A few years ago I fell in love with data. I learned Python and started playing around with machine learning algorithms — working to understand not just the technology but the math behind the various algorithms. As a physics major, the math was fascinating. As someone two decades past college, however, the math was also really hard. Thankfully, having a rock solid understanding of the math underpinning machine learning and artificial intelligence isn’t critical to applying the technologies to marketing and the customer experience. What matters is taking action and putting these technologies to work for you and your customers.
I had the pleasure of attending O’Reilly Media’s Strata Data Conference a couple of weeks ago. The conference is part of a group of conferences O’Reilly hosts covering data related topics. Per the conference web site, programming for the Strata conference is curated to emphasize how organizations can put “big data, cutting-edge data science, and new business fundamentals to work.” In truth, it’s almost impossible to discuss big data without addressing analysis — how the data turns into insights. For marketers and everyone else focused on the customer experience (CX), those insights originate with machine learning (ML) and artificial intelligence (AI).
I want to use this space to share some key takeaways I got out of the conference.
Here’s the TL;DR:
- Ethics Should Be a Principle Component of Your AI Approach
- You Don’t Have to Start at Zero — Stop Making It So Complicated
- Data Talent is in Tight Supply — You Will Have to Train Your People
- AI and ML bring Add More Value to Your Web Behavior Analytics
- Desktop and Mobile Browsers Have Enough Power to Run AI
- Chatbots Are Great for Specific Use Cases But Users Aren’t Happy about Them
Ethics Should Be a Principle Component of Your AI Approach
This isn’t the sexiest takeaway but it is, arguably, the most important. In the rush to make the most of the data they have, companies have implemented policies and procedures that are not mindful of the impact on people. The AI algorithms you build to predict purchases, assess customer risk or decide who gets interviewed for a job don’t live in a vacuum. They impact peoples’ lives.
The biases human beings harbor make their way into the artificial intelligence models used to make decisions. This happens both in the way sample data is collected and the choices made in how algorithms analyze that data. Moreover, one of the challenges of AI is so much of what is done with deep learning — a foundational AI technology — is done in a black box. There’s little transparency which means there’s not a lot of oversight on how models come to their prescriptive and predictive conclusions.
It needs to be said that ethics in data is not simply a function of assuaging consumers. Bias negatively impacts the performance of your models. When your model over penalizes race or wrongly classifies customers because a bias, it exposes you to unnecessary risk. That’s not to say bias won’t make its way into your models. It will. By being mindful of its potential, however, you limit the impact on your customers and your company.
You Don’t Have to Start at Zero — Stop Making It So Complicated
As I mentioned above, there are facets of machine learning and artificial intelligence that can be difficult — math being a big one. Some others include computing power, bandwidth, time to analyze and talent. I’ll touch on the last of those in the next takeaway but the others — including the math — have been solved for you with automated machine learning (AutoML).
AutoML, as defined by Wikipedia, is “the process of automating end-to-end the process of applying machine learning to real-world problems.” Rather than having to rely on experienced data professionals to assess and prepare and analyze data, AutoML allows people with limited experience to attain insights from data. It accomplishes this by providing tools and resources people can use to automate the analysis of data
Reliable AutoML platforms exist for you and your team to begin learning how to extract insights from the data (e.g., transaction data, web behavior, social media data, etc.) you have. Google, Microsoft and Amazon all have well-regarded AutoML platforms. Additionally,, companies like H2O.ai and dataiku also have AutoML software that you can use in-house to make inferences about your data.
The tools are there. You just have to get started.
Data Talent is in Tight Supply — You Will Have to Train Your People
The data scientist role is a fascinating one. Where many specialist jobs require a single specific skillset, the data scientist needs multiple skillsets. They must have an expansive knowledge of advanced mathematical concepts, a strong computer science capability and extensive knowledge of the industry they’re working in. To have one or even two does not make one a data scientist. It is the combination of aptitudes that makes people successful in the role. The thing is, these people are unicorns. Well, unicorns don’t exist. Data scientists do but they’re very difficult to find. If you’re waiting to start your data analytics journey on finding a data scientist, you’ll likely be waiting a long time.
All signs point to retraining existing employees as the best way — in the near-term — to build out a data talent pipeline. Your existing employees already have the domain expertise so that’s a plus. With training, motivated employees can learn the computer science capabilities needed to effectively analyze data and create models that can be useful in your business. Some employees may even may be able to learn the math but thanks to AutoML and its related technology, it’s not necessary.
One of the major benefits of training existing employees is, when you actually are able to hire a data scientist you’re more aware of what they should be doing. Data science work is sexy but it can be boring when companies are unable to provide the scientist with direction. Starting with the staff you have has the potential to create a better environment for everyone.
AI and ML Add More Value to Your Web Behavior Analytics
Companies have been very successful with using web analytics tools to monitor user and identify customer behaviors that lead to improved experiences. ML offers even greater opportunity for mining insights from that data.
Web analytics tools are invaluable tools for understanding how your customers use your site. They’re excellent at showing you what users did over some specified period of time. They’re not as good at identifying trends, explaining causation or predicting what users might do in a given situation. They’re not supposed to be good at those things. That’s what machine learning is for.
The platforms I pointed to earlier can help with this. Companies who use Google Analytics 360, for example, can export Google Analytics data to Google’s BigQuery, their analytics data warehouse. From there, machine learning algorithms available in Google Cloud can analyze the data to extract information that can help you make smarter decisions about your customers. This tutorial from Google (it’s technical but what’s important is how powerful the capability is) explains how they use the process to predict which content items should be used to personalize content for a user.
Desktop and Mobile Browsers Have Enough Power to Run AI
In one of the sessions I attended, the speaker made the point that many of the machine learning algorithms we rely on so heavily have existed for decades. Knowing what to do wasn’t the problem. Rather, the technology wasn’t robust enough to handle what needed to be done. Obviously, that has changed.
In fact, that change is greater than we realize. Thanks for the tremendous power of the computers in our pockets — our phones — ML and AI algorithms can be executed right in your users’ browsers. This is an important advancement.
Executing machine learning in the cloud introduces significant risks. There are privacy implications. There’s a cost associated with the resources you use. Additionally, there’s no guarantee the network will be reliable enough to support exchanges between your cloud and the users’ machines. Finally, even if the network is reliable, it may not be fast enough to facilitate machine-to-machine conversations as quickly as you might need. Being able to perform machine learning on mobile removes those obstacles.
Operating on the browser opens up new capabilities. With mobile deep learning, it becomes possible to detect objects (e.g., hands, faces, real-world objects, etc.) from the phone, tablet or desktop. Moreover, it also becomes feasible to personalize app functionality based on previous user behavior even if the user is in an area with an unreliable data connection.
Chatbots Are Great for Specific Use Cases But Users Aren’t Happy about Them
My bank has a chatbot that they really want me to use. When I login they ask if I have a question for her, Erika. When I end my session, they remind me she’s there. I’ll be honest. I don’t want to chat with her. I really don’t have any questions. More than that, though, I don’t trust her. I’m not alone.
A conference speaker reported of a study that indicated users don’t mind chatting with a chatbot as long as they don’t realize they’re interacting with a robot. People want to interact with — get this — people. That said, chatbots are plenty tempting…
Natural Language Processing (NLP) and Natural Language Understanding (NLU) have made huge strides in usability. It is those technologies — processing users’ queries and providing context around those queries so as to demonstrate understanding — that makes chatbots so useful. That usefulness makes chatbots very valuable for creating efficient support flows that potentially reduce both the time needed to resolve issues and the usage of human customer service people. All signs indicate, however, that the use of chatbots should be limited to basic use cases.
Humans are complex and the way we express our needs varies. Consider all of the ways someone can ask for support for their toaster-oven:
- Where can I find the manual for my toaster?
- Are there instructions for my toaster?
- Is there a book for my oven?
- I need a manual
- How do I cook a chicken in my toaster?
- Can I make bread in my toaster?
Users expect chatbots to be able to successfully respond to all of these queries. They expect the bot to understand every intent and to provide them the right answer regardless of how they ask the question. That’s hard for humans. It’s even more difficult for AI because it has to be taught what each individual query means before it can answer it properly (that hardly seems “intelligent”, huh?). By focusing on specific limited use cases, you increase the likelihood that you can protect your brand by meeting your customers’ experience expectations.
It may not be totally obvious given how advanced these technologies seem now but we’re actually at the infancy of using ML and AI to impact the customer experience. Naturally, personalization — true 1:1 marketing — is a huge part of what’s possible. That’s not the end of the story, though. These technologies empower marketers to infer a tremendous amount of information about their customers and their behaviors — information that will lead to more optimized selection of media channels, better targeted content and increased revenue.
Imagine being able to estimate a customer’s potential lifetime value when they first arrive in your store or on your site. How would that alter the experience you provide to them as they proceed through your funnel? What would it look like if you were able to correlate a leading indicator sentiment on social media to a loss of a customer and create an inflection point where you change the dynamic of that relationship? What if you were able to establish that people from a specific zip code who visited your site when the weather dipped below 50°F and then were sent a 20% off coupon were likely to spend 75% more than when the temperature was higher? How about if you could let a use gestures in his phone camera to navigate through your shop with augmented reality — bringing a real-world experience to a virtual environment — all while creating a “store” for the individual?
These concepts aren’t fantasy. They’re very possible with ML and will become more common. The question is, who’s going to do them first — you or your competitors?
** getting on my soapbox **
Did you know that most “unconventional innovation” occurs in cities? Lots of patents flow out of the suburbs but a not insignificant amount of research indicates the big, disruptive innovations are birthed in cities. The idea is, the close proximity of diverse people in cities cultivates new ideas and encourages a cross-pollination that leads to more impactful ideas … and execution.
What’s my point? What does this have to do with the Strata conference? There is amazing work going on with AI and ML. If you’re reading this, then you need to be out at conferences like the O’Reilly ones and others (like the upcoming AI Summit in NYC) where you can meet practitioners who are doing some of this work and bring that back to your company. The cross-pollination that drives innovation isn’t likely to occur if you’re only talking with people in your organization. So, be a busy bee and get out there, learn, share and be a part of the process that is so key to innovation.
*off my soapbox*