Machine Learning in the Enterprise: You Can’t Afford to be Wrong

(Lu Mikhaylova/Shutterstock)

As the final moments of Rutger Hauer’s tears in the rain monologue come to a close in Blade Runner, Netflix (or your streaming service of preference) has lined up some recommendations for your next viewing choice. From 2001: A Space Odyssey to The Matrix, the site’s algorithms find you similarly cerebral films that you may enjoy…or you may not.

The stakes are low in this situation. If you end up watching and disliking The Matrix, chances are you won’t cancel your monthly subscription; you will simply be more skeptical of Netflix’s algorithmic recommendations in the future and continue on with your day as if nothing happened.

In the B2C environment, machine learning is a constant presence in the end user’s experience. Recommendations that provide a variety of options for consumers are seemingly infinite. However, these recommendations serve as add-ons to maintain interest or upsell other products and have little to no effect on the consumer’s work or personal lives.

Machine learning recommendations for the B2B enterprise market are a different story.

These suggestions are provided to knowledgeable buyers, not emotional decision makers. Additionally, B2B relationships are maintained and built over time. There is greater financial risk of customer defection upon providing suboptimal recommendations. By combining the latest machine learning techniques with sales representative expertise, the objective is to minimize risk and provide a tool that is able to increase revenue from the existing base of customers.

In order to adapt to the demands of the B2B marketplace, organizations must take B2C machine learning principles and understand how they can work in their own environment.

Personalize for Customer

As an example, we can look to the algorithms used in self-driving cars. As George Hotz points out, current algorithms are based upon what an engineer believes to be safe, proper driving, not how a human would normally drive.

(etraveler/Shutterstock)

There is a fundamental issue with this mindset. By focusing on the gold standard, law-based driving style, car manufacturers are ignoring typical customer behavior and removing the grey area between the laws of the road.

For B2B enterprises, customer behavior needs to be the defining metric on which the algorithms are based. I have seen customers with tens of thousands of SKUs being sold to thousands of global customers across dozens of different industries. A uniform recommendation process for every customer will lead to a much lower success rate than desired.

The success rate can be improved by segmenting customers into groups, such as product, region, size, transaction history and market. Segmentation, however, is not granular enough. A good recommendation for one customer may not be applicable to all customers in the segment. This is where personalization comes into play.  Using customer attributes along with their purchasing history allows for the tailoring of recommendations to each particular customer.

Each customer sees completely different recommendations personalized to his or her needs.

Who is the End User?

When your recommendations incorporate data-based insights, the next step is understanding where these suggestions are going. This process is much easier to identify, but becomes more complex when acted upon.

After the analysis, where are the insights sent? I have seen a few different models, but the two most common are: sales rep oriented and end user/customer oriented. There is a vast difference between sending recommendations to your sales team as opposed to directly to the customer.

Sending suggestions directly to the customer can be high risk.  No matter how state of the art your algorithm is, it is still beneficial for a person with business knowledge to vet the recommendations.

No one in the organization knows more about the customers than sales. Their success is reliant upon in-depth knowledge of the customer’s industry, competition, behavior and needs. Collaboration with the sales team is an invaluable resource in crafting the algorithm’s metrics with the human element incorporated. In turn, they will be able to double check if the recommendation makes sense for the situation.

Use their experience as a guide to monitor if your data analyses are straying from reality.

However, since they are the specialists in their particular domain, they are often reluctant to change and may refute the need for help in completing their tasks. For decades, sales has been a human interaction business, so why should a “computer” be able to provide better information than their own experience? This distrust can be overcome by the fact that machine learning is able to find novel suggestions, especially for small to mid-sized accounts that may be overlooked.

Collaboration with the sales team is invaluable for fine-tuning algorithms

A sales rep’s experience is limited to their accounts and their region. Machine learning, on the other hand, is able to find patterns in large amounts of data. A recommender algorithm uses the experiences of all sales reps to provide both innovative and surefire recommendations.

Additionally, sales reps may have a large number of accounts. It can be difficult to go through all accounts by hand to determine what cross sells will provide the biggest pay off.  Machine learning can highlight what recommendations will lead to the greatest increase in revenue.  These insights help the sales team determine where they should focus their time and effort.

Know what is involved in your recommendation and who is receiving it. This foundation is key to minimizing suboptimal recommendations and increasing confidence in your system.

Was This Helpful for You?

B2C providers get direct feedback from their end users all the time. Customer satisfaction surveys are a pervasive part of the shopping or viewing experience. Their goal is to keep you on their site as long as possible to get you to buy and consume more. It’s obvious, but it can work. And yet, most people don’t recognize that behind the scenes, these short surveys or ratings or reviews play into the system and lead to updates, clarifications and streamlined user experiences.

Unlike the factors mentioned above, there is little difference here between the B2B and B2C worlds. With the rate of change in every sector of business today, your systems will become obsolete and irrelevant to your end users in no time. I say “will” on purpose. It’s an inevitability.

To survive and improve, feedback needs to be a core piece of the process. Whether it’s from the sales team or the customer, feedback is necessary on a regular basis; it cannot be an annual occurrence, or even bi-annual. Sure, updates take time, but at least knowing where the problems are helps keep the system in check and allows your team some time to adjust accordingly.

In order to ensure accuracy, find the primary pain points of your users, target them with your inquiries, and provide respondents with space for their own explanations. Receiving feedback in their own words will help fine tune future updates to their needs and wants.

We can all agree the goal is to provide recommendations that lead to better business performance. By understanding your customer on a personal level through customization and feedback, you can tune your recommendations to their needs and reduce suboptimal suggestions.

Ignoring these foundational best practices leads to suboptimal recommendations and bruised confidence in your system. If that happens, your recommendations truly do become like tears lost in the rain.

About the author: Ravi Mirchandaney brings more than 20 years of experience in the software business to Vendavo. Previously, he was Senior Vice President of Engineering at C3 Energy, a SaaS Grid Analytics software provider. Prior to C3, he held the position of Group Vice President of Engineering at Oracle, focused on CRM, and before that, Ravi was Vice President of Engineering at Siebel Systems. He has held research positions at Yale University and Shell Oil Research Labs in addition to being the author of more than 25 technical papers and patents in fields of parallel, distributed systems and databases. Ravi holds a Ph.D in Computer Engineering from the University of Massachusetts, with a focus on distributed systems.

Related Items:

Report: Machine Learning Driving AI

Behavior I/O: Using Machine Learning to Empower Human Learning

Unstructured Data Miners Chase Silver with Deep Learning