You have seen the Hollywood interpretations of artificial intelligence (AI). There is the evil AI we remember from HAL in “2001 – a Space Odyssey” through “War Games” to SkyNet in the “Terminator” film series. And there is the good or at least well-meaning AI in “I, Robot” and “A.I. Artificial Intelligence”.
But what about AI in the real world? Isn’t that something that was hot 15 years ago, but we haven’t heard much of since? Is AI actually used for anything practical these days? Well, yes and no. In fact, many of the technologies that stem from AI research are used quite frequently today, but they are rarely called AI. One example would be the electronic “customer service agent/avatar” seen on many web sites—they often use AI disciplines such as case-based reasoning, machine learning, and natural language processing. Or perhaps you remember the Tamagotchi ? Or control systems robots, or air traffic control, or…
What about business applications then?
If AI can find practical use in a wide range of applications, what about automating some of the decisions that we do manually in our businesses today? Or what about using AI to provide a better alternative to rule-based processing which we use so much of in our business applications?
Under the umbrella of IFS Labs, we put that very question to two students from the Chalmers University of Technology. Over the past few months, they have as their Master’s Thesis studied possible approaches, evaluated potential use cases, built a prototype for one of them and tested it on actual data from an IFS customer.
After evaluating a handful of scenarios, they decided to try applying machine learning to the scenario of selecting from which supplier to buy a certain part when there is more than one supplier that can provide that same part. Most organizations today either select supplier manually or use very simple rules such as always buying from default supplier unless they are out of stock in which case another supplier is used. It is reasonable to assume that such simple rules, or manual choices, will not be optimal for large volume purchases.
The idea was that after some initial training (where the most skilled buyers make the choice) the AI would learn how various factors such as price, lead time, delivery accuracy, and quality affects the choice. The AI would then be able to pick supplier automatically, or for those who are less willing to trust in machines show a recommendation for the buyer who then makes the final decision.
The findings? Well, with just a little training the AI managed to make decisions with low error rate. It was also stable enough not to be thrown by some errors or “wrong decisions” in the training data. The algorithms used also turned out to have some other interesting uses. For example, as a side-effect, they identify trends on suppliers, for example, if a certain supplier tends to get picked (for the reasons of price, quality etc.) more or less frequently. This can provide valuable intelligence for purchasers to see which suppliers to phase out, re-negotiate deals with etc.
Confirmed or Busted?
Is the idea of using AI for automation of business decisions confirmed or busted? In myth-busters terminology, I would have to say plausible.
The problem is not with the AI itself—the algorithms developed work well—but with the scenario and real life data quality. For this to work well (and be worth the while) you need a high volume of decisions where there are multiple choices and up-to-date values for all variables that may affect the decision. Taking the “choice of supplier” scenario as an example lack of up-to-date price or lead time information for all alternative suppliers would lead to decisions made on wrong assumptions.
Actually, the students successfully tested the algorithm on a purchase of home electronics, where exactly the same product can be bought from a large range of suppliers, and online price comparison sites provide accurate information about price, available quantity, supplier quality rating etc. Unfortunately, there are not many industries or product categories where this data is as readily available.
There are other potential applications, for example in manufacturing when allocating resources during production line planning, AI could aid in selecting the best resource for a particular job. This can be done based on various features, such as availability, reliability, production speed, deadlines and other requirements. The algorithm will learn preferences of resource allocation and apply them to new cases. The advantage of using the algorithm over traditional planning would be that it could consider more features than a human planner, and can learn from evidence rather than stick to fixed algorithms used by automatic planning engines.
Do I think it could work? Yes, I do. Do I know in what business process, and for what decisions, it would be truly useful in real life? No, I don’t. If you do – drop me a line. We might just go ahead and try it.
It would also be interesting to see if an AI based resource allocation system could compete with a deterministic automatic resource allocation algorithm in terms of performance while giving acceptably accurate results. Because deterministic automatic algorithms can have exponentially increasing complexity when increasing the number of available choices.
Good point, and a different potential use. The thisis project we ran this time was more geared to using AI as an alternative to human decisions (and generally AI algorithms are faster than humans). Might be a follow-up theis in there…