BACKGROUND IMAGE: iSTOCK/GETTY IMAGES
One of the biggest challenges of building AI apps is reconciling how the data scientists crafting AI models and customers using them often live in two separate worlds. Experts can spend much time optimizing for new predictions that may bring no practical value to end users.
Salesforce Einstein Discovery bridges that gap by helping business users experiment with creating AI apps on their own.
The resulting predictions may not be as accurate or scalable as those honed by a data scientist, but they do help users identify models they will actually use. Once a user identifies a good fit, data scientists can be given the resulting algorithm for further improvement.
At the Dreamforce 2018 conference, Einstein Discovery trailblazers shared some of the lessons they learned over the last year.
Understand the business case
Auction Nation, a salvage car trader in South Africa, quickly dove into Salesforce Einstein Discovery to improve its process for rating the prospects of wrecked cars. The company employs a group of auto experts who buy cars from insurers that are resold at a markup. Unfortunately, some types of damage are much harder to profitably repair and resell.
Errol LevinCOO, Auction Nation
Errol Levin, COO at Auction Nation, hoped Einstein Vision would make it easier for purchasers to recognize these lost causes more accurately in an effort to keep them off the company's inventory.
Most of Levin's staff did not believe it was possible to create an app that would do any better than humans, but using Salesforce Einstein Discovery, he experimented with the workflow himself. While the first results weren't perfect, workflow did improve.
It was important for Levin to dive in and discover these limitations himself.
"You have to understand the core business before you get to data scientists," Levin said. "They have to see there is a fundamental process to improve."
Start with a small team
Rebecca Greenberg, director of commercial systems and specialty analytics at Takeda Pharmaceuticals, wanted to provide actionable information to support sales representatives interacting with doctors. However, Health Insurance Portability and Accountability Act regulations protect much of this data, and it can't be shared across the company.
Greenberg experimented with Salesforce Einstein Discovery to look through mounds of data that was previously underutilized. She quickly found a bell curve in the cancellation rate of doctors ordering Takeda medicines. This finding helped identify doctors on the edge of cancelling, enabling sales reps to focus more time in connecting with these doctors. The resulting app sends field reps alerts on accounts that are at risk of loss while protecting patient data.
It was important for Greenberg to experiment with identifying the kinds of predictions she could make with a small team to be able to decide what to focus on.
"That small team let us see what parts of the tools were useful and where we would need to go with true data scientists," Greenberg said.
Clean your data
AI predictions are only as useful as the data that drives them, and if you have garbage going into them, garbage will come out. Managers experimenting with new models directly using Einstein Discovery can quickly identify where they may need to change their business processes to improve the quality of their data.
"This will make your garbage clear right away," Greenberg said.
Takeda already invested heavily in understanding, cleaning and organizing its data using tools such as master data management and working with a team of data stewards. But the process of building new AI models helped identify unexpected problems.
Jonathan Wray, former director of product management for Salesforce Einstein Discovery, said another Einstein user discovered its Chinese branch using an entirely different accounting model than the rest of the company. This discrepancy only emerged when an Einstein app showed China reporting sales several times higher than the rest of the world, even though managers intuitively knew this was not the case.
You may even consider reaching out to partners to help scale up new AI apps once a useful business case is identified. After Greenberg's early success, she worked with LiquidHub, a customer engagement firm recently acquired by Capgemini, to scale up the app. She said they are sometimes willing to help at little or no cost in order to build their resume of AI success stories.
Understand how tools behave
TouchCR, a consumer marketing analytics platform, turned to Einstein Discovery to improve the sales process for its sales reps.
Ritchie Hale, chief innovation officer at TouchCR, said he personally got involved in the initial experiments to better understand how they could improve the business. Once he had some insight into what worked, he was able to provide better insight to the rest of his team to scale up the use of AI.
His biggest insight is that AI is more of a journey than a destination, and you can expect a lot of failed ideas along the way. By not experimenting, companies can lose business to more nimble competitors.
"If you want to be competitive, you need to get insight for your market that you don't have today," Hale said.