Introduction

Good evening, everyone. Jonathan, you gave me a great way to do my transition because you did mention and talk about, you know, work and productivity and how AI is helping us, assisting or augmenting how you make it work for teams.

Productivity, speed, and the “AI myth”

and today that's what I want to talk about I want to talk about productivity and the perception of productivity a bit as a myth so who's using AI today just

show of hands okay does do you do you use it for what use it for coding presentation documents reading emails research all of the above do you feel it makes you faster, more productive? What do you get back out of it?

Time? The illusion of faster. You can do more? Productivity.

Does it give you back some time? Back to your point. Do you do more with it?

Do you use that time to do more of the same? I sleep. You sleep more?

You're part -time now? Okay. Do you make more money sorry do you make more money thanks to ai yes how many are making more money out of ai

okay it's pretty good it's pretty good i'll explain why it's so difficult and why why we think we're making money but actually very few actually making money out of ai you thought it through you had more time than me your entire career um and and that's kind of the argument today.

At least for a lot of us, AI is speeding up a lot of what we do. It doesn't mean it's improving or increasing the results. And I'll try to deep dive into that and explain the why

and explain how you can change it to actually create significant outcomes. And that means money, P &L impact, economic output. Okay.

What the data shows: more code, not necessarily more value

So we start with a study. MIT and Wharton have analyzed analyzed the work of 100 000 developers and what they found was kind of interesting pre -ai and

post -ai post -ai they're able to write eight times more code than before but they can ship only 20 more releases of software so the question is why if you can produce so much more code why can you not ship more releases and we got to assume that shipping more releases means happier users or more users spending more therefore you're making more money right no correlation more releases doesn't

necessarily mean actually you're making more money although you could so if we deep dive a bit more these are the steps right of software development high level from you can code faster write code faster.

But then the number of commits they observed in the research said, okay, it was more, but only significantly more, only slightly significantly more. And then pull requests, and then finally release it.

So you see that the process is decreasing as you're changing the output. It's like you're forcing more code into the process. But at the end, somehow, there's some kind of loss in translation that you're not able to release more the paper shows that there was

no impact on having more between more releases and more user adoption or acquisition or money so first of all you can produce more releases that completely uncorrelated with how much money you make in fact you download probably new apps 30 40 new app updates every night when your phone is plugged in and you're sleeping but it doesn't mean you're buying stuff right

so the number and the productivity here doesn't translate into money necessarily and we'll see how it can but the research here is showing showing that so what AI has helped do is the automate the very first step you could argue

before coding obviously there's thinking requirement gathering interviewing of of people, and you can automate a lot of that, you know, specification aspects, the documentation of it before you even get into coding.

But the biggest bottleneck of the software industry was and has always been about how fast you can type something good, some good code, right? So AI has eliminated completely that friction.

You can literally type, you know, a few thousand words a minute. So the typing is not a problem.

Now, you could also see why software languages have been evolving to create more abstraction layers so you don't have to type all every single function so there was a way to kind of work around that productivity problem but still you were still

Systems thinking: the Theory of Constraints

limited by how fast you can type so what i want to introduce here is the theory of constraints and system thinking if you look at software development end -to -end is as a production line that is creating a product from an input, an idea of an app, all the way to a release you can sell, and a release I can money, you think of it as a chain.

And what we'll show here is removing a single constraint on this system is not going to produce the right outcome unless you have picked a constraint that is effectively slowing down the entire system.

so what it means is in a company you can you can do now maybe more marketing campaigns right you can create more leads but your sales team cannot process more leads so you cannot convert them faster or say what marketing's been now super efficient sales is more efficient but you know what you have one person who's legal who's actually reviewing every single contract and guess what It piles up because they haven't been augmented.

So do you do more business? No, because you're not signing more deals, right? Same idea, right?

So another illustration, just so we're all at the same page, is the hikers here hold together, held together with a rope, right? You can imagine now, picture the first one trying to go a little bit faster than everybody else. He's going to be stuck, right? Because the rope's going to stretch and quickly he's going to be stuck.

The last one can try to go a bit faster, but guess what? He's going to bump into the other one, right? So the point is, this group of people can only go as fast as the slowest of the hikers, right? Which is exactly what we're talking about.

The constraint may be any of these. If you help one of these hikers to move a bit faster, the slowest one to move a bit faster, the whole system will be able to move a bit faster.

So going back to software development, this is what it means if today you say you can write faster but you're not able to integrate review approve release and adopt faster right now you don't see much economic output out of this process and again it works with the other processes i kind of mentioned right now not 100

true some of these steps here start to be automated but the point is if approval is automated you know why are you approving anything in the first place right the release management same, you can automate it, but if no one is controlling it, then at some point, this autonomous process with no human oversight to remove the bottleneck is actually also a problem, and it will backfire, right?

When “human in the loop” becomes the bottleneck

So when we talk about human in the loop, you probably have heard that a lot. The human in the loop is usually creating more problems because a human in the loop means human is reviewing what AI has already done.

And if you haven't designed your process properly, you create create a bottleneck just like this one, because humans cannot actually approve and review more than what the AI has done, unless it's been, again, managing by exception, and it's been well designed to do so.

So that can create more problems than solving one. And which is why also you see now, a lot of so -called AI pilots and initiative have had no impact on the P &L.

And and great demos that are showing how AI can do great things. Usually, it's been one department trying to do one initiative, another one doing the other one, and then everybody looks good.

But because it doesn't gel, because it doesn't optimize that chain that we talked about, there's no economic impact at the end.

So you can have a team of three who decided that it was really annoying to do all these marketing assets and templates, and they've optimized, well, a team of three in a company of 15 ,000. well, there's absolutely zero economic impact.

The team feels good. They created some good stuff. But if you look at it at the company level, no impact,

which is why, again, there was a study by MIT, controversial maybe, saying like, what, 95 % of the AI projects don't produce any value.

That's part of the problem is they haven't picked the right constraints. They pick whatever was easier to automate, which doesn't mean there was actually any relationship with the output.

The economics of AI: tokens, quality, and pricing risk

So, the first part was about the constraint, second part is about the economics. And the economics, you would have seen some interesting behaviors in the past few months, which was about token maxing, right? Token maxing.

Who's been a token maxer here? So the token maxing is about whoever is using the most AI is the champion, is the best employee in the world, right?

But if you think back about what I said, tokens are just an input into one of these steps you're trying to automate. And tokens is effectively money, right? More tokens doesn't mean better.

In fact, there's a study between Microsoft and Salesforce across 200 ,000 conversations showing that after the first turn of a conversation, so multi -turn conversation, back and forth, after the first turn, the quality and accuracy dropped by 40%. So more tokens, more or longer conversation doesn't mean better output in the first place. Second is the price.

Today, we are roughly paying a tenth of the price of a token if you consider the massive investment and subsidies being put into the infrastructure. So if your business case about automation today does not support a token price increase from three times to five times, it may be dead already. that means the moment it's the price is jacked up then you're gonna have a real

problem our eye problem and in fact people like to keep telling me oh but the price of the token is dropping yes but models consume more and more tokens and guess what you have zero control over how much they could how much token they consume for every request you make so they are more powerful they they

consume more tokens you have no control over that the quality tends to drop up after more tokens are being consumed and the price may change at any point of time. So the economics are still to be stabilized in that sense, right?

From doing work faster to doing work differently

So the question, I guess, what I'm trying to get to and to get you to think about is instead of asking about, can I do this faster? Effectively, can I do the same work today a bit faster?

It's like, can I do it differently? And where is the constraint that I can unlock so AI can have a real impact back to the chain of and think of the hikers, the group of hikers, right?

And so, and I think, again, Jonathan, you gave me plenty here. It's not about, again, automating what you do. It's really redesigning the work, right?

Because when you think about how a customer support service works, and suddenly you bring a gigantic automation into it, you need to think about a lot of stuff. You think handovers, when should actually an agent, an AI agent pass it over to a human agent, how it should actually be able to escalate or exit.

Are the emergency procedures to actually shut it off if things actually go south? So that requires more than just automate state A, B, C and replace it by AI.

So redesigning work, it also creates new opportunities because you realize like if you keep doing what you're doing just a little bit, a bit faster, it's always to the bottom because guess what everybody else is doing the same so now the five hours you have back people

expect you to be doing more work because you can because it's cheap so the problem is with that it's a race to the bottom unless you use it to reinvent how you deliver your business how you make money how you create new revenue streams maybe new services in a software industry for

Creating new revenue opportunities, not just efficiency

example i was meeting with a friend who's used to design a point of sale systems and he was like Like, well, in the past, I couldn't accommodate all the customizations my clients were asking for. Now, actually, and they were ready to pay for it. At the time, he didn't want to fork the product.

He was too much overhead. Now he can. So he can actually now have and provide a white glove customized services to his clients because AI has actually dropped the cost of making this change. And now it allows him to make it possible.

So it creates, again, a new way to look at your business and I think new opportunities if you want to rethink of it.

A practical way forward: start from the goal, then fix the constraint

So there's one thing to remember, you know, working backwards, always work from the goal. What is it that you're trying to achieve first? What is the metric you're trying to move? And then you work backwards from that.

You look at the constraint, you look what is slowing down the whole chain, and then you can apply AI on that specifically. By the way, not just AI, it could just be process redesign, operating model change, people change, skill set or competency change. AI won't solve everything in every company for every problem you have, right?

And one last point, AI can augment, AI can assist, AI can automate, and more and more can automate autonomously. And that means giving it a goal and be able to work on a task and on a series of tasks end -to -end.

Jonathan was mentioning earlier also about how to work and how to make AI work for us as a team. I think this is part of the answer here, which is think about the process end -to -end. Think about where the humans are and should continue to be, where judgment is needed, where empathy, emotional intelligence is needed. and use AI to actually augment and automate the steps that can be.

So sum up, again, AI can speed up a lot of stuff, but to improve results is actually not that simple. And I hope today it gives you a little bit of a glimpse of how you can tackle it, if not tomorrow, maybe Monday morning.

Conclusion: the marathon analogy—optimize the whole system

The last analogy is why this picture. So another last analogy, I guess, is how you pass the line and think of it, again, really as a marathon here. Yeah, you can see it.

So I'm an ultra runner. And to finish a race, it's not just the legs that have to actually move fast enough. It's the system, right? It's the system together that has to finish and cross the line.

I'm hungry. It's dark. I need a headlamp.

It's me at 3 a .m. in Japan last year, finish up an 80 -kilometer race. So again, the legs are actually the least of the problem usually. It's the mind, it's the hunger, it's the heat, it's whatever that is, but it's the system, right?

And the legs can move very, very fast. It doesn't necessarily get you to the finish line if you don't have the mind to it.

All right, thank you very much.