Laurent - stock.adobe.com
As companies strive to improve predictions and triage problems more quickly, customers said that using tools such...
as Salesforce Einstein AI will result in a lot of failure. However, the secret to success lies in learning from the different kinds of failures, creating value with the next experiment.
It's not just about experimenting with the algorithms, customers said. Companies need to experiment with data structure, too, as well as how AI is woven into apps and how AI initiatives appear to users.
Here are eight tips to help your odds of success rolling out Einstein artificial intelligence.
1. Start with a clear and simple project
A simple project enables your team to see how Salesforce Einstein AI works, so even if the project fails, it opens everyone's eyes to how AI creates practical value. Tire giant Michelin experimented with Einstein Vision to improve its ability to automatically categorize images of tire damage, identifying quality problems.
Danielle DeLozierGlobal product owner of Service Cloud, Michelin
"Once we looked at Einstein, we saw the value it could bring to our entire process," said Danielle DeLozier, global product owner of Service Cloud at Michelin. "This helped prime the pump for getting more AI development ideas in our development pipeline."
2. Measure before and after
Start an Einstein artificial intelligence project with a concrete, measurable goal, and create a metric for determining if a particular AI use case and implementation deliver value. Michelin uses Service Cloud to address customer issue time. In the past, the process for resolving issues was lengthy, and it was hard for Michelin support agents and customers to understand where they were in the process -- quality leaders at Michelin want to address most issues in less than a day.
Metrics include time to resolve issues and the number of issues to immediately address. The tire damage app can then be measured against whether and how much it improved the ability to resolve issues.
3. Focus on the value rather than problems
As it turned out, the tire damage app did not deliver the same level of accuracy in categorizing problems as a quality inspector inspecting the physical tire. It did, however, deliver results much faster than physical inspection and automatically classify problems, enabling everyone involved in the process to work with the same data. Persistent problems can now be identified and addressed more quickly. So, even though it didn't perform perfectly as expected, the team found value in the Einstein artificial intelligence tool and worked to maximize it.
4. Identify simple fixes
Sometimes, giving Salesforce Einstein a little real-world help does the trick, instead of sinking more resources into DevOps. Michelin's tire damage app had a difficult time classifying images of black tires with black treads. Using chalk outlines of the tires, the company was able to drastically improve results. This worked much better than trying to tweak the algorithm.
5. Augment, don't replace, humans
Another company, recruiting firm CPL, wanted to improve recommendations of job candidates for its staff of 3,000 recruiters across Europe. Recruiters winnow a database of approximately 2 million candidates to the five best fits for a particular job -- the better the fit, the less time a client spends vetting hires and asking for more options and the faster CPL and recruiters get paid. The best recruiters can tweak these search filters to quickly find good results.
Kevin Sweeney, CIO of CPL, decided to focus on augmenting everyone's ability to quickly identify good candidates by integrating Einstein artificial intelligence. This enabled people to spend more time talking with the best leads with less time and effort.
6. Augment, don't replace, humans' favorite tools
When CPL launched its Salesforce Einstein AI recommender, it appeared in its recruiting dashboard as a option. CPL told recruiters it could more quickly identify candidates they should look at by clicking on this button. Many recruiters complained the recommendations were not as good as the ones made using other tools. "We positioned this as something extra to what they already had," Sweeney said.
7. to spend time sharpening AI
Sweeney said writing Einstein AI apps is different than traditional IT apps, because AI delivers probabilities rather than clear answers. "With coding apps, you know what the answer should be," he said. "AI is a bugger to debug because you don't know what the right answer is. It took a lot of time to work out if a particular model made sense."
8. Keep experimenting
Identifying ways to streamline the process of experimenting with Salesforce AI models is important. This can include making it easier to ingest data, score the weights of data used in a model and improve the workflow around using the models. Even when a model works well the day it is implemented, results can get worse as the business or market changes.
"We rescored our model 50 times," Sweeney said, "so there is no real concept of failure."