When Salesforce recently unveiled Einstein, its big move into artificial intelligence, it took many people by surprise. On the one hand, there were partners such as InsideSales and Apttus who have made much of their own AI credentials and who did not expect Salesforce to wade in to the space with such bravado. On the other hand, larger cloud players with deeper pockets, including Amazon, Google, Microsoft, Facebook and IBM, have all been investing in AI for several years and may have wondered why Salesforce was suddenly jumping on the bandwagon.
Both sides may be forgiven for asking whether Salesforce really understands what it’s getting into. Enterprise customers will want to know the answer to that question. So here’s a quick take on Einstein and whether it really is ‘rocket science’, as the saying goes.
What we mean by AI
First though, a quick overview of what we mean by AI. A really helpful Fortune article has just been published that describes some of the recent history of AI developments and why, to quote the article’s online standfirst, “deep learning is suddenly changing your life.” It breaks down AI into three layers:
Artificial intelligence — the broadest term, applying to any technique that enables computers to mimic human intelligence, using logic, if-then rules, decision trees, and machine learning (including deep learning).
Machine learning — The subset of AI that includes abstruse statistical techniques that enable machines to improve at tasks with experience. The category includes deep learning.
Deep learning — The subset of machine learning composed of algorithms that permit software to train itself to perform tasks, like speech and image recognition, by exposing multilayered neural networks to vast amounts of data.
It’s the third layer, deep learning, that is throwing up the big advances in AI. Interestingly, this was a point highlighted by Salesforce’s SVP data science for Service Cloud & Communities and CTO of search data science in a briefing we covered back in March — an early hint that Salesforce was serious about investing in AI. Today’s rapid advances are being fueled by the growing availability of data that’s readily available to feed to increasingly powerful deep learning server farms.
Quite often, though, when people talk about AI they’re talking about the capabilities it enables rather than the underlying technology. Deep learning essentially is a mechanism that allows computers to learn to recognise patterns, whether those are words, faces, behaviors or trends. That can be applied to the user interface, allowing us for example to converse with our computers, to visual processing, which for example allows robots to move around without colliding with objects and people, or to business processes, where computers learn what has worked well and can guide us to improve performance. So chatbots, autonomous vehicles and predictive lead scoring are all examples of AI-powered applications.
AI for CRM
In Einstein’s case, it takes this pattern-recognition capability and applies it to the specific realm of CRM applications. The AI platform has learned from analyzing all the billions of data points in Salesforce’s archive of sales, service and marketing transactions. That learning is then packaged up into specific tools that users can apply to certain processes within Salesforce’s CRM applications, such as lead scoring or audience segmentation. John Ball, Salesforce senior vice president and general manager of Einstein, is careful to qualify those capabilities:
The AI is for CRM, it’s AI for Salesforce. We’re not trying to be a general purpose AI platform.
The implication is that if you have a highly customized Salesforce application, Einstein may not be able to help you very much. That’s because there won’t be enough data in the system for it to work with. What the team has concentrated on is building a user experience that makes it easy for admins to harness the data science for specific CRM functions.
Einstein also includes developer-level services, but again what’s available initially is limited to specific capabilities — predictive vision, predictive sentiment and predictive modeling. These are important services with a wide range of applications to CRM, which is why they’ve been chosen. If developers want to some other AI capability, they must either wait for future releases, or use the PredictionIO service that will be available running in Heroku Private Spaces to build their own models.
Limitations to the model
But developers should bear in mind that the value of those models will be limited by the amount of data they are exposed to. As Dave Elkington, CEO of InsideSales, a Salesforce partner that has built predictive analytics for the sales process based on more than ten years of data, told diginomica, the data matters far more than the algorithm:
It’s not unlike any network effect. Has Facebook got the best user interface? Probably not. Does Google have the best search algorithm? Maybe, maybe not. It’s irrelevant.
Once you have enough data associated with that network effective data, it becomes the value of Google, it becomes the value of Facebook. Ultimately the value of LinkedIn and Twitter. Ultimately that’s our value.
It’s also important to apply the model correctly. Kirk Krappe, CEO of Apttus, which has been applying Microsoft machine learning to the quote-to-cash process, says that when you’re working to optimize a business process, it’s important to really work out what the customer wants to achieve in order to apply the AI correctly.
For true optimizing of process — like managing discounting better, getting better insights into discounting and the machine telling you what you should be doing and why — it requires a protracted process.
I don’t think it’s a throw it out there and suddenly everyone’s going to develop thousands of use cases. Given what we’ve gone through, I’d be surprised. Unless they’re very lightweight use cases … There’s a lot of involved things that you need.
Others note that enterprises will want to have tools that allows them to rigorously test the effect of applying Einstein’s AI, for example by comparing success rates when using it against a comparable set of use cases that aren’t using it. This is a classic methodology for testing the effectiveness of a machine learning model.
Salesforce has assembled an impressive team of AI experts through a trail of acquisitions, mostly in the past 18 months. These people are ‘rocket scientists’ and clearly know what they’re doing in the AI field. I get a sense that, in amongst the marketing claims, they’re wise to both the capabilities and the limitations of the science.
But at a time when machine learning has hit the ‘peak hype’ apex of the famed Gartner hype cycle, there’s a danger of people getting carried away in their anticipation of what can be achieved with this new technology. This is a classic pattern with any promising emerging technology — before people really understand it, they tend to overestimate what can do for them. The next phase of the cycle is that they get disappointed when it fails to live up to their overinflated expectations, and they end up losing faith just as it starts finally to prove its worth.
Net-net, Einstein is an important addition to the Salesforce platform that does bring important new capabilities. When Salesforce chief product officer Alex Dayon says AI “is really going to transform software for the next decade and beyond,” it’s not an idle boast. The impact of modern AI techniques is going to be massive — but it’s not going to happen all at once. Einstein brings some powerful new capabilities, but it is just at the start of its journey, and enterprises are at the start of their journey of learning what we can and can’t do with it.
Image credit - Atlas launch from Cape Canaveral, Florida © Mike Brown - Fotolia.com
Disclosure - Salesforce is a diginomica premier partner at time of writing. Salesforce has not funded my travel to attend Dreamforce.