How Spotify uses AI to sound out new efficiencies
Artificial intelligence (AI) is transforming the way we work across industries. From back-office finance departments to front-facing customer support, AI's power lies in its ability to discover and introduce efficiencies. In unison with human experts, its potential knows no bounds. This is exactly where Sidney Madison Prescott, global head of intelligent automation at Spotify, and her team of engineers and solution architects come in.
In our TELUS International Studios interview with Prescott, she explains how her team seeks out Spotify's most mundane, repetitive tasks with the express purpose of establishing new efficiencies. Prescott deploys a wide range of technology to get the job done, including robotic process automation (RPA), machine learning and AI. This quest for intelligent automation and simplicity drives greater data quality and can help ensure that Spotify employees — or "Spotify-ers" as Prescott refers to them — are focusing on tasks and projects that add value. Enabling these efficiencies means a quicker time to market and new and improved products and features, according to Prescott. Ultimately, it all comes together to better the customer experience (CX) of Spotify users.
Using AI to introduce efficiency
Perhaps the best way to learn how Spotify uses AI to improve real internal processes is with a concrete example from Prescrott:
"One example is when we go into the Spotify application and we look at all of our different audio ads that run within the application, we have to test those ads to ensure that the audio integrity is there. This is a process where we take each ad and we test to make sure […] that everything is up to spec for that particular advertisement. This is a manual process, but we have been able to automate this process and leverage robots to actually check each advertisement to make sure that it's running properly, that it is showing the right or the correct specs in terms of the visuals and that the audio levels are appropriate. This is a really perfect example of something that you would wouldn't think that humans are doing today, but in reality, it's a very integral part of the responsibilities of the team."
What is intelligent automation?
We know that AI's effectiveness depends on the sourcing of large amounts of data that is used to train AI algorithms to complete tasks. A significant amount of time and effort must be spent by data experts to ensure they're not only selecting the correct kinds of data, but also labeling it appropriately. When all is said and done, what does "intelligence" in the context of AI actually mean?
"From my perspective, in terms of the team that we have today and the technology that we are leveraging, we really are looking at the distinction between: Is a particular automation cognitive, in that it can facilitate different ways of learning? And is it able to recognize different characters? Does it have facial recognition? Things of that nature.
"Now, there's a separate part of the intelligent automation piece, which is also the robotic process automation [component]. And those robots are not cognitive. Meaning that they cannot execute anything that has not been predefined in their developer build workflows."
Prescott posits that intelligent automation really "comes together" when RPA is combined with AI or machine learning engines. "So it really is the amalgamation of these different tools coming together to be able to facilitate a specific set of business outcomes, which we would typically relegate to, let's say, human intelligence rather than a machine," explains Prescott.
Make strategic decisions with the full picture of your digital workforce.
Is AI really learning, or just memorizing?
The human annotators who manually label data play a critical role in ensuring that AI technology understands its training data samples. With such an understanding, patterns can be learned to recognize similar patterns, predict future pattern results, correct false assumptions and even build extensive vocabularies. Without these humans leading the way, no machine learning algorithm would be able to accomplish such a feat. Which leads to the question: Are machines really learning, or are they just memorizing? Are we able to tell the difference?
"It really comes down to how you break out the different categories of machine learning, deep learning, and artificial intelligence. So if you think about machine learning, it really is a subset of artificial intelligence. And so it's basically creating systems that can have the ability to learn and improve over time with experience. Now, machine learning models need a large amount of data very, very much like the human brain. We need a lot of data continuously over a lifetime in order to build up our understanding of the nuances of, say, a chair or the concept of education, the concept of computers. And so machines are very much the same. We are able to basically build up their understanding of a particular concept through the amount of data.
"I think one of the other things I would say is in terms of learning versus memorization: I believe it's a bit of both. And we do this as humans as well, whereby we learn quite often by memorizing. And it's very interesting, even if we think that we aren't memorizing something, we are. So, for example, memorizing where a particular Starbucks is located on your corner or memorizing the way to get to a specific location — that is a form of learning. And I believe that machine learning can basically facilitate the same levels of understanding, whereby it's a component of a little bit of memorization, but that memorization actually enables the machine to then learn different patterns, learn different potential outcomes that will help to solve problems.
"And so the last part of that is when we start moving into another subset of machine learning, which is really your deep learning. And that's where we are attempting to create a very similar computer network that mimics the neural networks of humans. Deep learning is where we get into the potential ability to delve deeper into a little bit less memorization and a little bit more learning of patterns again from a wide variety of data that is fed into that particular neural learning network," says Prescott.
The challenge in the future of machine learning and AI will be for engineers to enable machines to do a better job extrapolating from their datasets. Instead of matching memorized patterns or symbols, newer machines may be able to create abstractions based on the patterns they are taught or learn. To get there, humans remain critical by providing a continuous loop of feedback. Designing for this harmony can ensure that the small improvements made behind the scenes can make a big impact to an organization's back-office operations, CX and beyond.
To gain more insights on the impacts of intelligent automation from Sydney Madison Prescott, check out the episode of TELUS International Studios podcast.