Supply Chain Visibility & Artificial Intelligence

As some of the other articles on the site have discussed, I feel that there is a coming dominance of information over the purely material-management aspects of supply chains. And, at other times, I’ve mentioned my belief that artificial intelligence will play a major role in the shift. In this article I’m highlighting specifics behind that belief. The distinguishing line between tasks which must be completed by humans vs. machines is dynamic, changing more based on machines’ increasing capabilities than on changes in human abilities.  At a point in the near future, much of what is now “human-only” will become “human optional”, and finally “human impossible”. Data or calculation intensive tasks are an example of work that has recently shifted over these lines.  For those who are having trouble imaging robots doing supply chain analytics, I promise to start slow…

How to define Artificial Intelligence?

Although it irks some supply chain professionals, creating and using definitions are a major part of the problem solving we do. So, what is artificial intelligence (I’ll just use the acronym AI from now on)? For many people, AI would be a machine which can fool a human or group of humans via a conversation into thinking the machine is a human. This is, more or less, the Turing test. And for those who don’t know, it has not yet been achieved for extended periods of time. Of course, it probably matters a good deal who the human being are, what topics they discuss, in what medium (phone call, email, instant messaging, etc), and how long they can take to reach their decision. From the human side of the table, AI is about “passing” for human. From the machine side, it’s just more complex searching and calculation than we can yet produce at speeds necessary to avoid giving away the fact a machine is posing as a human.

So machines don’t yet socialize passably as humans. But, the track record is good for them taking over other areas which were considered human domain. For example, in the 1920s Nevil Shute had a team of British aviation engineers do full-time calculation to verify (as opposed to create) the viability of a structural design for military planes. The team worked full time for two weeks (around 1,200 hours of top-intelligence effort). By 1950 the same task would be accomplished in less than one minute on a special-purpose computer. By 2000 the verification would be done in less than ten seconds on general purpose computers. What’s more important, the computer wouldn’t just verify but would now design the airplane frame directly. The idea that humans are preferred calculators in today’s environment is laughable: in a side by side comparison the human is overpriced or simply incapable of solving the harder problems.

Today, new patented structural designs are created by computers. Does that make the latest computers intelligent? In the past, it certainly might have. But at each step, the definition of what is “human-only” intelligence has shifted. The most common example of this is in chess. In a very real way, chess was considered to be its own Turing test. It was supposed to be uniquely human to be able to play world-class chess. Well, we were wrong: computers can play as well or better than us. Maybe in the future they will do other things as well or better, such as striking up conversation or providing therapy to grieving individuals after they lose a family member.

AI seems to be a catch-all term which covers any cognitive task which appears, at the moment, impossible for anything but a human. Behind this statement is the more important assertion that what we call intelligence is usually a form of calculation. If they are not the same thing, they are usually substitutable. In other words, sufficiently powerful calculation can substitute for intelligence. This is what was seen in chess-playing, where human intelligence and machine intelligence compete using very different cognitive structures. Yet, given the task, they appear to achieve a similar level of performance.

Finally, let us note that because “AI” is based on current capabilities of machines relative to humans, it’s a moving target. And that target, that which we call “AI”, is going to pass over a great proportion of today’s supply chain jobs within my lifetime.

Intelligence is also Modular

If the “intelligence = computation” feels too simple, we should add that intelligence is not a uniform quality. Current cognitive science suggests that intelligence is modular. For example, dogs are particularly adapted to understanding human intention. They are the only animal which can interpret a human pointing to hidden food as an attempt to share information beneficial to the dog. Our nearest animal cousins (chimpanzees) are perhaps more generally intelligent but they lack the module of intelligence which allows them to understand human intention. As another example, humans have an intelligence mode related to coordinating to throw an object with our hands. No other animal on earth can throw an object well enough to use it to hunt prey. Of course, machines can do this using advanced calculation methods. This is another example of an intelligence equivalence being achieved via calculation.

Learning and Creativity: The Retaining Walls of Human-Only Intelligence

The two intelligence modules which are the current barrier between human and machine task dominance are (1) the ability to learn, and (2) the ability to behave creatively. Other notable exceptions include jobs of deep responsibility (president of a company, for example) and jobs which require human empathy (therapist, etc). For very low-value jobs, there are also differences in the ability to manipulate physical objects. I’m not joking when I say that humans are usually much less expensive to move a box, bag, barrel, etc with their hands than a machine designed for the same purpose.

But let’s stick with purely knowledge-worker jobs in the supply chain field. It is creativity and learning which makes human intelligence more effective at key supply chain jobs as of 2011. Let’s explore each module of intelligence in detail.

 

Learning:

Human being have most of our learning-intelligence built into us below the level of conscious thought. Individual capability to learn is varied, but we are measurably good learners. It may be our most powerful feature as a species. For example, our nearest animal relative (the chimpanzee) will learn a useful but complex physical technique after being exposed to it for around seven years. Humans will learn a physical technique after only seeing it being done in front of them 1-2 times. Machines with good AI software fall somewhere in between the chimpanzee and the human, but usually only on a small subset of processes for which they are designed to learn.

19

 

In a supply chain context, there are many tasks which require constant shifts in process. At present, the supply chain industry software available would not be able to sense and learn from these factors enough to replace humans in analyst positions (i.e. entry-level). The supply chain analyst tasks require a human-level intelligence to understand what changed and how to adapt to it. For humans, these may be rather mundane aspects. For example, deliveries from a supplier tend to be late if they don’t have a reminder email two weeks before. Because humans make almost infinite unconscious cognitive attempts to find these kinds of relationships, it seems obvious to us. Software AI just isn’t yet at this level as of 2011.

Creativity:

Sometimes humans don’t just learn from reality: sometimes we make an imaginary world in our mind and, through scenarios played out on it, discover a totally new approach to solving existing or undiscovered problems. In supply chain management, I believe the ability to negotiate a contract for existing products or standard services (like brokerage rates) may rely on creativity. Negotiating contracts for non-standardized services or partnerships certainly requires creativity. Humans may not be amazing at creative tasks: it is a rare gift, in general. But we are clearly superior in this kind of intelligence when compared to machines. Only with a well-defined problem are current AI software finding novel outcomes of value.

A taste of the future…

Machine learning may be the first of the two main roadblocks to be overcome. A computer & software which could learn from website or other written content could be valuable, but not immediately game changing. Bear in mind that there are a lot of perfectly capable learning programs out there in the form of humans. An intelligent human capable of learning new supply chain processes or content costs between 12-40k USD per year (depending mostly on geography and working language). Getting a human with deep expertise is 2-4 times more expensive. Given a computer’s ability to run constantly, it might be justifiable to say the 1st generation to pass the human-level of learning would be worth a low multiple of a human salary. So initial software costs might be in the ~500k USD range (assumes a tolerance for three year break-even). If your company bought one of these AI machines it would be a (probably weak) member of the team, similar to any new-hire human. It would NOT run the whole supply chain department and axe all the staff jobs. In time, as new generations of the AI are deployed, something truly game changing will occur. This is because machine learning will cross human learning capabilities fairly slowly. Remember that intelligence is modular and, as a result, machines may exceed humans in some forms of learning while lagging in others. The real breakthrough occurs when all necessary forms of learning are dominated by AI rather than human intelligence.

Author Josh Storrs Hall, in his book “Beyond AI: Creating the Conscience of the Machine”, describes this key turning point as “Epi-human AI”. Given the ability to learn in many formats, without a pause, and on fast hardware or networked hardware, the Epi-human AI would be similar to a thousand of the brightest staff in your company all compressed to one.  It would involve an AI capable of reading an average book (with high comprehension) within several seconds. This AI could complete a college course, with all coursework, in 10 minutes or so. In total, it would “produce the equivalent of a human’s lifetime intellectual output, complete with all the learning, growth, and experience involved, in a couple of weeks.”[i] For obvious reasons, this will be a turning point for all human endeavors, including supply chains.  An Epi-Human AI would totally dominate what we currently have as supply chain departments. In smaller fields, an Epi-Human may even lead or manage a stand-out business.

Before moving on let’s say something about creativity. I’m not an AI researcher, but I have not seen a lot of discussion pointing to AI creativity being a near-term strength. For a longer time into the future, this may remain a differentiator.

Back to the Present

This article has gone out a little further from our usual down-to-earth tone. Let’s try to get back onto firm discussions about supply chain visibility using technology available in 2011. For reasons I discussed in other articles, the supply chain visibility field is facing an onrush of data. Our ability to manage data to the advantage of our supply chain and company will become a significant, perhaps even singularly important, part of supply chain visibility. It’s within the context of a mounting tsunami of data and the need for data-management that we must expect increasingly “intelligent” software to be deployed. I’m going to go ahead and make a bold statement here. The main users of supply chain visibility will probably switch from people to computer programs. This is because (a) people are just not good at managing the data quantities that will be available, even with good pre-processing “tools” and (b) the adaptive & learning ability of the “tools” which help humans deal with vast data will begin encroaching into the steps humans actually complete alone. If the human staff always needs “tools”, and use them in predictable ways, are humans really necessary in the process? What this means for supply chain visibility is that our solution design should generally treat the “user” and “usability” carefully. If the “user” is a group of human staff, the solution should cater (and at times compensate) for their human decision making styles. If the “user” is a software agent, then the visibility application and processes must cater to that user’s particular needs as well.

To be clear, this is NOT a science fiction scenario. There are already fully deployed, proven business cases where the visibility application leads directly into a semi or fully autonomous decision making application, which in turn ties into an execution application. Together, these form a kind of OODA loop as shown below (see link here for OODA loop definition).

20

In other fields this is also a proven setup. Fraud detection for credit cards, as an example, has the exact same loop: visibility of transactional space along many dimensions (time of day, geographic location, spending type, etc) ties into an AI agent which identifies suspicious transactions and makes a fully autonomous decision about how to react. The decision is then executed by sophisticated systems designed to limit or halt transactions in the global finance system. High frequency trading, military logistics, and many other areas have these systems in place.

Monday morning wrap-up:

As with all the articles on the site, I’m ending today’s discussion with some bulleted points about how this information can help you immediately, on Monday morning at the office.

  • The term “Artificial Intelligence” is hard to take serious for many people, partially because it is a moving target. What is considered “human-only” intelligence levels have changed drastically in the last fifty years, and will pass over many supply chain jobs in the next 30 years.
  • AI does exist, but usually as soon as a task is possible by software it is denuded of its “intelligence” and just thought of as ”computation”. Chess, speech processing, and the as-yet-unachieved general topic conversation are examples.
  • There are some important tasks humans can do much cheaper than a machine, which have nothing to do with intelligence: we’re pretty good at picking up small, random shaped objects from obstructed areas. A general purpose machine that did this would be very expensive to build and maintain.
  • But, in “knowledge work”, the two key places where human computation exceeds software are in creativity and our ability to learn random processes quickly. As machines achieve and then pass humans in these areas there will be world-changing effects.
  • Supply chain visibility, depending as it does on massive data management, is likely to be targeted for early AI deployments. If the supply chain visibility application isn’t using AI, it is just as likely that the user of the SCV application will be a software agent instead of a human.
  • Even today there are deployed, proven examples of visibility applications which serve up their intelligence to another application, which makes very fast and autonomous decisions about how to react to the new situation to meet business goals.
  • In whatever way you touch supply chain visibility, keep an eye on AI applicability. If you produce visibility processes or solutions (as a consultant or software provider), this is part of the future solution architecture. If you are buying or running visibility processes, ensure you have a plan in place for integrating AI into your business process.

Leave a Reply