The last one on AI for a while
At least by yours truly
Friends,
I hope that all is well with you and yours, and that this e-mail finds you on a boat with shoddy connection, in the tropics, three months after I sent it.
In today’s newsletter, I revisit a prediction about artificial intelligence and summarize my thoughts on the matter. For premium subscribers, we also break down Conviction Narrative Theory (CNT) for decision-making under radical uncertainty, discuss the geo-political implications of Trump’s stance on Taiwan, what is next for the US economy, and why many economists at the moment are adding little of value.
Now accepting keynotes for 24Q4-25Q2
Every year for the last decade or so, I have created three main presentation decks. For 2025, however, I have (for the first time) added a fourth due to popular demand. They are:
What to Do When You Don’t Know What to Do: How to turn change into a competitive advantage. (Based on the new book by the same name.)
Leadership in Times of Uncertainty: How to steer an organization through a sea of change. (Executive audiences only.)
Resilient Retail: How to survive and thrive in the modern marketplace. (Based on the 2025 follow-up to the highly praised 2022 white paper The Gravity of e-Commerce.)
AI Beyond the Hype: How to understand the narratives, risks, opportunities, and best uses of a new technology.
If you want to book me for your event, corporate speaking slot, or workshop, merely send me an email. To make sure I am available, please do so at your earliest convenience; my availability is limited and the schedule tends to fill up fast. More information may be found here.
A couple of updates before we go-go
First, foremost, and most importantly, today is my wife’s birthday. Happy birthday, my love!
Next week will mark this year’s start of the summer guest post series. We have a remarkable list of people lined up, each with a broader piece for everyone to read and an exclusive section of practical insights and takeaways for premium subscribers.
Without giving anything away, you will absolutely love what they will be bringing to the table.
On a rather less jubilant note, a new pet peeve of mine is the frequency with which many strategic management authors will simply make shit up so long as it behooves their argument. Declaring in certain terms how decisions are made, for example, without bothering to study the basics of cognitive neuroscience or decision theory just ensures that anything that follows is impossible to take seriously.
Speaking of decision-making and making shit up, I must admit that I am growing ever less patient with behavioral science. Not only is the field repeating many of the same mistakes that we have seen in psychology in recent years, but it is criticizing conventional economics while itself relying on the very same core assumptions.
I mean, at least strive to do better.
Moving on.
My view on artificial intelligence
TL;DR: the tipping point remains in the distance
A couple of weeks ago, I was approached by a retail executive who had gone back over their notes from the presentations that James and I gave at Cannes (The Gravity of e-Commerce in 2022 and Under Pressure in 2023). They had been impressed; we were proven correct on every single prediction that we had made. However, one particular piece of analysis had, they thought, perhaps not gotten the attention that it deserved: current-gen artificial intelligence, we had argued, would be a value chain vampire for the foreseeable future. In their organization, that had quite clearly become the case.
Beyond being flattered that someone would not only make the observation but also take the time to reach out to tell me, their comments made me compare our notes with recent research. What I found was revealing.
To clarify, what James and I noticed was that horizontal tech stacking increased post Covid due to escalating financial pressures and demands for efficiency. But although it was feasible that any individual app might help, we found that the likelihood was negatively correlated with the combined app portfolio size. That is to say, the more pieces of optimization software a firm used, the more inefficiencies (or at least higher collective costs) it would see. To make matters worse, the more money a firm spent on apps, the more time its employees would spend doing things on apps, and the less time they would spend doing things that made the firm money.
Artificial intelligence in its current form and most common application, we argued, was yet another app added to the puzzle and thus a potential value chain vampire. Although it could help, it was far from guaranteed that it would help. In fact, as far as we saw it, AI was more likely to underdeliver on its promises in the near-term than not.
Since our talk, more and more companies have, obviously, started to use gen-AI solutions. However, adoption is not yet anywhere near what some would have you believe. McKinsey’s “Global State of AI report”, for example, claimed that 65% of respondents were regularly using the technology and saw “material benefits”, both in terms of “cost decreases and revenue jumps”. A LinkedIn survey (conducted together with OpenAI) came in at 75%.
If using Google or Bing equates to using AI, the figures could perhaps be accurate. If it does not, however, they are dramatically exaggerated. The US Census Bureau’s continuous tracking of firm usage has the number at around 5%. In Canada, it is 6%. These numbers make more sense; adoption takes time. Despite all the hoopla around e-commerce, it still took more than 20 years for 20% of US retail to move online, for example.
As for the potential returns, a Goldman Sachs paper found that they were broadly negative; the companies that had “the largest estimated potential change to baseline earnings from AI adoption via increased productivity” had seen their shares vastly underperform the broader stock market. The technology, the investment bank concluded, carried no prospect for investors of extra profits. If anything, it could distract executives from more important matters. MIT professor Daron Acemoglu added that “given the focus and architecture of generative AI technology today... truly transformative changes will not happen quickly and few, if any, will likely occur within the next 10 years”.
To be perfectly clear, neither James nor myself would make such definite statements; the adjacent possible is inherently unpredictable. A breakthrough could hypothetically happen tomorrow. But we agree that it is unlikely to, at least for large language models (which are closer to “authentic” artificial intelligence though they firmly remain machine learning for now). The reason is rather simple: transformational innovations have historically not taken off until the technology is democratized, at which point the adjacent possible grows exponentially. Although ChatGPT grew very quickly from day one, it was a) very limited in scope (i.e., hardly transformational), b) enabled by legacy systems (i.e., did not need a new infrastructure), and c) largely a mirage (most early users have not been back since - growth today is stagnant).
Until the barriers to entry come down (gen-AI currently sits in the exigent quadrant in my 4E model of market dynamics), the broader impact, such as that on employment and productivity, will thus continue to be minimal. The numbers back the point up; rich country unemployment remains close to an all-time low and real output per employee is stagnant.
Having said that, there is a current shift towards smaller language models designed for specific purposes such as financial modeling, software programming, legal analysis, and so on. Although they do not have the intelligence of their larger siblings, they can simplify everyday tasks just as well (if not better). They are also significantly cheaper to produce, which makes the market more exposed.
From a microeconomic perspective, my recommendation is therefore to treat artificial intelligence investments as any other. If the technology makes sense to your business, by all means, try it out. Just keep two things firmly in mind.
First, it lies in our human nature to overestimate the objective difficulty of what we find subjectively challenging and underestimate the objective difficulty of what we find subjectively easy. Advanced mathematics, for example, is hard to do but relatively easy to model. If it can be modeled, it can be performed by an AI. And so, we overestimate its capabilities. Meanwhile, many of the things that my two-and-a-half year-old daughter is now learning are non-algorithmic in nature and therefore impossible for an AI to learn. Put into pragmatic terms, this means that the systems are easiest to implement in situations where automation already helps; contexts that are highly constrained.
Second, the size of the bet matters greatly. As much as vendors will want you to make a big one because they have, it makes little strategic sense to ignore the potential cost of failure or underperformance. As we have previously discussed, avoiding losses is important due to the volatility tax; a hidden “tax” on organizational investments caused by the negative compounding of previous investment losses. The larger the loss, the longer it will take to recover. The financial realities of AI are no magical exception.
Artificial intelligence is at the moment a business opportunity primarily for a select few vendors and a great many consultants. BCG, for example, has publicly declared that it expects 20% of its 2024 revenue to come from helping big companies work out what to do about gen-AI. In other words, ignore the hyperbolic sales pitches, do your due diligence properly, and approach artificial intelligence critically. Just like you would (or should) with anything else.
Until next time, have the loveliest of weekends.
Onwards and ever upwards,
JP
This newsletter continues below with market insights and in-depth content exclusive to premium subscribers. To unlock them, an e-book, and a number of lovely perks, simply click the button below. If you would rather try the free version first, click here instead.