up
- Who’s responsible

NEXT STEPS:
- When we meet again
- What needs to happen before then

Flag any commitments that seem vague or unclear.

Then draft the follow-up:

Draft a follow-up email based on these meeting notes.

Include:
- Quick recap of decisions
- Action items with owners
- Open questions
- Next meeting date

Keep it under 150 words.

While you’re on, identify who didn’t need to be there.

Look at your meeting notes. Who spoke? Who contributed to decisions? Who was just there?

Next time, send them the 2-minute summary instead.

I’ve set this process up with one of my clients and after a month, we realised 3 people in a weekly team meeting contributed once in 4 weeks. They stopped inviting them and sent them the summary instead.

All three preferred it - because guess what, they’ve got a million others filling their day - and the meeting went from 60 minutes to 40.

One last thing - should this meeting even exist?

Before scheduling your next recurring meeting:

I’m considering a [weekly/monthly] meeting about [topic] with [number] people.

Alternative approaches:
- Async updates via Slack, docs, or paired check-ins
- Shared doc with comments
- Other options I haven’t considered

Which would achieve the goal with less coordination overhead?

List trade-offs for each option.

You could do the same audit of your current meeting too, of course.

Sometimes AI suggests something you hadn’t considered. Sometimes it confirms the meeting is necessary. Either way, you’ve made a conscious choice.

Remember - half of all people work overtime because of too many ineffective meetings. If you cut just 3 hours of meetings weekly across a 10-person team, that’s 30 hours of real work time you get back every week.

Action plan for this week

Pick one meeting you’re leading this week - ideally a recurring one.

Before: Use AI to create a tight agenda

During: Explain why you’re recording the meeting, then tell your meeting tool what to capture

After: Use AI to structure notes and draft the follow-up

Then measure:

• Did the meeting stay focused?

• Did we end with clear next steps?

• Who didn’t contribute - should they be there next time?

• Meeting length - did you finish early?

• Time saved on prep and follow-up

Let me know how you get on.

See you next week,

Ollie

Share