Preparing your Enterprise for AI: Lessons from predicting sales for B2B

Artificial Intelligence (AI) lurks behind consumer applications, often without the end-user’s knowledge. From identifying images to recommending friends to serving the right ad, web-scale data has rendered many old algorithms (e.g., neural networks) potent and capable of beating humans at similar tasks. But what about the Enterprise, where the last big wave of intelligence was Business Intelligence (BI), and it is still playing out twenty years later?

In this post, I will explain why AI is not BI, and the old habits of the Enterprise both from the supplier and the buyer side will need to change before these technologies are widely adopted. With Salesforce’s Einstein, Oracle’s Intelligent App Cloud, Amazon Machine Learning, and countless other horizontal AI solutions now in the market, it is worth understanding how organizations can extract value from AI.

Business Intelligence was the last successful analytics killer-app in the enterprise. When a newly minted data-scientist first encounters BI, her reaction is akin to Groucho Marxcommenting on Military Intelligence. BI systems are built on middle-school math and while they are very useful, they are not analytically hard. Middle school math is “horizontal” in that you can apply the basic ideas – summation, averages, deviation, etc. to any Enterprise without any knowledge of the specifics of the business.

It is no surprise then that the AI solutions from incumbent players have followed the BI playbook. Make a set of standard toolkits available and make the customer responsible for the output of the AI as well as the business outcome of the deployment. This approach does not work.

My company Lattice built an AI-based marketing & sales solution from the ground-up. We first introduced an AI-based X-sell/Up-sell recommender for B2B sales & marketing in 2010. Since then we have expanded our solution to include a number of other use-cases including finding new prospects, serving targeted ads, out-bounding to key accounts, scoring leads, and retaining customers. Along the way we have also deployed our solution over 200 times, usually at companies ranging from high-growth mid-market (e.g., Mulesoft, Hootsuite, CSOD, Dropbox) to dozens of Fortune 500 companies (e.g., Dell, Amazon, Suntrust Bank, Tektronix, HireRight, Thomson Reuters, Verizon, etc.) . Through the course of engaging with the end-users and continuously refining our solutions, we have learnt five important lessons about creating measurable value for our customers.

  1. AI applications need cross-platform data – Peter Norvig pointed out several years ago that mediocre algorithms with more data beat great algorithms with less data. Therefore AI applications that limit themselves to a single platform will invariably be beaten by applications that can make use of cross-platform data as well as web-scale data. For example, at Lattice we have built a global index of over 250M businesses that we track for changes and events using web-based and offline signals. Each of these signals is fed into predictions of future purchases. For 90% of our deployments, these “external signals” are much better predictors than signals derived from CRM or Marketing automation alone
  2. End users need explanations not orders – A sales rep is never satisfied with a purchase probability because it is not actionable information. He wants to know why the score is what it is, why was this account picked over others, and what he should talk about when he calls. The recent surge in AI was caused by the spectacular success of one particular method – deep learning. Deep learning based AI is ideally suited for high-dimensional problems, e.g., image recognition, but it doesn’t explain its rationale in a human-understandable form. Any AI will have to explain itself to an end-user, or lose their trust – this limits the usefulness of deep-learning based AI to sales & marketing problems.
  3. Data-Scientists need controls – While an end-user may be satisfied with a score and an explanation, the data-scientists in the organization will want full control. Black box algorithms may work in the SMB segment but don’t work as well in the Enterprise. In fact, our platform took a giant step forward when we opened it up so data-scientists could use their own algorithms on our data. AI is supposed to amplify human intelligence, not replace it.
  4. AI has to be right – Neils Bohr – a contemporary of Albert Einstein – famously said, “it’s very hard to make predictions, especially about the future”. Unlike the BI world in which the analysis is focused on understanding the past, the AI applications have to outperform human guesses about the future. At Lattice, we take responsibility for the quality of predictions, not just the quality of software. We monitor every model in production at every customer through a set of automated tests. Any drift in models is detected automatically and alerts are sent out to the customer success team. This will require a massive change for traditional BI vendors who just had to deliver bug-free code that could do middle-school math. Will their predictions be as accurate for a $1000 product being sold through the channel, as a $100,000 product being sold direct? “One size fits all” does not work for AI applications.
  5. The vendor is responsible for the business outcome – Rob Bernshteyn, the CEO ofCoupa has pointed out the secular trend from SaaS to VaaS (Value as a Service). Nowhere is this trend clearer than in the world of AI applications. We have learned through experience that aligning closely with the business process, embedding the solution deeply into the workflow and being responsible for the outcome are the only ways to successfully deliver AI. The BI vendor model of dropping the software, having an SI integrate it, and making it the business’s responsibility to derive value does not work anymore.

We are excited about the announcement from Salesforce today. As they take Einstein into the Enterprise, we look forward to jointly educating the market on the potential of this technology. In the spirit of collaboration, we hope that our lessons are useful to other AI enthusiasts – both buyers and suppliers.

Written by

Shashi Upadhyay
October 5, 2016