By: Mark Anderson
Microsoft Dynamics 365/CRM customers have a number of machine learning tools at their disposal thanks to the solution’s proximity to technology from Azure Machine Learning Studio and the Cortana Intelligence Suite (formerly the Cortana Analytics Suite). MSDynamicsWorld has covered a few of these capabilities already, such as our article earlier this year on ranking CRM leads using Azure ML and Microsoft’s progress in offering pre-built tools for CRM users like Customer Insights.
However, for all its promise as a strategic asset, machine learning is sometimes viewed as a black box. Machine learning (ML) is not just a feature that you can select by checking a box in a CRM menu somewhere. And it’s not a feature that one activates and never thinks about again. Rather it’s a technique that power users and decision makers learn over time how to best apply to their own workflows. Over time, ML helps a business application like Dynamics 365 sales, marketing, and customer service apps learn and improve their predictive capabilities through access to growing sets of data, plus some trial and error.
Machine Learning in CRM – best practices
Stefano Tempesta, VP of Engineering at EF Education First and a Microsoft MVP, offers some tips for users to best apply machine learning tools within CRM.
First, understand that machine learning is really a tool built on data. So before a user begins reaping the rewards of machine learning, she needs to understand the data sources that the ML technology of Cortana Intelligence Suite or Azure will be drawing on.
“First you want to identify your data sources as much as possible,” Tempesta says. “If your information is missing relevant data or is inconsistent, then the training process will not be accurate.”
The second tip concerns input data that is incomplete or patchy. Perhaps some members of the sales team don’t fill in their CRM fields completely, or maybe one of the data sources is unavailable at times. In that case, he says, “(Azure) Machine Learning provides powerful tools for splitting data, generating other variants that can fill in the gaps, even making up new values by looking to average conditions.”
All of these data transformation tools appear in Azure ML under the rubric of Data Transformation / Manipulation.
Microsoft has considerable documentation, tutorials, and videos detailing the steps one takes after gathering the data that will train a machine learning model.
Tempesta’s third tip applies to a user who is on the other end of the machine learning model. There’s always the temptation, he says, to over-populate one’s models with many data sources and then attempt to extract a wide array of insights from these various inputs.
Rather, he says, keep a clear head about what data is going into the model. Don’t just throw in a bunch of data because you have it. And keep the insights you extract from the machine learning model reasonable, given the inputs you provided. Machine learning ultimately is about teasing out complex patterns from large data sources. But it’s not a magic wand. It’s only as “smart” as the data sources you train the model with.
“Keep it simple,” he says.
Machine Learning in CRM – use cases
The quality of machine learning available in a company’s CRM sometimes comes down to that company’s ability to use CRM, says Ryan Chynoweth, data science consultant at Northwest Cadence in Bellevue, Wash., A machine learning model, in other words, is only as good as its input data.
“If somebody doesn’t track a sales pipeline, and all they do is after the fact enter the sales they got, … then they call us and say, ‘Can you predict of all these people we contacted, who’s the most likely to close?’ If the only data set you have is the people who closed, then you can’t predict who’s not going to close,” he says. “Your use of CRM dramatically influences what kind of predictions you’re going to be able to make.”
Adds Steven Borg, Northwest Cadence co-founder, “It’s hard to convince sales people sometimes” to provide full sets of data in their CRM, not just the bare minimum, in this case, of sales that closed. “When we look at machine learning and artificial intelligence data analytics, where we are today is a supplement to humans. … You can’t replace your sales people with CRM integration with some data science program. But what you can do is make them more effective sales people.”
“It’s workflow prioritization – what should I be working on right now, what’s the most important,” Chynoweth says. “You can use machine learning to say this is your #1 priority, this is your #2 – if the data is there, and if it’s available to use.”
In addition to the CRM lead-ranking example Tempesta previously detailed for MSDynamicsWorld, he says he’s also run an Azure ML experiment that uses airline routes and prices from online sources to predict the fluctuations of ticket prices. Tempesta says he’ll be presenting this experiment at Directions EMEA in October.
Which brings Tempesta to his final tip: If you can do something with a spreadsheet, it’s not worth the extra time and trouble to train an Azure ML model and extract insights from that.
“You can’t do this with an Excel spreadsheet, because the criteria to analyze are so changeable, that it’s really impossible for a human to analyze this information row by row,” he says. “You need the power of machine learning to estimate when this flight ticket price change will happen.”
…