Machine Learning in the Enterprise
What is machine learning?
First, I should emphasize that machine learning isn’t only about machines learning. It’s ultimately about having computers learn more like we do – by trial and error, and through experience, rather than having to be told explicitly what to do — and that opens up vast new realms of possibilities.
So does it sound amazing to you or slightly frightening?
Both! First, it’s an amazing opportunity — in the future, we can look forward to all kinds of products and systems that automatically getting better, over time, and as people use them. I know that sounds like science fiction, but we’re already seeing examples of this inside organizations.
But, as with any powerful technology, there’s also the possibility of misuse and abuse. Machine learning involves delegating decisions to machines, so it’s important to have robust governance in place, and to make sure that people remain firmly in charge.
How does machine learning work?
Machine learning is the term we use for any type of algorithm that can automatically update itself based on data. It can be an automated version of the kinds of statistics you learned in high-school, or it can be based on neural network approaches that are closer to the way our own brains work.
Without going into the technical details, machine learning generally “works” through algorithmic trial-and-error, receiving data and getting rewarded for more of the “right” types of answer.
So, focusing on the amazing part: what can we use machine learning for?
The first big opportunity is a massive increase in productivity. Wherever there’s a complex, repetitive decision as part of any process, we now have an opportunity to automate that. McKinsey believes that in some business areas such as finance over 70% of processes can eventually be automated in this way.
For example, we worked with a large chemicals company to introduce machine learning into their finance function; specifically, something called “invoice matching”. When they first implemented their business systems several years ago, only 40% of the invoices that they sent out to customers were able to be automatically matched to the payment information they received via their bank. So, the majority of the time, the reference numbers were different, or there were two invoices for one payment, or two payments for one invoice, etc. and a human being had to step in and sort out what was going on.
Over several years, they put in place a rules-based approach and managed to get the matching rate to 70%. But with machine learning, based on all the historic data they had gathered over the years, we were able to get the matching rate to over 94%.
And what’s wonderful is that it’s still improving: any time there’s an exception, it gets kicked out to a human being, who figures out what is going on, and then that knowledge goes back into the algorithm to do a better job next time. It’s a great example of a business process that is automatically getting better, over time, and as people use it, thanks to machine learning.
And I know that invoice matching might not sound very exciting compared to things like robots — but this company has hundreds of thousands of invoices from hundreds of entities around the globe, they’re really excited about how much time and money this new approach is going to save! And that’s just one small example among dozens of others in finance, and hundreds across a typical organization.
Second, it’s not just about efficiency; it’s also about augmenting human intelligence. For example, we can provide sales people with intelligent guidance: which prospects should they call first, who is most likely to purchase? What products are they most likely to be interested in? How long with the deal take? Ultimately, it’s about using pattern matching to help all of the sales people be as good as the best sales people in your organization.
Finally, it’s not just about profits – we can use machine learning to make the world a better place. A great example is Rainforest Connection, a not-for-profit organization that works to reduce deforestation because of illegal logging. They have pioneered the use of “bio-acoustic” monitoring: they use old GSM phones connected to solar panels to relay the sounds of the forest to the cloud. And with machine learning, they’re now able to detect the subtle sound changes that indicate the loggers have arrived, such as the changing patterns bird song — so that they can alert the forest rangers even before they fire up their chainsaws.
It also raises big questions, though. Such as: will we all be out of a job in 50 years?!
The good news is that — so far at least — machine learning is about displacing work rather than replacing workers. We’re typically finding that people embrace machine learning because it gets rid of the most boring and repetitive parts of their jobs so they can move on to more interesting and strategic work.
And in terms of the larger context, it’s always easier for us to see the jobs that will be negatively affected than it is to imagine the new jobs that will be created thanks to these new technologies. The history of technology shows that we should be cautiously optimistic — the surveys we’ve done of companies that are adopting these technologies indicate that they’re generally hiring new people rather than laying people off.
But of course, as with any new technology, it does mean that some workers will need to adapt. The good news is that we can also use machine learning to help people in that process. You know when you buy something on Amazon, and it says “people who liked that product also liked this product”? Well, we’re doing the same thing for training: using all the data we have about employees, we can do things like “people like you who enjoyed this training course also enjoyed this course”, or “based on your career plans, here’s the most recommended next step”.
So we can potentially use machine learning to soften the changes due to machine learning.
What kinds of new jobs will be created?
We will obviously need lots of data scientists that can help create and guide these new algorithms.
But it will also help people do jobs that otherwise they wouldn’t have been able to do. Let me give you an analogy. A long time ago, if you wanted to work in a store, you’d have to be able to do mental arithmetic in order to figure out the bill. Now the electronic till does all that for us, and stores can hire people based on their customer skills, not how good they are at mathematics.
Artificial intelligence will make work more human – instead of rewarding people to behave like robots, there will be a premium on things that only humans can do, such as flexibility, creativity, and leadership.
How do we detect mistakes made by a machine and how do we protect ourselves against unintended consequences?
First, it’s really about sophisticated pattern matching, and so it never will be 100% perfect. But of course, people make mistakes, too, and we have systems in place to deal with that: audits, controls, and checks. With machine learning, we need to continue doing the same thing.
One important thing to watch out for is bias in the data – the machine learning algorithms will happily copy and replicate that bias when it makes decisions. This is why the best and easiest place to start is things like finance and logistics processes, where there is less chance of bias in the data.
Apart from machine learning, I’m hearing more and more about RPA or Robotic Process Automation — how does this relate to ML?
The easiest way to think about RPA is to imagine somebody that regularly has to do a series of steps as part of their job, across multiple systems. We can look over their shoulder and and record the different steps, so they can play them back any time they want. It’s clearly a big efficiency gain.
If machine learning is about automation inside an application, RPA is about automation across multiple applications. And of course, combining them both is even better – by integrating machine learning, we can introduce intelligent RPA that is easier to use, more flexible, and more robust.
If companies want to start using machine learning functionalities in their business process, where could or should they start?
The first step to identify any complex, repetitive decisions in your existing business processes. It’s an easy way to get the benefits of machine learning with little or no disruption to existing ways of working.
Second, it’s all about the data. If you have large amounts of high-quality information, you can implement machine learning fairly quickly. When projects go wrong, it’s typically because of missing, incomplete, or inconsistent data. Make sure you’re gathering the information you need now to benefit from these new opportunities in the future.
How should they deal with the ethical questions raised by machine learning?
First, I want to emphasize that machine learning is an amazing opportunity to make the world a better place, by more efficient use of resources, and by augmenting human intelligence. We’re working with organizations around the globe that are reducing suicide rates, improving health outcomes, and making farming more sustainable.
But machine learning is about delegating decisions to algorithms, and they don’t really understand what they’re doing. If you’re using machine learning in any way that has an impact on people, you must think hard about things like governance and transparency. Organizations should have an ethics advisor to help them make the right choices in this area.
At SAP, we’ve recently created an external ethics committee composed of ethics scholars and other experts, that can help guide as we implement machine learning in everything we do.