Episode Summary: Over the last year, we’ve covered a lot of marketing applications, including a survey of the landscape of machine learning in marketing applications and which industries will be affected first. But marketing doesn’t tell the whole story when it comes to B2B sales. At some point, we need to take these clicks and turn them into appointments, for example.
In this episode of AI in Industry, we are joined by Vitaly Gordon, VP of Data Science and Engineering at Einstein, Salesforce’s customer relationship management application driven by artificial intelligence.
We speak with Vitaly about where AI is serving a role in sales enablement today and how the CRM and sales tool ecosystem might be different in the near-term future; how will salespeople be able to leverage AI to make themselves more productive? Vitaly paints an interesting picture of where he sees the low hanging fruit and the unique challenges with sales data and B2B data that are quite different from the challenges those in the B2C world might deal with.
We hope that B2B business leaders will find this episode useful for their AI initiatives.
Subscribe to our AI in Industry Podcast with your favorite podcast service:
Guest: Vitaly Gordon, VP, Data Science and Engineering – Salesforce Einstein
Expertise: Data mining, machine learning, Hadoop stack, various programming languages
Brief Recognition: Vitaly Gordon is the VP of Data Science and Engineering at Salesforce Einstein. Prior to this, he was the Director of Data Science at the same company. Earlier in his career, he served as a data science lead at LinkedIn and as Senior Manager for Data Science at LivePerson. Vitaly earned a BS in Computer Science from Ben-Gurion University of the Negev and an MBA from Technion-Israel Institute of Technology.
Interview Highlights
(02:20) Let’s cover the state of AI in sales enablement. Can you get us up to speed as to where AI is serving a role in this area?
Vitaly Gordon: Sales enablement covers a myriad of sales processes from lead generation to closing a sale to making sales teams as efficient as possible. We have definitely worked on a number of solutions in that space. It is interesting to see that the approach needed to solve these problems in both the sales and enterprise space is very different from the consumer approach that I and other data scientists come from.
There is a lot of work we have to do not just in solving the actual problem but also solving problems that our customers care about, such as privacy, security, and related topics.
(04:00) Since you mention the difference between B2B and business-to-consumer (B2C), let’s dive deeper into this. Some of the people listening are aware that much of the data science action is in the big consumer companies. How different was it to find the right kind of data to solve the B2B problems versus B2C?
VG: It was a very transformational shift. I was very surprised that some things I took for granted in the consumer space were different in the enterprise space. One example is access to data. As a data scientist, I used to get familiar with the dataset I was working on, which is something that as a vendor giving service to customers, they’re not as comfortable with that as consumer data. There is a whole array of privacy of how can we do machine learning without access to the data of the business transaction. Some of it is the most sensitive data of our customers and they are correct in their opinion that they don’t want people outside their organization to access it.
(05:40) There need to be ways to move around this proprietary stuff and still have enough to train on [without risking data security]. Is that a problem?
VG: That definitely is true. The way we approach this problem is by investing a great deal in automation. There is a new area that is rising in machine learning called automated machine learning, which means doing machine learning without a human in the loop or with minimal involvement.
Privacy is just one of the problems. The other problem in the enterprise space is the sheer scale because enterprise and sales enablement for Fortune 500 companies are very different from the mom-and-pop shop: Different processes with different data sets.
If the standard now for building a machine learning offering is to have a team of data scientists sit and work on it for weeks or months, obviously that doesn’t scale for hundreds of thousands of enterprises and multiple processes. That is why automation is just the key to being successful in the enterprise.
(07:45) What can AI do in your ecosystem today?
VG: I would say we’ve had a lot of success in optimizing the job of sales individuals. If you think about sales, a lot of their work is process automation and following a playbook: not actual sales. They spend time on a lot of things that machines are very good at. One of the best examples is ranking leads.
There is a lot of data around these leads. There are also historic results of which leads have panned out for that business and which ones have not. This is an example of a classic machine learning problem. We should employ machines to organize our call queue and not spend so much time on chasing and calling them.
Unfortunately, what happens in business is this first in, first out routine that leaves much to be desired. This can just be the entry point, and then you can take it to more advanced stages in the sales enablement pipeline.
(10:30) It seems we have a lot of inherent challenges here. What are those big hurdles for you?
VG: These challenges are part of why enterprise AI is so hard, and why we haven’t seen as much innovation compared with the consumer space. It comes from low amounts of data, which is a challenge in itself and requires techniques that come more from anomaly detection rather than pure classification. For some businesses, success might be one out of a thousand. And the data that is being used to understand if that lead would pan out is completely different.
What makes this extremely difficult is that everytime a customer signs up, we have no idea who this customer is, what business their in, and how their pipeline looks. This is why automation is key.
(12:47) Is there a vision to find clusters where there is overlap [between sectors], and there are transferable lessons?
VG: This is an area that we have definitely been thinking about. An example of where this could work is if you take the approach the way Tesla did to intelligent cars, where all the cars share data between them. The benefit you get from it is your car gets smarter based on the behavior of a different car.
The price you pay for this is that your own personal driving data is used to make all the cars better. That enabled them to have massively more data to solve that specific problem than other car manufacturers that don’t take that approach.
We are thinking about whether we should have a similar industry-based approach where companies can opt into such an engagement to make everyone better. But again, even companies in the same exact industry still behave and manage their data…
As an anecdote, I can tell you that Salesforce as a platform comes with a scheme where you put your data in. Salesforce is also customizable. Eight percent of the data in Salesforce is the customizable part. Every business that we see, even though our product comes with a very well defined schema, 80% of that data is not in that scheme but in the extension that customers create based on that schema.
(15:30) Look five years into the future, where is AI going to play the most prominent role in the CRM and sales enablement space?
VG: I have a very futuristic view on this. I believe that pretty much everything will change, and I think it comes with our expectations as customers. A phone that was invented 11 years ago is now a smartphone. And now we have smartwatches. I think very soon we will stop adding the tag “smart” to things and will just call everything else “stupid.”
A lot of the interesting products today are still in that stage, where it’s more about data collection or data organization systems. One of the analogies we have for these systems is cops, where at the end of the shift, cops have to stay to type their report and file it. That is the stage many enterprises are at today, but the entire world is changing and people are no longer comfortable with manually typing information or even thinking about the auxiliary stuff they need to do in their day-to-day lives.
The way the world will change is that people will just do their jobs, and we already have a lot of sensors in place that know where we are and what we do. I am talking about things like email, calendars, locations and meetings. We can use all that information and start creating this business knowledge base that a lot of enterprises are not really actively asking people to spend time on as part of their work week. That is the first stage of this completely transforming the data entry.
The amount of data will just grow. We have to get away from letting people try to find the relevant data to do their job. It has to completely reverse itself and the machines will basically automatically surface the information that we need.
(19:35) How does that vision transfer to sales?
VG: Your salespeople will wake up in the morning and they have a smart speaker in your house. The same way that people use the smart speaker today, they ask the speaker about the news or the weather or what time it is. What about asking the speaker “what should I work on today?” Everyone has a job that consists of making decisions. We need to decide what is the most important thing we need to do today and what is not.
Think about that in sales enablement. You might be working on multiple deals in peril, and deals in the sales enablement space take some time but there isn’t dally activity. Then you can say, I just received news that there is a leadership change in this company. You should know it because it might affect your deal. Or you might have an email in your inbox, that you could have triangulated into a deal that you were working on.
If that email has a negative sentiment, you should go and check it out, because you might have a problem where you need to intervene. That is an example where the right information is surfaced to you at the right time.
Subscribe to our AI in Industry Podcast with your favorite podcast service:
Header Image Credit: Self Management Group
