Video Demo: Using AI and Machine Learning with AEM to Impact Customer Experience
Tim Donovan, October 14, 2021
This year, Adobe Summit was held virtually. 3|SHARE Managing Partner Tim Donovan and Roche Pharmaceuticals' Digital Strategist Joerg Corsten teamed up to deliver presentation that included demos for three user types using Roche's use cases and explored the question many businesses have around whether or not to use AI and Machine Learning to change the CX and how to do so.
For each demo, Joerg introduces the website and Tim offers a look at how this technology can be used by from the perspective of the CMO and AEM Content Author, and how looking at the emotional responses of site visitors can inform the business so it can make better content decisions.
WATCH THE PRESENTATION
Tim Donovan (3|SHARE): [00:04:13] Thank you, everyone for joining us today for our session on AI, machine learning and customer experience. Before we get started some introductions. My name is Tim Donovan. I'm a managing partner at 3|SHARE. We've been delivering Adobe projects for around 10 years. Now. One of our biggest partners is Roche Pharmaceuticals.
We have implemented over 25 sites for them to date. And we're lucky enough to be joined today by Joerg Corsten, from Roche Digital Excellence.
Joerg Corsten (Roche): Thanks Tim, for having me with you today.
Tim Donovan (3|SHARE): It's a pleasure. Thanks, Joerg. And one last introduction. I should mention that this is not in fact my cat or my dog, but what you see here is an attempt of machine learning to classify this face.
And it's quite confidentit's a dog. But I'd like you to set this particular example aside as we look further at machine learning, AI and its possible applications for [00:05:13] the customer experience and customer journey.
We've been collaboratively rolling out websites with Roche for a number of years now. And with AI and machine learning an increasingly hot topic, we thought we'd do some blue sky thinking with Roche around how it could be used to impact their customer experience and journeys.
So while diving into the realm of decision trees and neuro-linguistic programming, we realized there were three key questions:
Should you actually change your customer experience based on AI? Are there practical and impactful applications to help people across the organization and not just your data scientists? And what are some of the common pitfalls to avoid when looking at machine learning and AI?
We realized when looking at improving the customer experience and journeys, it'd be best to frame it from three different perspectives.
The implications of using machine learning and AI are far reaching and the applications of the technology has the potential to support many people in different ways. So we've split this [00:06:13]presentation into three personas.
First CMOs, faced with an overwhelming choice of technologies that have the potential to drastically shape up customer experience orchestration.
Secondly, Content Authors increasingly pressured to generate timely, but personalized content to support a delightful customer experience.
And finally, Site Visitors whose attention is increasingly under attack and these days expects a personalized experience throughout that brand interactions and journey.
For each of these personas, we've also created a small illustrated technical example, hopefully bringing to life what we see as potential applications of the technology.
So first, CMOs. When we talk to CMOs they're quite often excited about how AI can help them in particular with their CX journeys, but they soon become overwhelmed with the sheer number of possible applications for the technology.
The first opportunity is always to ensure an understanding of [00:07:13] exactly what machine learning is; that is giving machines access to lots of data and letting them learn. And the differences to AI, which is a much broader domain; looking at getting machines to carry out tasks autonomously in ways we'd consider smart.
The biggest challenge for CMOs however, is where to invest. Investing properly in AI right now is not cheap. It can be a big bet for CMOs. So where do they place that bet to ensure the best return for their dollars?
This is where CMOs were investing their spend last year. And this is just the top seven. There are countless other AI related areas they're investing in. You can see it's fairly split between these personalization insights and targeting. This firehose spending approach to AI is not uncommon at the moment.
Where we've seen expensive failures is when bets are placed on certain AI technology without even considering the overall strategy. A great example is conversational AI, in [00:08:13] particular building skills for devices like Alexa. This has had a lot of investment in the last few years but we've struggled to find any organizations really making a return on their investment. So when considering AI, it is extremely important that a CMO focuses on strategy first and then assesses what practical applications of AI should drive the strategy, not the other way around.
So, where do we see the best place to make that bet in the coming years? Deep fakes have been in the news a lot recently and some of you may have noticed in my previous slide, the deep fake of Nicholas Cage as the Tiger King. Indeed for legal purposes, I have to mention that none of these actually are Nicholas Cage.
Why do we see generative adversarial networks, aka, the process that produces deep fakes as particularly impactful to customer experience? Well, looking at budgets, content creation is one of the most costly areas for a CMIO. If we can harness this process correctly and have machine learning and AI directly create impactful content for the [00:09:13] customer experience, the financial impact and savings for CMOs is potentially huge.
Now onto our first demo and over to Joerg to introduce the first Roche site we're going to take a look at.
Joerg Corsten (Roche): Thank you Tim. So one of our websites is called ForPatients. This is our platform where we support patients and caregivers whilst they're suffering from a disease and whilst we need to take critical disease treatment decisions together with their doctors. It is of crucial importance that patients can access complex information
in a more simple way in order to feel empowered to take these decisions. We want to support them with content that is written in the lay and patient oriented way that is meaningful for them and is not perceived by physicians to be intruding their key competencies. So this is our answer to the patients. But Tim, can you tell us more about how we do this?
Tim Donovan (3|SHARE): [00:10:13] Absolutely. Thanks Joerg. Okay. So let's take a look at the site in particular, the disease area overview. Let's see what articles are actually available. As you can see, there's a lot of great content on the site for patients to find out information about various diseases. We're going to take a look at the hemophilia information page in this example.
For each of the diseases, content authors have produced some fantastic articles, painstakingly, written, and curated to be easily understood. And you can see here not only is it written content, but each disease area is illustrated with nice diagrams to help your understanding. But how can we support this content production and creation process?
Let's jump to our AEM instance. First, a bit of background to this. To drive this, we've created a standard AEM cloud configuration item. And this is what connects the AI content generation service. We did it this way because when looking at services to perform the actual content generation, there are quite a few [00:11:13] choices available
so we have multiple options. This configuration allows us to quickly switch between the different services available and quickly test the results in situ. Here we are in the standard AEM content tree and we're going to take a look at authoring the hemophilia page. So you see the standard author will page in Adobe Experience Manager here made up of . Components that were 3|SHARE implemented for Roche.
One of the standard components is the rich text editor. So how are we exposing the deep fake content generation type functionality to authors here? We've extended the normal rich text component with an inset AI content button that calls our backend service. What we ask authors to do is simply seed the text component with at least a few sentences.
And then the more the author can provide, the better the results. So here we see the author has entered just two sentences and as then going to click the insert AI Content [00:12:13] button. At the moment, we're dependent on third party services to actually do the content generation. So there's always going to be a bit of a delay returning that content.
And at the moment it's fairly generic text that's returned. What we're hoping to do in the future is used locally powered machine learning and AI that is local to AEM to generate the content rather than relying on cloud services. In this way, not only as a content generated more likely to be relevant to Roche and written in the appropriate tone, but hopefully it will remove any delays. I will get the content actually generated faster. And there we go, some content has been returned for the author without them having to put in too much.
Right now it's a long way from directly producing actually usable content in every situation. And sometimes the text that comes back is actually jibberish. It still requires a degree of review and editing. And of course, any medical information will always need to go through approval before release, but it certainly can be useful at assisting content authors with [00:13:13] ideas or prompting them when they get writers' block. And that's the end of our first demo. Hopefully you can see from this, that when CMOs look at their budgets in future, less money will need to be allocated content generation and more can be invested in other areas of the customer experience and customer journeys.
Now on to Authors. The site Author's content comes first. Or maybe coffee comes first and then content. Everyone has heard of content velocity being important to organizations. As we've heard, content creation is a huge barrier for many of our clients. And what can we do to help overcome this? We realized we needed a way to ensure what authors are producing will result in a great customer experience.
And one way to achieve this is to include them in more feedback loops. But how can we do this and how can machine learning help? We took a look at sentiment analysis. It's been around for a while now. Indeed some of the earliest scientific work on analyzing exactly what [00:14:13] people say and understanding the sentiment from it dates back to publications as early as 1969.
But as we looked at applying sentiment analysis to some of the feedback Roche get directly from site visitors, we started to see some of its shortfalls. Generally it analyzes text to give you a score ranging from one to minus one, or it could be diluted further. So you might get positive, neutral, or negative, something similar. But it's not always so good at understanding the emotions of the feedback.
Consider this review of a page on a site. "This page looks great." Okay, sounds super promising. And then I actually read it. Well, I wish sentiment analysis understood sarcasm, but it's not quite. there yet. Nevermind. We can probably just ignore this user's comment. "I will find where you live." There's clearly a lot of emotion conveyed in this feedback,
some of it not so subtle. See sentiment analysis is good for a general idea of what users are experiencing along the customer journey, but it won't necessarily [00:15:13] convey the emotion. How can we address this? It's time to get emotional AI and machine learning can now be deployed in ways rather than simply labeling customer feedback as good or bad.
It can create categories of emotions from the data. This is known as emotional analysis. On the right, we see Plutchik's wheel, developed by famous psychologist, Robert Plutchik. And it maps out the various fringes and degrees of human. What we need is customer feedback mapped onto this wheel so we understand the actual emotions your customers are conveying to you and make adjustments based on it.
For example, if the feedback you're getting is your customers are bored, then you may want to invest in some more exciting content. If the feedback's excitement, then you might want to continue what you're doing and maybe promote the Content Author responsible. So to the question, should I change my customer experience based on AI? Absolutely. If the changes are the result of meaningful feedback.
For our next demo. We're going to look [00:16:13] now how we could possibly surface some of this information to content authors at Roche.
Over to Joerg to introduce our second Roche site.
Joerg Corsten (Roche): Thanks Tim. So Medically is an important building block for us. Every year we conduct multiple readouts of all clinical trials.
We release protocols, infographics, posters, and full scientific papers to physicians. And so those things are an important part of a crucial scientific exchange to up-skill doctors and to enable their scientific exchange. We do cover between 80 and a hundred medical congresses a year. And we want to tailor the experience that we offer on this hybrid platform to the needs of the physicians. Based on these interactions, we want to learn and further improve how we present these information to them. Over to you, Tim.
Tim Donovan (3|SHARE): Thank you very [00:17:13] much. So when you visit this site, this is a pop-up that you'll see if you linger on the site for about three minutes or more. And it asks you for feedback on the particular page, you're looking at. Things like if the content is useful, if you'd recommend the content, and then it asks for your thoughts about the page. It's this freeform text field that we'll use to drive the contextual and emotional analysis.
At the moment, the results of this form go directly to Survey Monkey and the content authors don't always get to see the feedback yet, so they're not involved in that important feedback loop. Again, back to our AEM author instance, this time of course, taking a look at the Medically site. We're going to take a look at the neuroscience section of the site and some of the user feedback about conferencing pages. Let's jump into the list view.
What we've done here is taken the feedback from the form and passed it through some sentiment analysis processes. In this example, we are simply showing the [00:18:13] authors a score between zero and five of how well that content has been received which is useful but limited. So we began to experiment with other ways Content Authors could visualize the information.
We switched to showing emojis for awhile. This is a bit more accessible for content authors, and it's quite obvious how well your page has been received based on the associated emoji. But ironically, there's not so much emotion behind the emoticon. This is where we switched to doing emotional analysis.
Here you can see, again based on site visitor feedback, the emotional sentiment and categories that are based on Plutchik's wheel of emotions. This is far more accessible and useful to an author than a simple sentiment score and allows the author to investigate why the content made the visitor feel that way. We can clearly see here that there is one page in particular that needs some attention and Site Authors might want to investigate why this page is having such a negative effect on the customer's [00:19:13] emotions and experience.
We then thought about other possible uses for this. So we're looking at doing something. We are calling it, lookalike modeling. You may have heard this term used in analytics before but we're now going to apply it to the content. By using machine learning models am we'll be able to deduce which pages are similar.
We can already do this today with images. So why not entire pages? In this example, we can see that the page on the left is very similar to the page on. And if I'm busy as a Content Author, making changes to the page on the left, why not have AEM provide feedback on what types of emotion your new page may elicit based on the results from similar content?
This is exactly what we're imagining. Here you can see while the Content Author is busy creating the content, we're notifying them that their new page looks similar to another page, how similar it actually is, and assuming the other pages already live showing the emotional feedback the other [00:20:13]page has received.
And that wraps up our second demo. What you've seen as a great example of how we can use the feedback from AI and machine learning, even before a page has gone out to an audience, to sharpen up and change the content in ways that improve the overall customer experience and customer journey.
And finally on to Site Visitors. When we started talking to Roche and other organizations about how to meaningfully impact a customer's experience, we realized there were two key areas to address.
The first was around targeting individuals and not audiences.While many organizations do some degree of personalization or targeting often based on decision trees that whittle down a customer base of millions to still a few hundred thousand, few organizations are really doing hyper personalization at scale.
And this is a problem. 52% of consumers unlikely to switch [00:21:13] brands if a company doesn't personalize to them. That's half a customer base leaving. If you don't start communicating in the right way, in the right tone and at the right time. Not addressing this can be catastrophic for your organization. So the need to target and personalize to individuals, not an audience becomes key and machine learning and AI promises to address this.
But then we took a step back and realize what really drives machine learning. At its core machine learning relies on big data, huge sets of data that models can churn over to arrive at a result. And therein lies our second challenge to address.
Siloed data across business units continues to impede AI usefulness in CX. Roche is not unlike many organizations when knowledge, information and data is siloed across multiple business units. If you can't appropriately harness this information it's application in machine learning and AI are limited. Machine learning in particular is dependent on large amounts of clean [00:22:13] data.
That is data that doesn't contain errors and is in a suitable format to be ingested for learning algorithms. Often organizations undertaking machine learning projects spend most of the time simply going through the unglamorous exercise of sifting and cleansing rows and rows of information.
Furthermore with the cloud, cheaper than ever storage costs and more ways every year to capture information organizations are becoming overwhelmed with too much data.
Where do you begin when your organization's collective information consists of 5 million PDFs and 2 trillion data points about your customers? How do we go about turning big data into something useful, those insightful tasty nuggets that will drive customer experience? Well, the answer is with more machine learning and AI. We realized at Roche, if we can use machine learning and AI to ingest the enormous amounts of data and information they hold and turn it into something specifically tailored for individual site visitors, we can unlock [00:23:13] the value in that big data and put it towards improving the customer.
Now over to Joerg to introduce our third and final Roche site for today.
Joerg Corsten (Roche): Thank you, Tim. You know, as you probably spotted already, we're learning a lot about the content needs of our customers and we're trying to address them really upfront as much as possible. We provide a lot of information online. Um, but still as it can think, there is always one more question a customer have may have for us. At which temperature do I need to store a medicine? Is the product lactose-free? You know, is the medicine safe because it looks different compared to last time I've used it? So roughly we receive around 180,000 of such questions worldwide a year. And I think we want to offer all of the customers an easy way to leave the question with us.
But just think about it. [00:24:13] Wouldn't it be great if we could already answer the question at the moment where having this question? And this is actually the hook that we tried to test out now and think about how we can do that in a better way. Maybe Tim, can you show what we are planning?
Tim Donovan (3|SHARE): Thanks, Joerg. So here we can see one of the forms that capture some of those 180,000 questions that make their way to Roche employees
every year. You type in your questions, select who you are, physician pharmacist, patient, et cetera. And it will fire off an email to somebody at Roche. We'd very much like to incept at this point in the customer journey and see if we can instantly help the site visitor avoiding the need for any email to be sent to Roche at all.
And this is Roche's big data challenge slash opportunity. For every medicine Roche releases to the market, there are literally thousands of PDFs available both internally and externally with all sorts of information written from various perspectives here, [00:25:13] we see just some of the commonly available ones from patient information to trial results to Congress papers covering the science in deep detail.
None of it's currently easily surfaceable on the website as anything other than PDFs.
To solve this silo data problem, we are investigating the Adobe PDF extract API. This leverages Adobe Sensei and it's rich machine learning and AI capabilities to not just blindly extract the text from PDFs, but to actually understand the content structure and extract information in smarter ways than possible before. This is perfect for the sorts of data that has embedded in PDFs in particular, when it comes to tabular data.
We can then index that information, present it in pretty much any format and use it to drastically improve the customer experience.
What we're hoping to achieve is by having all the PDF data available, we can do some early interception of site [00:26:13] visitor questions. Hey, for example, we've extracted a PDF to Jason, so it can be now used to power a recommendations.
We can see that based on the question of site visitors asking instead of an email being sent to someone at rush, we can now automatically suggest sort of appropriate responses to that question. Hopefully this early interception will avoid an email being sent to Roche at all. And this concludes our third and final demo.
As you can see if there's one major thing AI and machine learning can do for your customer's experience, it's solving your big data problems to ensure your customers, or in fact, your employees have access to the right information at the right time.
And that is the end of our presentation for today. So here are the three key takeaways. Number one: take a strategy first approach to picking the right AI and machine learning technology. Don't spray and pray with your budget plan strategize, then [00:27:13] invest. Number two: provide real time feedback to Content Authors to improve the way they shape the customer experience and customer journey. This can drastically uplift the content that is introduced throughout the customer journey and results in better content, even before it goes. And number three: unlock machine learning usefulness by getting a hand on big data across your organization. Get a handle on your data today. If you really want to unblock its usefulness in machine learning and AI tomorrow,
AI and machine learning presents some of the most exciting opportunities in our industries. And I'm eager to hear and see what you all do with it in the future. If you'd like to contact us, here are our details. Please reach out to us on social media or email with any questions. Thank you to Joerg and Roche for joining us today.
And thank you audience for your time. Take care.[00:28:13]
Get in Touch
Questions about this or do you have a problem 3|SHARE's Adobe Experience Cloud experts can help with? Contact 3|SHARE and we'll put you in touch with the right person. Just visit our Contact page and submit the form.
Tim Donavan is a Managing Partner at 3|SHARE based in the United Kingdom.