Every business has customers. Every customer needs care. That’s why CRM is so critical to enterprises, but between incomplete data and clunky workflows, sales and marketing operations at most companies are less than optimal.

At the same time, companies that aren’t Google or Facebook don’t have the billion-dollar R&D budgets to build out AI teams to take away our human efficiencies. Even companies with the right technical talent don’t have the petabytes of data that the tech titans use to train cutting-edge neural network models.

Enterprise AI shouldn’t be impossible

Salesforce hopes to plug this AI knowledge gap with Einstein. According to chief scientist Richard Socher, Einstein is an “AI layer, not a standalone product, that infuses AI features and capabilities across all the Salesforce Clouds.”

The 150,000+ companies who already use Salesforce should be able to simply flip a switch and deploy AI capabilities to their organization. Organizations with data science and machine learning teams of their own can extend that base functionality through predictive APIs like Predictive Vision and Predictive Sentiment Services, which allows companies to understand how their products feature in images and video and how consumers feel about them.

The improvements are already palpable. According to Socher, Salesforce Marketing Cloud’s predictive audiences feature helps marketers hone in on high-value outreach as well as re-engaging users who might be in danger of unsubscribing. The technology has led to an average 25 percent lift in clicks and opens. Customers of Salesforce’s Sales Cloud have seen a projected 300 percent increase in conversions from leads to opportunities with predictive lead scoring, while customers of Commerce Cloud have seen a 7-15 percent increase in revenue per site visitor.

Achieving these results has not been cheap. Salesforce’s machine learning and AI buying spree includes RelateIQ ($390 million), BeyondCore ($110 million), and PredictionIO ($58 million), as well as deep learning specialist MetaMind — of which Socher was previously founder and CEO / CTO. Marc Benioff spent over $4 billion to acquire the right talent and tech in 2016.

The competition in enterprise tech is intense

Even with all the right money and the right people, rolling out AI for enterprises is fraught with peril, due to competition and high expectations. Gartner analyst Todd Berkowitz pointed out that Einstein’s capabilities were “not nearly as sophisticated as standalone solutions” on the market. Other critics say the technology is “at least a year and a half from being fully baked.”

Infer is one of those aforementioned standalone solutions offering predictive analytics for sales and marketing, putting them in direct competition with Salesforce. In a detailed article about the current AI hype, CEO Vik Singh claims that big companies like Salesforce are “making machine learning feel like AWS infrastructure” which “won’t result in sticky adoption.” Singh adds that “machine learning is not like AWS, which you can just spin up and magically connect to some system.”

Socher acknowledges that challenges exist but believes they are surmountable.

Communication is at the core of CRM, but while computers have surpassed humans in many key computer vision tasks, natural language processing (NLP) and natural language understanding (NLU) approaches fall short of being performant in high stakes enterprise environments.

The problem with most neural network approaches is that they train models on a single task and a single data type to solve a narrow problem. Conversation, on the other hand, requires different types of functionality. “You have to be able to understand social cues and the visual world, reason logically, and retrieve facts. Even the motor cortex appears to be relevant for language understanding,” explains Socher. “You cannot get to intelligent NLP without tackling multi-task approaches.”

That’s why the Salesforce AI Research team is innovating on a “joint many-task” learning approach that leverages transfer learning, where a neural network applies knowledge of one domain to other domains. In theory, understanding linguistic morphology should also accelerate understanding of semantics and syntax.

In practice, Socher and his deep learning research team have been able to achieve state-of-the-art results on academic benchmark tests for main entity recognition (identifying key objects, locations, and persons) and semantic similarity (identifying words and phrases that are synonyms). Their approach can solve five NLP tasks — chunking, dependency parsing, semantic relatedness, textual entailment, and part of speech tagging — and also builds in a character model to handle incomplete, misspelled, or unknown words.

Socher believes that AI researchers will achieve transfer learning capabilities in more comprehensive ways in 2017 and that speech recognition will be embedded in many more aspects of our lives. “Right now, consumers are used to asking Siri about the weather tomorrow, but we want to enable people to ask natural questions about their own unique data.”

For Salesforce Einstein, Socher is building a comprehensive Q&A system on top of multi-task learning models. To learn more about Salesforce’s vision for AI, you can hear Socher speak at the upcoming AI By The Bay conference in San Francisco (VentureBeat discount code “VB20” for 20 percent off).

AI research is hard, but operationalizing workflows is harder.

Solving difficult research problems is only step one. “What’s surprising is that you may have solved a critical research problem, but operationalizing your work for customers requires so much more engineering work and talented coordination across the company,” Socher reveals.

“Salesforce has hundreds of thousands of customers, each with their own analyses and data,” he explains. “You have to solve the problem at a meta level and abstract away all the complexity of how you do it for each customer. At the same time, people want to modify and customize the functionality to predict anything they want.”

Socher identifies three key phases of enterprise AI rollout: data, algorithms, and workflows. Data happens to be the first and biggest hurdle for many companies to clear. “In theory, companies have the right data, but then you find the data is distributed across too many places, doesn’t have the right legal structure, is unlabeled, or is simply not accessible.”

Hiring top talent is also “non-trivial,” as computer scientists like to say. Different types of AI problems have different complexity. While some AI applications are simpler, challenges with unstructured data — such as text and vision — mean experts who can handle them are rare and in-demand.

The most challenging piece is the last part: workflows. What’s the point of fancy AI research if nobody uses your work? Socher emphasizes that “you have to be very careful to think about how to empower users and customers with your AI features. This is very complex but very specific. Workflow integration for sales processes is very different from those for self-driving cars.”

Until we invent AI that invents AI, iterating on our data, research, and operations is a never-ending job for us humans. “Einstein will never be fully complete. You can always improve workflows and make them more efficient,” Socher concludes.

Mariya Yao is the Head of R&D at TOPBOTS, a strategy & research firm for artificial intelligence and bots. 

This story originally appeared on Www.topbots.com. Copyright 2017