On Real Patterns, Inventory Control, and AI

It’s been almost two years since I last wrote about LLMs (recall here). Since then I published a paper on how to integrate the epistemology and ethics of AI (with Federica Russo and Jean Wagemans (here). But despite the fact that this paper is cited at a very healthy clip (even called “influential” in print!), I have not tried to muscle my way into the eye-catching debates about AI. And, except for deploring the deteriorating quality of good-old-fashioned Google, I have been a passive bystander in all the developments.

A few weeks ago, after attending the Eastern APA, I took a transatlantic day flight back home. I had upgraded to Premiere Economy. As it happened, I was sitting next to the managing director (CEO) of one of Europe’s largest non-food private retailers, but with operations in other parts of the world. We had an eye-opening half-day conversation. What follows is modestly edited to protect the identity of him/her and their company, which I will call STUFF henceforth.

Now, the success of all retail turns on a small number of key variables: the product(s) in the marketplace, the location, the quality/knowledgeability of salesforce, cash-flow and inventory management. Of course, in some businesses marketing, after-sales service, knowledge of one’s customer, and, for example, product development matter, too. This post is not a business manual.

But even in businesses where the margins are healthy, inventory management is underappreciated and rather essential to survival. Essentially, it’s always a precarious balance between avoiding shortages, whereby you miss opportunities, and gluts, which are very costly.

One, especially, tricky area in inventory control is how to handle returns. In fact, I already knew this. For most of my life my dad (who started out as a designer in his parents’ family business) was in the women’s ready to wear business as the Dutch intermediary for foreign manufacturers. One side of his business was sales in advance of the consumer fashion cycles. I wouldn’t call it exactly glamorous in his market niche, but it did involve a nice showroom and models (and all that).

But the other side of his business was dealing with returns from stores. This involved the gruffy world of transportation companies, customs offices, and giant trucks. For some reason I associate all of this with the smell of Diesel exhaust. Sometimes this involved unsold items, but not infrequently returns were the effect of late deliveries, mistaken deliveries, or problems in the delivered product (fabric, zippers, etc.). All of this in an age before barcodes. So, not to put too fine point on it a good chunk of his business involved knowing who to call that would take a whole bunch of out-of-date summer dresses or winter coats of his hands quickly.

Now STUFF is in a clearly defined market niche. But it operates in dozens of product areas. For various reasons and in a different number of ways STUFF makes it easy to exchange products. So, products are coming in and out of the store at a healthy clip. And at that point I exclaimed that it must be hellish to manage inventory, and (thinking of my dad) what to do with all the returned product. To which my travel companion, the managing director, replied that for STUFF the real problem is to figure out (i) in which store the returned albeit sellable product should end up, and (ii) how to get it there in an expedited fashion.

On (ii) STUFF basically keeps it simple and runs a centralized hub-and-spoke system in each country. The hub is a giant warehouse, the spokes are the stores. There is continuous experimenting with the number of trucks and routes, but the days of incoming and outgoing traffic are fixed. It is not the most fine-tuned system, but the stability allows for robust expectations in planning and, crucially, for company-wide inventory protocols and controls. This is wholly unremarkable.

STUFF invests a lot in understanding their customers and enhancing their loyalty. But even so it’s difficult to experiment with (i) in a fairly cheap way and get robust and actionable knowledge. The heuristics STUFF staff used would be out of date quickly or not robust enough to guide major decisions.

However, AI solved this (i) problem for STUFF. The AI was developed mostly in house over a two year period. How they acquired the skill and technology to do so involved non-trivial corporate creativity. Some of it sounded borderline shady to me. But it helped they happen to have a PhD in statistics on staff.

In particular, they quickly learned that AI would notice patterns in their per store sales and returns data that other techniques simply did not catch in real time. But initially — and this will not come as a surprise — it would also spit out junk every so often. So, after an extended period of supervised simulation, they decide to follow the suggestions of the AI in a limited number of stores. STUFF was astonished by the initial results. And so quickly scaled up the use of AI to the whole business.

To give you a sense of the order of magnitude involved. STUFF estimates it’s doing more than 10,000 profitable extra transactions a week, than it otherwise would, in a major European country alone in virtue of letting AI help it figure out in what stores returns should end up. That’s a lot of extra income. But it also means cash is less tied up in stationary inventory, and there is also less need to discount products aggressively as they stay in inventory over time. There are multiplier effects within the company’s operations not the least the need for management to coordinate all of this stuff. (Although STUFF clearly does not like splurging money on anything; the managing director was not sitting in the front of the plane, after all.)

So, I have no doubt AI is revolutionizing the behind the facade operations of a lot of bread-and-butter businesses. If properly trained and supervised, it is capable of picking up surprisingly powerful patterns in very messy data that become actionable signals. Of course, this is made possible also by the fact that the data is highly curated, and the whole AI program is oriented toward a clearly defined problem. Presumably a lot of the major existing tech and retail giants like Amazon and Walmart, financial services firms, and specialized scientific labs are much further down the learning curve than STUFF.

I don’t think anyone knowns in how many domains finding actionable and profitable patterns in existing data is out there. But presumably AI will contribute to productivity improvement throughout society. Of course, depending on context background conditions do change sometimes (and sometimes at higher speeds than others); that’s just to say the kind of intelligence involved in locating real patterns in statistical noise is not all-purposive.

In reflecting on this, I thought it would be fun to experiment and ask Claude, Deep Seek, and GTP 4o to find me papers that use Dan Dennett’s (1991) ‘Real Patterns,’ or ideas close to it in framing research in AI cognition or business applications. And, unfortunately, while the answers were definitely way much more plausible than a year or two ago, citations and references are still invented sometimes out of whole cloth. And would invariably end up at the point that one of the bots would tell me, “Given this correction, it seems that finding direct applications of Dennett's real patterns concept in AI cognition research is more challenging than I initially implied.” So, I would still urge caution to people in using AI to fix citations or to point to original sources outside their area of expertise.

Be that as it may, where are we?

AI’s contribution to the bottom line of STUFF is really impressive (even if predicted by folks a decade ago). And I would be amazed if there aren’t stories like this in lots of places. So, that’s really not nothing—something big is really happening in the world economy. (And clearly given market valuations that’s old news, too.) And this will reshape society in all kinds of ways.

Since on my Dennett inspired view pattern recognition is a building block of consciousness, I don’t think discussions about the genuine possibility of AI consciousness are silly. But I am not convinced yet that outside tightly controlled data environments and quite directed/specific/standardized/supervised tasks we should expect much more from AI.