Episode Transcript
[00:00:03] Speaker A: Nada. So, I mean, following up on this, really, you've got a back catalogue of books and research papers about forecasting and resilience.
What concrete practices within businesses or within teams within these silos that Pawan was referencing there, what gets people to move from reacting to events towards identifying risks, planning those scenarios that you talked about before, uh, so before they, their operations are affected, what's the best way? Is there a benchmark people can aspire to?
[00:00:36] Speaker B: Yes. So I'll, I'll give you a few examples of a few practices, but what I want to do, if I may just for a second underscore what we've already said, because I think it is so critical. It's the issue of the data and dirty data and the silos. And as someone that has been doing this for a while, as this team has as well, it is somewhat frustrating to see that we're still talking about this, right? That these internal silos that don't have visibility with, with one another and that aren't talking to one another, these are some really fundamental problems. So now as we see this advancement in AI, agentic AI and all of it in terms of what's happening, we haven't even gone back in terms of restructuring, making sure we have flows inside the organization, let alone within our know, with our partners. But back to what you're asking, I think it's really important to distinguish between visibility and then early warning capability.
So right now more companies have more data than they've ever had before, but they still react late. And I think in large part it has to do with the fact that they haven't designed the process to turn that signal into action. And it goes back to what has already been said. I'd say there's, there's a lot of things I can add to it, but I think I can think of at least a few really concrete examples. One, companies have to really move and look beyond transactional data.
Traditional data that we use relies heavily on orders, on shipments, historical demand.
But those are really lagging indicators. Right. And I think we need to, as organizations really begin to look at external signals that are going to give early risk notices, this would be supplier behavior, port congestion, weather, geopolitical events that are happening on a moment by moment basis. So we have a lot of other kinds of things.
So it's a different kind of data.
Two, I think we need to move, and I've already mentioned this, from a single forecast to a range of possible futures.
So rather than saying here is what we think will happen, I Think teams should be saying, okay, what are the three or four plausible ways this could unfold? And then how would we react to each one of these? And then I third, and I think you've alluded to this is I think it's really important for organizations to predefine trigger points. Right? It's not enough to identify risk, but you need to have trigger points that are set in advance. For example, you could say something like if lead time increases by 20%, if a supplier's reliability falls between a certain threshold, if geopolitical risk crosses a certain level. So you have these thresholds, you know, very similar to what we've done in quality control for decades, but you have them preset and what, what they do is they trigger automatic reaction. I think it's really important to do that. So I think what we need is to shorten the distance, if you will, between that signal understanding and be able to take action.
[00:04:08] Speaker A: The difference between that trigger, that risk trigger that you just talked about, can we look at that through the reference point of your book, your machine, which argues that AI isn't necessarily replacing people. It's about there's a future where we have a genuine integration, where each side gets better, together, you're stronger. That type of integration, I'm interested in how that works in terms of forecasting risk. Is there a case from that book maybe, or a real life example where you've worked with a shipper or a manufacturer, where you've watched that integration work or maybe fail in terms of those trigger points?
[00:04:45] Speaker B: Yes. Let me start with a general example and then I'll give you one that has to do more with shipping. But I think the central argument in the book, the human, is that the winning model was never going to be human versus machines. And I think that is something that is more true than ever. So all along we've been saying, what's better? You know, is AI going to replace humans? And so forth in the book, and there have been two editions now we're going to work on a third one. And then I'm actually working on another book, the Agentic Enterprise.
All of that, the foundation and really the crux of my work for a really long time, decades now I'm getting older here, is that they have to work in tandem. Humans and machines have to work in tandem. I saw this decades ago when I, you know, as a young doctoral student that was, you know, we had the birth of neural nets and all of it. And we cannot replace decisions with the information, the data, because at the end there Are humans and trust and other kinds of things. That's the essence of what the human sheen is about. We have many, many cases in the book.
One, just a good example that I want to start with, it was Unilever because I think it's just a really good example of something where you could take things that are very rote, you give it to machines. And in this case, I think the example that we started with was the hiring process where basically they're using technology to handle this high volume early stage screening while preserving human involvement for the consequential decision. This is I think an overall meta example that I think is really important because human decision making, expert judgment, the kind of things that listeners have, they have experience, it's a scarce resource. As humans, we do get tired, right? We get decision fatigue. So we have to preserve decision making for the things that really matter. So what Unilever did is they basically redesigned the process so technology AI was going to process the screen through the candidates in a very automated way, identify patterns, narrow the field. But then when it comes time to hiring, it remained up to the human. That is a really important thing. And then when we look at other examples, principles in supply chains, I've personally seen this with Maersk, with Amazon, the amount of information that is confronting an operator is simply beyond the human scale. It's beyond what an operator can process. You could have thousands of shipments moving. I've seen this with Maersk. Changing demand, port, conjunction, inventory position, you've got supplier issues, weather, all of this is coming at them. They can only process so much. This is where AI can be extraordinary. It can monitor all of that. Then what it does, it detects anomalies, it can predict where a bottleneck might emerge and then it gives alternatives and you pass it over to the human, a human that has the experience, the judgment. So we really then ask the question, should we reroute the shipment? Should we pay more to protect this particular customer? This is where human judgment comes in. It's the humans that understand the customer supplier.
But you can't waste that judgment, Mike, on tire people out on the road things. That was the lesson in the HU machine. And what I'm seeing now is, is it is more true than ever.
[00:08:42] Speaker A: Well, where does the human machine end and agentic AI begin? And how does that factor through into the future of risk forecasting?
And a secondary question to that is there's an awful lot of AI companies out there promising all sorts of different things. I mean, you've been an expert witness in cases along these lines, what should people be wary of when vendors tell them that AI can make these decisions for them?
[00:09:10] Speaker B: Oh boy, once you get me started. So I have done a lot of expert witnessing.
I haven't recently only because I commit so much. Mike and I really try to do a good job and we get paid well for that. But let me tell you, you earn every cent.
For me personally, the expert witnessing was especially valuable because you get to see under the hood, right? Everybody talks about how everything is great, but I can't tell you how many times I've seen companies that have bought software packages. I'm not picking on anyone or any consulting company to be surprised as to how many hidden costs there are. I could tell everybody out there, I can almost guarantee you you will encounter hidden costs. So you need to really plan for that, add ons, even silly things like I've seen companies say, but what if we want to query the algorithm and in this way versus that way, and then you get the usual, well, we can do that, but it will cost you extra. So that is definitely something to be cautious of. I've already mentioned, it has been said the data issue, the process flow, those are things that are really fundamental. I think so many promises are being made right now, especially as AI moves from gen AI to agentic AI, which is the next stage of the integration model. It's not the end.
What I am working on, I can tell you I'm working on this right now, I was working on it yesterday, I'm talking to companies, is where the human enters the process.
So the human machine and everything that I've said up to this point is completely valid, it's true, more than it's ever been. But the issue now becomes where does the human come in? So with traditional AI, we often imagined a person sitting next to the algorithm. The AI makes some kind of a forecast. The human decides whether or not to act. You know, it can press, you know, accept or not. Agentic AI changes all that because now the system can do everything, can take action, it can observe the outcomes, it can continue acting.
So the agentic AI doesn't invalidate the human, but it makes the design of this entire process more consequential. It also brings in the question of where does the human come into play early on, checking in the data or somewhere throughout the process.
And that is something again, like I said, I am literally working on at the moment. And it's going to vary in terms of the size of the company, the decisions that are being made. The human may no longer approve every shipment rerouting or every inventory adjustment. So it's also going to be about now setting up those boundaries.
What level of uncertainty is acceptable? When do we escalate the model to the human level, and when do we automate? These are really hard things to put into processes. And this is where I think the key is redoing, readjusting processes. One of the things that as an operations, I'm operations supply chain person, we're pulling out the old playbook, not just in scenario planning, but also in terms of workflow analysis and workflows and bottlenecks, because we have to actually rethink what that process is. And where do we put the humans and where do we put AI? Not to create bottlenecks, but to actually oper optimize the entire decision flow. And it's really hard.
One other thing, if I may, that I want to add that I think is really important is also to include when does the AI stop acting?
So I can tell you, AI is going to continue generating endlessly and creating, as we all know, a lot of slop.
And we've all seen this, right? You get one iteration, two iterations, and then it just continues. And then it sort of, you know, just really is a downhill process.
One of the things companies need to do is also put in when is good enough and when does the system stop acting? So all the things that I mentioned, the trigger points, when do we escalate and when do we stop?
