Salesforce talks about its Einstein Analytics product for financial services and new tools to combat AI bias

At TrailheaDX 2019, Salesforce’s Allison Witherspoon spoke with TechRepublic about the new industry-specific Einstein Analytics products and the new Trailhead modules designed to help
developers
tackle bias in AI. The following is an edited transcript of the interview.

Bill Detwiler: Tell me a little bit about the new announcements and new developments with Einstein Analytics.

Allison Witherspoon: Einstein Analytics, we’re really focused on delivering very tailored insights to every user no matter what your role, department, or industry. And we’ve really been on this journey since we launched analytics in 2014 to deliver those kinds of very role-specific purpose-built insights. And we started out by building a kind of analytics for every role—sales analytics, services analytics, B2B marketing analytics. 

We kind of switched gears over the past couple of years to deliver that same kind of mentality to industries so that we’re taking a vertical approach. And the first industry that we’ve launched an analytics product for is financial services so that every wealth advisor, every retail banker can see the things that they care about where they work, which is financial services cloud. So you can see things like deposits, loans, fees, client goals, withdrawals, assets under management… all very, very tailored made for financial services.

Bill Detwiler: And with analytics, there are lots of different types, right? You’ve got prescriptive analytics, you’ve got predictive analytics. Talk a little bit about what types of analysis that customers can get from the Einstein platform.

Allison Witherspoon: So with Einstein Analytics for financial services, it is the full spectrum of analytics, all the way from descriptive and diagnostic (what happened, and why did it happen?) to predictive and prescriptive (what’s going to happen, and what should I do about it?). And that’s really thanks to the AI-powered insights that are injected into our analytics platform. So when you purchase Einstein Analytics, you get Einstein Discovery with that product. And that’s our smart data discovery tool. So you get to leverage all of those predictions and recommendations in context. So what that might look like for Einstein Analytics for financial services, for example, would be now wealth advisors and retail bankers can do things like predict churn. Which clients are most likely to churn? Which clients are most likely to have large deposits and increase their assets under management? So very specific predictive insights for those folks.

Bill Detwiler: And what are the tools that the customers get to take action once they get those insights?

Allison Witherspoon: Einstein Analytics has a very rich action framework as we call it built into the platform, so right from any dashboard, right from the point of insight, you can take action back into Salesforce. You can do things like log a task, create an event, post to Chatter, communicate. So really increasing collaboration and communication from the point of insight back to Salesforce.

SEE: 60 ways to get the most value from your big data initiatives (free PDF) (TechRepublic)

Bill Detwiler: Can that be automated? So for example, if you’re a broker, if you’re an agent, and you get information that says, “Hey look, this client is likely to leave the practice, to take their money elsewhere, to move out because of a certain event or because the platform has predicted that,” is then there an action… then they are prompted automatically to take an action, right? So Salesforce, admins, and developers can set the system up so that the brokers and the agents are then prompted to take an action right then and there automatically, right?

Allison Witherspoon: The full breadth of analytics—all the way from descriptive to prescriptive—so we can have recommendations in the context of what action they should take. We like to keep that kind of decision up to the user. But we do have tools built into the Salesforce platform, like Process Builder, kind of workflow automation tools that allow you to set up those triggers, if you want, based on the insights that you’re getting. So because analytics is built on the Salesforce platform, you get access to all of the goodness of the Salesforce platform.

Bill Detwiler: You mentioned it a little bit earlier when we were talking about AI, talking about that’s what allows it to do a lot of the decision making. Talk to me a little bit about how AI is built into the analytics plan.

Allison Witherspoon: AI comes with the analytics platform now. So like I mentioned, when you buy Einstein Analytics, you get AI with that out of the box. And what we see with our customers is the line between analytics and AI is really blurring between BI and AI. It’s really becoming this kind of one intelligence experience. And that’s really what we’re striving for with Einstein Analytics is one UI, one UX where you get an intelligent experience start to finish, whether you’re looking at a dashboard with kind of historical information or predictions or recommendations or that kind of automation side by side. And so with Einstein Discovery, our smart data discovery tool, we’re powering the predictions and the recommendations side by side with more of the historic kind of learning.

SEE: Managing AI and ML in the enterprise (ZDNet special report) | Download the free PDF version (TechRepublic)

Bill Detwiler: And how much of a challenge was it, if at all, to integrate AI into the Einstein platform? With analytics, was it always the plan to have AI added into it? Talk a little bit about incorporating AI into the Einstein platform.

Allison Witherspoon: Yeah, so we’ve kind of taken a combination of an organic and an inorganic approach to the way that we think about building our AI. So we’ve had a team of data scientists that have been building the Einstein platform and a lot of the AI and machine learning models here for the past probably four or so years. But we’ve also made some acquisitions along the way, which I’m sure you’re aware of, that has allowed us to grow kind of in spaces that we maybe hadn’t thought about before that our own team wasn’t working on, especially with things like deep learning. So being able to take unstructured data and do things like image recognition, natural language processing, and now doing speech-to-text thanks to Einstein Voice. So it’s been a really healthy balance of kind of organic and inorganic.

And Einstein Analytics, specifically Einstein Discovery came to us from an acquisition in the space. So we were able to add that smart data discovery capability to our analytics platform once we were hearing from our customers that that was something they wanted. Because once again we’re a company that thrives on customer feedback, and we heard from our customers that they didn’t really care whether they were using analytics or AI or machine learning or deep learning—it was just this kind of sea of terms to them. 

Bill Detwiler: Tell me a little bit about the new sort of trust initiative that you’re building into the Einstein platform through AI.

Allison Witherspoon: Trust of Salesforce is the number one value, so it’s no surprise that we’ve kind of taken the same mindset to our AI product as well. And about a month ago, we launched a kind of series of features inside of our products, inside of our Einstein offering that really helped reinforce that message of trust. It means that it needs to be transparent. So we need to show the end consumer of the AI why predictions are what they are, and we call those predictive factors. So being able to expose predictive factors and being able to expose the underlying R code of a model if a user wants to geek out and kind of see under the hood. So that’s kind of the first piece is transparency.

The second piece is having responsible AI. So that means being able to prevent bias from entering models very early on, kind of whether that lurks in the model itself or the data that you’re feeding into the model. And so we do things like a flag for bias protection. When a builder of AI using one of our tools is building a model, how can we tell them which data fields might potentially lead to bias?

And then the third pillar is kind of accountable AI. And that is the idea that there’s this kind of feedback loop that’s happening, and how do we at all times show how the model is performing, expose model metrics in a model scorecard so that the user can determine if they want to actually turn that model on, deploy it, and expose the predictions to the end user?

So that’s kind of the idea of trusted AI, and we’re actually shipping a lot of those features across the Einstein offering. And just this week, we launched our responsible creation of AI Trailhead because we really believe that no matter what your skill set with Einstein we’re actually enabling kind of citizen data scientists to build AI, to build custom AI. And so if you don’t have a PhD in data science, how do you start building AI models that are free of bias, that are fair, that are trusted, that are ethical? And so this trail is a great place to start to understand what is AI, where can this bias be lurking, and to kind of skill up and educate yourself so that you build models that are truly fair and trusted.

SEE: Salesforce handily tops Q1 earnings targets (ZDNet)

Bill Detwiler: So most AI, most algorithms, are only as good as the data you put into them. And a lot of organizations have data that’s spread all over the place. It’s disparate, it’s not clean. So a lot of organizations when they go to implement an AI project, they need to do a lot more data cleansing, data management, data organization than maybe they think. What are the tools within Einstein to help organizations do that, or are the integrators helping them do that as they deploy the product to the customers? How does that part of the process work?

Allison Witherspoon: With Einstein analytics platform, one, we can bring in data from any source. We have a library of connectors that allows you to easily connect to third-party data sources, because we hear from our customers all the time their data doesn’t just live in Salesforce—it lives in ERP systems, HR systems, finance systems—and so we need to make it easy for them to bring the data into Salesforce. So that’s the first step.

The second step is the data prep process and really creating kind of smart data prep tools so that our Salesforce admins or data analysts, whoever are using Einstein Analytics, whatever persona, they can actually clean their data, cleanse their data, transform their data with kind of AI-powered recommendations. So we’re bringing the AI to the data prep process where we’ll actually do joins and fill in missing fields and all of that kind of stuff very intelligently to make that process that much smoother.

Bill Detwiler: Okay. And when you’re talking about the ethical use of AI and inherent bias built into the equations, what are the triggers within the tool? But beyond that, are there flags and triggers within the tool to help them know, “Okay, when you’re bringing in this dataset, it may give you information that you want,” for example, but it also may cause unintended bias to creep into the system as well? So how does the system do that beyond what you’re doing with the new Trailhead in terms of training?

Allison Witherspoon: And I’m glad you brought up that point because I really do think it’s kind of a two-pronged approach. We have to do the education, but we also have to build these triggers into the product, and that’s exactly what we’re doing. With Einstein Bias Protection in Einstein Analytics, we can actually flag for fields when you’re building the model that might lead to potential bias. So what would happen is the builder of that model, whether it’s an admin or an analyst, would actually set up fields that they think could potentially lead to bias, like zip code, like gender, like race. And then Einstein Discovery will actually look through all those fields and find similar fields, correlated fields that could potentially be a proxy for those fields that have already been flagged by the user. And you’ll actually get kind of little exclamation marks in triangles that pop up in the product while you’re building the model that flag you for that potential bias in the model creation process.

Bill Detwiler: And with financial services, there is a lot of bias already inherent in the financial services system.

Allison Witherspoon: Yes.

Bill Detwiler: You mentioned zip code.

Allison Witherspoon: Yes.

Bill Detwiler: How are the customers who have maybe always used zip code or their clients already come from a specific zip code or specific zip codes geographically located around them or whether it’s national? So how hard of a job is it to convince the customers not to rely on those fields, which have been predictive in the past, which have worked for them in the past on a purely ROI or a purely financial or purely to the bottom line but have also incorporated bias in the system?

SEE: Salesforce’s Parker Harris offers CXO insights for managing complex IT integrations and handing the inevitable crisis (TechRepublic)

Allison Witherspoon: And this is a great example because in financial services specifically, we see a lot of this kind of bias perpetuated in models, especially with zip code, not giving loans to certain individuals because they come from a zip code. And back to this kind of the point of these two tracks, we can educate on Trailhead, we can put these prompts in the product, but we can’t force people to build a model a certain way and we’re never going to. 

The best that we can do is educate our customers at every turn using avenues like Trailhead, using our platform. We have an amazing, amazing colleague here at Salesforce named Kathy Baxter—she’s our architect of ethical AI. And so she’s really passionate about this, but building these prompts into the product, educating folks on Trailhead, but not kind of being that Big Brother that’s going to automatically build a model for them, automatically take out certain fields, that will never be kind of our stance.

Also see