Have you ever found yourself struggling to find just the right spot to go for an impromptu dinner with friends? Today I dust off some long cherished memories, and ask myself why it’s so easy to read a menu - but so damn hard to pick a dish. I personally believe the latest fashion to call everything 'Generative AI' is wrong, and even harmful - and I bet a story might help you understand just why!

Picture this - a city center late in the early evening, bustling with passers-by, and then there's that group of five or six young folks, standing on a street corner, their group pulsating, expanding and contracting, taking five steps to the left, and then ... hopping 6 back to the right, perhaps for an hour. What they’re doing? Well, this is a group of brilliant physics students, and no, they’re not acting out the movement of atoms in an electric field, or showing how celestial bodies get sucked into a black hole. A much more mundane problem is keeping them dancing in the street: where should they go for dinner?

Yep, this is in fact about me in my university years, and I can tell you the problem certainly wasn't a lack of dinner options. Ghent, the city where these gastronomical struggles played out, had a true cornucopia of choices: from plain sandwiches to exotic international flavors, from budget-friendly burger joints to Michelin-starred culinary heavens. And it also wasn’t a lack of brainpower holding us back, these were some of the sharpest minds I've ever known, and while physicists may have a reputation for struggling with the practicalities of everyday life, surely students are experts in the art of getting fed, right? So why the hassle?

Well, had this group found itself suddenly in, say, a Chinese city, the challenge would’ve been clear: we would struggle to decide where to eat, simply because we wouldn't be able to decipher all the signs and menus. But such simple information digestion wasn’t our problem, we could all read the signs calling out "Pita Kebab", or "Thai food", and we knèw the neighborhood, we had sampled every place more times than you can count. Our real issue was that our question was in fact ambiguous: What is the best place to eat? For one of us, "best" meant finding a spot with low-fat options because of their diet. For another it meant a hearty, calorie-packed meal, because they were prepping for a night of post-dinner debauchery. And yet another was running low on budget for the week, and just wanted some cheap filler. So each of us considered this in their own unique context, approaching it with their own one-of-a-kind strategy, and the resulting ambiguity is a much different challenge than the reading of a menu card.

Organizations also apply strategy in a specific context, but in fact, both the ambiguous questions ànd the handling of rich information cause headaches. Imagine an insurance agent receiving a claim by email. First, the information needs to be digested, reading the email to extract who the client is and what type of claim is it, to then forward it to the right unit to deal with it. And that unit needs to decide whether to honor the claim, but this involves more than just checking contract terms, and indeed is an ambiguous decision: perhaps you want to honor the claim simply because this is such a profitable client? In practice, because traditional rule-based automation offers very little in the way of handling these two types of tasks, most companies still fall back to humans, leading to costs in terms of efficiency, speed, and scalability, and this is where Artificial Intelligence can help.

The information handling tasks are done by humans day in day out, like seeing, reading, hearing, recognizing. These tasks are generic: the mental processes involved typically take less than one second of mental effort, and they are highly similar in different people regardless of the context like which organization you work for. If you're European, you don't need a new set of eyes to look at a picture taken in Africa, and no one needs to learn to read all over again when they change jobs to a new company. One big reason for this is that the information involved, the data, are highly standardized: images are images, texts are texts, and sound is sound. Finally, discussing the correct execution of these tasks is rather straightforward: an image represents a tiger, or doesn't represent a tiger, even if stylistic variations exist. Interestingly, these information handling tasks almost always have an inverse: you can read an email, or you can write one. You can look at an image of a train, or draw one. You can create computer code for a task, or describe the task a given code performs. This is of course what “Generative AI” is really all about, and its recent rise has been terribly disruptive for all of these information handling tasks, allowing great automation almost out of the box.

The second type of task however, taking those ambiguous decisions, is virtually the opposite of information handling. These problems require real thought and are often deeply unique. Most people never have to predict clogging of public sewers, or set an optimal price for a washing detergent, or identify a tax return as fraudulent. And even when two organizations happen to struggle with the same issue, their solution will differ because of context and strategy. The appropriate mortgage rate for a prospecting bank client will depend on context like the interest rates set by central banks, but also on the strategy of the bank: a bank aggressively expanding its client base will eagerly offer a sharp rate, but one simply focusing on drawing value from existing clients, won’t. These ambiguous problems often rely tabular data with meaning that is highly context-specific. For a mass retail bank a client with half a million euros is a rather big catch, but the same client may not even qualify for joining a private bank, so the same number means a drastically different thing to different organizations. This influence of context and strategy means that decision makers need to adjust their mental processes when changing jobs, and it also often makes it hard to debate 'correctness', after all, the appropriate mortgage rate isn't hard-coded in the DNA of that bank client. Finally, for an ambiguous problem it is often hard to imagine the inverse: perhaps 'describe the type of client that would get this mortgage rate'? But that's rather clumsy, no?

Most processes in organizations include steps of both simple information handling and ambiguous decision making, so to get rich value from straight-through automation, you need to master both, with Artificial Intelligence. Still, based on the above, I cannot emphasize enough the importance of recognizing how fundamentally different the two are: in the types of data they use, in the drastically different approaches, and in the skills they require. Any AI leader that doesn’t account for these differences into practice is setting themselves up for failure. For example, the information handling tasks are so generic that some foundational model from OpenAI or Google can often perform them close to perfection. Hardly any organization really builds such systems from scratch anymore, and the skills your organization really needs involve technical integration and customization, leading some to proclaim that the machine learning engineer has replaced the old-school data scientist. But for the ambiguous decisions, the good people at OpenAI do not know your specific context or strategy, meaning you will still need data scientists to embed your specific needs into tailored algorithms. Also, the risks involved differ, since one relies mostly on generic vendor solutions, while the other is often home-brewed. More and more, these two types of AI are drifting apart, require different governance and culture, and already today teams that do wonderfully well on one type, may fall horribly short on the other.

These two types of AI also follow different paths to value: automation of information handling typically brings big efficiency gains to rather simple, repetitive processes like customer support, while the ability to automate ambiguous decisions also allows you to bring straight-through automation to even the most unique and valuable processes at the heart of your organization: where to invest for an investment bank, when to do maintenance for a utilities company, or what to eat for a hungry student.

In the wake of the rise of chatGPT there's now a fashion to call every type of AI “generative AI” - after all, it sounds great, no? Well, I advise every AI leader to avoid this trend, and to distinguish clearly the two types of tasks: information handling (now often done wonderfully well with generative AI) and ambiguous decision-making. After all, they’re drastically different in the skills they require, in the burden they place on your organization, in the risks they entail, but also in the way they create value for you.

Thank you for listening to this second episode, I hope you enjoyed it, or that it at least got your thinking going. Because, this was just one of a 101 stories to cement your AI leadership, and many questions remain. Luckily, we still have a few episodes to go – 99, if the title is anything to go by. I felt I had to explain these two types of AI this time, because next time, I would like to offer some historical perspective about just how mature the field of AI is - or rather, the two fields of AI.