Salesforce Has Cracked The Code On Demystifying Machine Learning For Mortals (includes audio)

The categories of computer technology known as artificial intelligence (AI), machine learning (ML), and data science have a serious public relations problem. Despite their ability to be game changing for many organizations and individuals — I mean mind-blowingly game changing -- they live on the fringes of esoterica where, relatively speaking, barely a handful of people know what they are or how to put them to truly beneficial use. Just their exceedlingly academic names alone belie their applicability to the mainstream and to organizations of any size. There’s nothing like some really crappy monikers to alienate the masses.   

However, in its year-old Einstein portfolio of intelligent services, Salesforce may very well be the first to crack the code to this thorny problem and one of the key reasons it’s able to do so, despite a blog post that claims that Einstein is “AI for Everyone,” is that it’s not trying to boil the ocean as other AI, ML, and data science vendors are wont to do. What Salesforce really means is that Einstein is AI for everyone that uses Salesforce. Whether Salesforce is in turn for everyone is a different question. 

Currently, the Einstein portfolio consists of roughly 20 add-on AI modules, each of which is purpose built to flush important insights or function out of one or more parts of Salesforce’s ecosystem. More are coming. Whereas some modules work with Salesforce’s Marketing Cloud, others are tuned to work with the Service Cloud and so on. 

Other AI vendors target a very wide range of applications, giving you some combination of raw ingredients and expecting you to figure out the rest. For example, there are quite a few providers of horizontally applicable natural language processing (NLP), very much a machine-learned skill for a computer to have. If, after some introspection, you realize that NLP is the sort of thing that will benefit your business (this alone requires some imagination), then you must sort through all the steps it takes to deploy that NLP according to your vision. This may or may not involve the additional barrier of marrying your choice of NLP to the lexicon of your business. 

While this quasi-DIY approach works amazingly well so long as everyone involved is clear on the connection between the what the technology does, the journey to completion, and the measurable business outcomes, these horizontal applications place a lot of the burden to succeed on the organization. The process can be intimidating and without the right personnel, organizations that might otherwise benefit from the rocket science might be dissuaded.

Although Salesforce has horizontally broadened its portfolio from its early days, most of its offerings — its marketing, sales, and service clouds for example — are still true to its roots: the vertical of customer relationship management (CRM). It’s a category that Salesforce knows like no other and it’s because of this expertise as well as what it knows about the design of its own technology that users — and I mean business people (not geeks) — of Salesforce technology can benefit from ML, AI, and data science with mainly mouse clicks instead of coding (and in most cases, it should be noted, some additional service fees).  

In fact, as Salesforce Einstein vice president of product management Marco Casalaina explained to me during last month’s Dreamforce, the company’s annual conference in San Francisco, many Salesforce customers are already benefitting from the Einstein portfolio and they’re not even aware of it. It’s working some of its charm in the Salesforce search engine, free of charge. 

I requested the interview (the audio and full transcript of which is available below) after listening to Casalaina’s counterpart for Salesforce's Marketing Cloud — vice president of product management Raji Bedi — give a Dreamforce keynote in which he said "Einstein democratizes access to data science.” In my mind, it sounded like code for “we fixed that serious PR problem.” But I wanted to hear more to better understand how Salesforce may have pulled it off.

Casalaina explains some of Einstein's benefits in terms that anyone can understand. For example, if you're a Salesforce customer, the Einstein modules know exactly how to apply themselves to your data. There’s very little to figure out. No models to build (one of the typically intimidating factors of machine learning). You don't need to have a vision. No code to write. Not even much of a journey.

For example, if you activate one of the modules, it will sift through the thousands of opportunities that may have been created by your last five marketing campaigns and it will tell you which of them has the highest likelihood of turning into a customer. Or who is most likely to open that next marketing email blast (which in turn can help to minimize your unsubscription rate). Casalaina claims that Einstein can figure out how to do these and other tasks (he covers more examples below), even after you’ve customized your instance of Salesforce in ways that make your data look substantially different from that of other Salesforce customers. 

If the claims are true, then I would a have to agree that it’s a very clear example of democratizing machine learning; making an otherwise intimidating (or difficult to access) technology accessible to an audience that might never have considered it before.  

For developers, Einstein’s talents are available by way of API. But, not outside of the Salesforce ecosystem. For example, developers can build apps that rely on Einstein’s talents in the course of working with Salesforce data, and then market those apps through AppExchange (Salesforce’s third-party app store). I asked Casalaina if there might one day be something like an EinsteinExchange; a third-party “talent store” where other AI, ML, and data science vendors could add unique capabilities that Salesforce itself doesn’t offer.  

The answer, I think, was “maybe.” According to Casalaina, "How that plays with Einstein is yet to be seen I think, but there we will be ultimately a market place for third party providers of data [and] third party providers of what we call feature engineering.”

For now, Salesforce will limit Einstein to Salesforce-branded and engineered AI, ML, and data science capabilities. Given how hot the AI, ML, and data science space is, my guess is that the Einstein portfolio will grow its list of talents through acquisitions as well as internally engineered solutions. So, if you’re an AI startup with a killer spin on CRM data, maybe there’s an opportunity if you nail something unique that Salesforce hasn’t considered before. 

Finally, there are lessons to be learned from Salesforce’s journey with Einstein. To the extent that Salesforce is successful with the offering, the most important decision that Salesforce ever made was to confine its scope to the Salesforce ecosystem. In keeping to this narrow scope, Salesforce, based on its expertise in sales, marketing, and other services, is able to identify and build for the AI capabilities that its customers will most easily deploy, appreciate, and profit from.

Other purveyors of similar technologies should consider how to focus on similarly limited horizontal ranges (aka “verticals”) to ease the burdens of deployment while raising the chances of positive business outcomes. Nintex, a Salesforce partner, is bringing specific AI talents to bear in its worklow automation tools. "The idea," Nintex senior VP of tech strategy Ryan Duguid told ProgrammableWeb, "is to identify where people, decisions, and tasks are delaying workflow completion" and to automate those. "How do I give you Tony Stark? Take the [human in the process], and chuck him in a suit?" Duguid asks. It sounds strikingly similar to how Salesforce is isolating processes within its ecosystem where the cost of human solutions would be disproportionate to the value of the outcomes (or that just intolerably delays them). It's a formula that solution providers in other verticals should waste no time in replicating. 

Here’s the audio and full text transcription of the interview.

Editor's note: ProgrammableWeb has started a podcast called ProgrammableWeb's Developers Rock Podcast. To subscribe to the podcast with an iPhone, go to ProgrammableWeb's iTunes channel. To subscribe via Google Play Music, go to our Google Play Music channel. Or point your podcatcher to our SoundCloud RSS feed or tune into our station on SoundCloud.

Tune into the ProgrammableWeb Radio Podcast on Google Play Music  Tune into the ProgrammableWeb Radio Podcast on Apple iTunes  Tune into the ProgrammableWeb Radio Podcast on SoundCloud


Full Transcript of David Berlind's Interview with Marco Casalaina

David Berlind: Today is Tuesday, November 7th 2017. I'm David Berlind, this is ProgrammableWeb's Developers Rock Podcast [sponsor message]. I'm here at the DreamForce Conference, it's the big annual conference that's put on by Salesforce and I'm joined by Marco Casalaina, he is the Vice President of Product Management for Salesforce Einstein. Einstein is Salesforce's artificial intelligence machine learning platform that gets integrated into all of the various other platforms [00:00:50] that Salesforce offers and Marco, first of all, welcome to the podcast.

Marco Casalaina: Thank you very much, good to be here.

David: I was just listening to Raji Bedi speak about the Marketing Cloud, he is the Vice President of Product Management for that and you guys are like kissing cousins. You're in Einstein, he's in Marketing Cloud and he said something that I thought was a really interesting quote. He said, "Einstein democratizes access to data science." What does that mean?

Marco: It means that what we're doing with Einstein is making artificial intelligence accessible through a larger set of users, people who aren't trained or classically trained anyway in data science. We can do that in large part because of the fact that Einstein is not a general-purpose AI toolkit. Einstein is AI for the Salesforce platform and because it's constrained in that way, it allows us to present a more usable user interface that's accessible to a large audience.

David: [00:01:50] Why couldn't they just have this before? I mean data science is out there, it seems like you hear about it all the time, we've heard about artificial intelligence for decades now. What was the tipping point that has popped the cork and now it's available to everybody? What's the big deal?

Marco: That's a good question. It has to do with critical mass, I think, and part of it is critical mass of data in the cloud because obviously you can't have Ai without data, there's no data science without data. It also has to do with critical mass of compute. Now compute is so cheap that we can do a bunch of stuff just brute force to automate a lot of things that a data scientist would normally do manually. Because compute is so cheap and we can use it so liberally, we can apply that to automate a lot of what Einstein does.

David: Are you saying that artificial intelligence, machine learning the tasks that are associated with those technologies are very compute intensive?

Marco: I am absolutely saying that, they are very compute intensive and in the past, [00:02:50] I mean one of the things that data scientists often do when they're building predictive models manually is use their intuition to cut stuff out, to omit stuff, to omit data that they think won't matter because they can't necessarily compute all of it. My point is that today we can use all of it. Rather than just using intuition, we can use automation, we can use brute force and compute power to be able to determine what is and what is not predictive.

David: A lot of this has to do with the machine learning bit of it and machine learning is a fairly esoteric concept for a lot of business people and business people are the kind of people that, generally speaking, use the various Salesforce platforms. How do you bridge the gap between something that's so complex and an audience of users that doesn't necessarily have that technological capability built into them or they haven't been educated that way?

Marco: Absolutely, everybody is talking [00:03:50] about AI and machine learning and most people don't even know what a prediction looks like. In part, if you look at the evolution-

David: You just said the word prediction, which is the first time you mentioned an application of this technology, so don't be afraid to talk about that too as one of the applications.

Marco: That's a very good point and it's true that you can use artificial intelligence for a number of different things and we'll talk about a few of them as we go through this podcast but one of the most common ones is to make a prediction about a future behavior or a future event. If you look at the evolution of Einstein over the last year because Einstein as of this podcast has only really been around for about a year, we started with these applications because one way that we can make it accessible to users is to make it do something that they're familiar with.

In the Marketing Cloud as we discussed it's about targeting. I don't want to send emails to people who are not going to open those emails. I want to predict who's going to open the email before I send it to them, and by the way, who's going to open it to unsubscribe because I want to put them on the suppression list. [00:04:50] These kinds of predictions are natively comprehensible to users of Marketing Cloud and software like that. They want to do email targeting and that kind of prediction makes sense to them.

In the Sales Cloud it's about lead scoring, about opportunity scoring, it's about understanding is this a good lead for me, predict for me, yes or no? Is this person going to be my customer or not? We started with Einstein with these applications that were very concrete applications of artificial intelligence to problems that our users were facing and now a year in we're starting to expose the Einstein platform itself. Now that our users are starting to get accustomed to what AI can do, now we're starting to just let them do it in a custom way with the My Einstein platform.

David: Are you making modules available that do a variety of these applications? I'm guessing there's a lot of different applications of artificial intelligence, machine learning — not just predicting things. Are these [00:05:50] sort of you pick à la carte what it is you need and you pay for that, how does that work?

Marco: Einstein runs the gamut of what Salesforce does. Einstein is threaded through most of what you do in Salesforce today, it even powers our search engine now. You may not even notice it, but our search engine now is using machine learning to deliver search results to you and you have no idea. There are Einstein features that you can use for example Marketing Cloud Einstein is just included in the higher-level editions of Marketing Cloud and then there are also individual modules that you can use.

For example, Einstein Language Services. Einstein Language Services we expose as an API and you can use that to do intent extraction, you can use it to do sentiment analysis but it also forms the basis of some of our higher-end products, for example, Einstein Bots. Einstein Bots is using Einstein Language Services, so that when you give it an utterance, it's able to figure out and this is not exactly a prediction, [00:06:50] it's what we call an intent extraction or a classification.

It's pulling out of that utterance the information that it needs. If you say, "Where is the equality session at DreamForce?" It can pull out and say, "Okay equality, session," and it can use that to react and come back at you with a response that makes sense to you. We're exposing both the underlying services in some cases as APIs, in some cases as declarative platforms as Salesforce often does. We're also exposing higher level pre-built applications like Marketing Cloud Einstein, Sales Cloud Einstein that solve specific user pain-points like putting these technologies together.

David: In a nutshell, you've taken the data science, you've baked it into a variety of à la carte offerings that are available to all of the existing Salesforce customers. They don't have to be data scientists to say, "Okay, I want intent analysis or I want predictions or whatever." They just pick that off the [00:07:50] shelf of the modules if you say, if we want to call it that. Then they apply it to their data and out comes a prediction and they didn't have to hire a million dollars' worth of PhDs to get it done like they might have had to do 5 or 10 years go.

Marco: Yeah, that is the idea but training is the new coding. What I mean by that is what you need is the data. Let's say for example that you wanted to put up a bot, today we launched the new Einstein Bots framework and people will start doing that with the bots but you need to train these bots in your language and every business' language is different. If you have not up to this point had a chat channel or some channel by which you were gathering data in that format that looks like what people will type to a bot, you're going to have to accrue that data. These things are not magic, they're science and they need to learn.

David: How difficult is that for the average Salesforce user to train your off the shelf machine [00:08:50] learning modules to work with their specific set of data, their lexicon as you imply.

Marco: It's a good question, it depends on which type of product they're using and where in the stack they're using it. For example, if you're just using the raw language services API then you're going to need to bring a labeled dataset in, you're going to need to bring a bunch of utterances and labels that might say ... For example, let's say you were building a bot that's dealing with shipping and orders and stuff like that. Then you're going to have to bring a dataset that says, "I want to know where my box is," and that would classify it as shipping, stuff like that.

If you're using some of the higher-level applications, well we're actually cobbling the data together for you based on the data you already have in Salesforce. Let's consider, for example, Service Cloud Einstein. Service Cloud Einstein is doing tier one triage of cases. A case comes in and when a case comes in, a service case comes in its usually pretty thin there's a subject and a description.

Then later on your agents will work that case and they'll figure out, "Okay what product [00:09:50] does it pertain to, is it high priority, is it escalated," all those kinds of things, right? They are implicitly labeling the dataset. If you fire up Einstein Agent, which is part of Service Cloud Einstein that does tier one triage on cases, it will automatically take all these pre-labeled datasets that you've already been building, all these cases and put it together and learn just from that, so you don't have to put together the data. Again, it depends on where in the stack and what specific services you're using.

David: In some ways there's a natural advantage here, you're making the machine learning and artificial intelligence technology available but you as Salesforce have a fundamental understanding of your schemas, your datasets that sort of thing, so it makes it a lot easier for you to apply that and then as Raji said democratize it to at least a bare minimum Salesforce's existing customers.

Marco: Yeah, that's one of the things that make Einstein work. Every Salesforce customer customizes [00:10:50] Salesforce, they customize the heck out of it. Everybody has their own data model and yet the metadata is always there. Just by virtue of using Salesforce you are always specifying metadata. You're telling us this is not just a var char field, it's an email address or it's a phone number or it's a zip code. and if it's a zip code then we know certain things.

We know that there is no 94105.5 between 94105 and 94106, and by virtue of knowing these things that's how we can automate a lot of this machine learning stuff. The same is true of relationships between tables. Again, by virtue of customizing Salesforce you're implicitly specifying this metadata that tells us the relationships between these tables. We can use that information to build these models automatically for you.

David: These models are clearly available within the Salesforce ecosystem. You mentioned earlier that some of this is available through API, are the models also available outside of the Salesforce ecosystem?

Marco: Not really [00:11:50] because as I said earlier Einstein is not a general-purpose AI toolkit. Einstein is AI for the Salesforce platform and it's meant to run in the Salesforce platform. It may in some circumstances consider data that is not in Salesforce. It can consider external data when building this model but nonetheless it's meant to run in the Salesforce platform.

David: You mentioned the APIs earlier, do developers access through the APIs or you're just internally consuming them in order to marry whatever module the customer picks to their data?

Marco: Yes. We're taking an Amazon model here, we are building higher level services but we're also exposing the lower level services as APIs that we use because we've recognized just as Amazon does that in a lot of cases those lower services can be useful to a lot of folks.

David: You've got an API architecture there to reuse that technology across your entire infrastructure not necessarily to expose it to [00:12:50] external developers or publicly?

Marco: For example, consider Einstein Language Services and Einstein Bots. Einstein Bots is using Einstein Language Services to do intent extraction, whenever you type anything into a bot it's using Einstein Language Services but as a developer you can also just use Einstein Language Services itself as an API.

David: You can externally outside of the Salesforce cloud?

Marco: It's a cloud services, it's a cloud API so it's still using the Salesforce cloud but no not unless you can call it as an API just as you would call a service from any vendor.

David: Okay, great. What about the opportunity for other vendors in the machine learning space who are building very specialized models to democratize access to whatever that kind of model is. Is there the equivalent of Einstein exchange, can they add their technology to some marketplace to expand the palate of models from which Salesforce customers [00:13:50] can pick?

Marco: It's going to be the other way around actually. Initially, with the release of Einstein Prediction Builder that's where we start to ... In general, the Salesforce ecosystem is driven by this packaging system. You can make an app in the Salesforce platform, package it and sell it and now we're going to start to be able to add intelligence to those packages with Einstein.

You'll be able to add Einstein Prediction Builder fields to your package. For every one startup in the Salesforce ecosystem that's doing AI or machine learning there are probably ten that have applications that are useful but don't have any intelligence in them. Initially, we're going to start by basically offering that capability to our ISVs. To say, "Hey now you can add intelligence when you're using Einstein and you can be powered by Einstein.

Meanwhile, yes, it's also true that there are a number of startups even now that are using some of the APIs that we've exposed, for example NeuraFlash is one. NeuraFlash also has a bots framework also [00:14:50] powered by Einstein Language Services and they're focusing on certain verticals.

David: I'm still getting at there's just a whole machine learning economy developing out there and we're seeing all kinds of crazy models being developed and then offered. Let's call it democratization of machine learning in general and those models are being made available via API to anybody who needs them and they can apply them in any dataset, it doesn't have to be the Salesforce dataset. Will there be something equivalent of AppExchange, where they can compete for the affections of Salesforce users to apply that technology to their dataset? It very well could be models that you guys are not offering à la carte and it just expands the opportunity.

Marco: Absolutely, and in fact they do that today. Not only do they do that today, I myself [00:15:50] did that once upon a time. I was at a Salesforce AppExchange partner seven years go in which we did build predictive models and we pushed them back into Salesforce just using Salesforce's standard public APIs and sure lots of companies can do that and will do that.

How that plays with Einstein is yet to be seen I think, but there we will be ultimately a market place for third party providers of data, third party providers of what we call feature engineering. That is like being able to come up with new and innovative variables and derivations of the data that can be useful for Einstein to work on. I think that will happen but right now we're really focused on this core use-case of adding intelligence to Salesforce and adding intelligence to our partner applications. That's where we're going right now, where this evolves well we shall see.

David: The million-dollar question or hopefully not a million dollars is how much? How much more if I'm a Salesforce customer [00:16:50] must I spend in order to get some of this intelligence applied to my data?

Marco: As I mentioned, a lot of this is already sprinkled throughout Salesforce and our users barely even know it. In some cases, some of the Einstein capabilities are included in some of the editions of the products as I mentioned Marketing Cloud. In general, at Salesforce all of us have to make this vision statement. My vision statement is two words, it's Einstein everywhere. My goal is to put intelligence everywhere, you should not be able to touch Salesforce without in some fashion being assisted by this intelligence. Some of those things will be extra in some cases.

David: How much extra?

Marco: It differs by product I can't name them all. There are now 20 different Einstein capabilities that are already live.

David: Let's say I'm a typical user and I want predictions. How much more, is it a percentage of what I'm spending now? We talk about the democratization of machine learning and artificial intelligence here. Democratization implies availability [00:17:50] of something very rocket science like at a dramatically reduced cost. Give us some idea of what that is.

Marco: Relative to the cost of trying to do these things manually it's super cheap. For example, yes if you type something into a search box in Salesforce the Einstein capabilities baked into that cost you zero dollars.

David: How much though if I add it to my existing dataset come on?

Marco: If you're talking about Sales Cloud Einstein lists at $50 per user per month.

David: It's on a per user basis?

Marco: In some cases, again there's many different Einstein capabilities Einstein Language Services, the API, if you use it just as an API it is charged on a consumption basis. I don't remember the exact number but it's a certain amount of dollars for a million predictions at shot. It's cheap in any case but I'm not even going to quote the number because I can't remember it off the top of my head.

David: It's cheap compared to how much it would have cost if you had to go out and get your own data scientist, you can say that for sure.

Marco: It's cheap compared to trying to build a similar model in TensorFlow because if you try to do that [00:18:50] ... TensorFlow is a great toolkit don't get me wrong but it pales in comparison to the ease of use of some of these APIs that we now provide.

David: Are your models based on TensorFlow?

Marco: We use a number of open source toolkits, TensorFlow is one of them. In deep learning cases we do use TensorFlow, sometimes we use Spark. We use all kinds of different toolkits to power the use cases that we have.

David: Last bit here, which is there's a little bit of a gap when it comes to understanding how these models that are out there, third party models ... to a lot of people that would be a third-party model anyway.... how that can take your business to the next level. It's one thing to understand all of the data that you have. It's another thing to understand or wait a minute there's this capability that these models bring that I've never had before at a greatly reduced cost. How do you educate everybody on the availability of that and how it could matter to them?

Marco: [00:19:50] It is tricky and one key thing to note is that what Einstein is about, we're not selling you models per se. These are not pre-built models that are just built on everybody's data. What we're really doing is it's a technique and when you apply Einstein to your data, it is building a model from scratch based solely on your data and solely for your business. How we can get the word about that out there? Well, that's a good question. I hope folks like you can spread the word about what that means but a lot of it really comes down to being able to demonstrate business value basically in some sense word of mouth.

To get customers like US Bank up there and talking about how they've been successful with Sales Cloud Einstein or to get College Forward talking about how they've been successful with My Einstein the platform. That strikes me as probably the best way to get that message out.

David: I don't think the problem is unique to Salesforce. In general, the idea of democratized [00:20:50] machine learning creates a whole new level of business that everybody can aspire to. The problem is they have to put their imagination to work to really understand how to make it work for them and ultimately envision what the benefits could be in terms of return on investment.

Marco: We can hopefully jog their imagination by showing them customers of ours similar to them that are solving similar problems.

David: You could put like popup up when they go into their normal Salesforce interface it says, "By the way we threw a little Einstein at your data and it came back and told us that these are the five top opportunities that are going to come through, how would you like to try Einstein platform and get more of this?"

Marco: We can do that but we have to be very careful. In the AI world you do have to be careful about what you just automatically apply. In Salesforce, trust is our number one virtue and our customers trust that we're not just going to fire up automated processes without their explicit consent. We probably wouldn't [00:21:50] do that because you always tread this fine line in AI and especially now that AI is getting so good. You tread this fine line between "hey that's cool"and "you're Vladimir Putin." You've got to be super careful about that.

David: Okay, well, Marco Casalaina thank you very much for joining us on this podcast.

Marco: Thank you very much for having me.

David: It's great to have you. Well, that concludes this podcast. Again, I'm David Berlind I'm the Editor and Chief of ProgrammableWeb you've been listening to one of our developer rocks podcast, we'll see you at the next podcast. If you want to tune in, you can find us on sound cloud at soundcloud.com/programmableweb You can also find our video podcasts up on YouTube at youtube.com/programmableweb thanks for joining us.

David Berlind is the editor-in-chief of ProgrammableWeb.com. You can reach him at david.berlind@programmableweb.com. Connect to David on Twitter at @dberlind or on LinkedIn, put him in a Google+ circle, or friend him on Facebook.
 

Comments