Machine Learning and the Future of Client Success

Every day of my career I’ve been consumed by client success…

I’ve thought about how to deliver better, clearer, more personalized support — and how to do so while keeping costs manageable. Historically, that meant maintaining the best possible knowledge base and troubleshooting guides, hiring and training top client care agents, and building systems for self-service — all the while monitoring those same systems for anything that left clients feeling abandoned or exasperated.

Those strategies are as critical today as they’ve always been, but the challenges are growing:

• How do we quickly understand how symptoms correspond to solutions, especially as products proliferate?
• How do we digest increasingly complicated troubleshooting data and share it with agents?
• How do we leverage what we know about our clients to fast-forward the process of getting them the information they need?
• And how do we unite the different portals (phone, chat, email), while building out new methods of engagement for mobile, social, and even virtual reality?

These challenges aren’t specific to any one industry. Whether a client is reaching out to the cable company, the pharmacy, or the vendor of their enterprise software, they want to feel heard and helped immediately. That means delivering client support that combines personalization with speed and precision. I’ve written about these issues before, but let’s dive deeper.

Data First

As machine learning comes into its own, more and more of us are seeing its potential to transform how we organize and deliver client support — in ways that could benefit clients and support providers alike. It starts with the data…

As a baseline, we want to take advantage of the data from disparate sources to train models for making predictions about what clients require now and in the future. Many organizations I work with have volumes of this proprietary data that’s simply untapped — data that can be mined to make meaningful predictions. In general terms, the data falls into three categories — with some overlap:

1. The data about the clients themselves

2. The data about their contact with provider support

3. The data that you can use to solve their issues

The first category comprises routine demographic and organization data, but also includes detail about a client’s preferred contact portal, their privacy stipulations, how long the person or organization has been a client, their contacts within your organization, which products and services they use, and how those products and services are configured.

The second category — data about their contact with support — might seem more narrow, but it’s remarkably diverse: which product or service is causing issues, time and date of the call, how long the person has been waiting for a response, how long it’s taken to resolve issues with this client in the past, whether the person re-opens closed issues, how long since they last contacted support, who handled their previous interactions, and so on.

qAnd then there’s the third category: Data that you can use to solve their issues. We typically think of this as Knowledge Base articles, troubleshooting guides, white papers, FAQs, and forum postings — but increasingly we can look to unstructured data around previous interactions with clients: the text or transcription of previous support tickets, and even data scraped from social media sites where your own organization, products, and services are being discussed.

(In many organizations, we would include a fourth category for public or externally available data: weather forecasts, road congestion, interest rates, and so on.)

Enter Machine Learning

Machine learning is about feeding data into models that offer reasonable predictions. For client support, those predictions give client care agents the opportunity to drill down quickly to the client’s specific need.

But doing that means more than looking deeply at the current client’s profile. It means comparing across all clients to aggregate an understanding that lets us say, “clients like this run into issues like these.” For example, clients who call each month to verify their payments could bypass the phone tree to get routed directly to a person who can put their minds at ease. Or a more complex scenario: Machine learning algorithms might determine that clients who have two particular products installed and who rarely call for support and who are calling in the middle of the night tend be to reaching out for advice about a particular configuration parameter related to an upgrade.

What better way to build rapport with those clients than to answer their phone call or chat inquiry with the answer that they need already in hand. And even if machine learning doesn’t get us to the answer right off the bat, it could still arm us with intelligent guesses and help us adjust our questions to quickly route clients to the information or expert that can help them. (I think visualization could play a key role here, but more about that in a bit.)

The Challenges

Building rapport, fostering personalization, and reducing time and cost are wins for clients and support teams alike, but what are the challenges?

We know that machine learning relies on volumes and volumes of data, and the ideal for training the models is clean, clear, structured data. We’ve seen that all manner of client and support data could be fed into the models — if it’s available. Here’s where established organizations can have an advantage, especially those who’ve been careful to maintain deep history of client interactions. That proprietary data — often maintained on private clouds behind the firewall — can act as a superpower for those organizations, especially if they’re able to set up high-performance machine learning systems that can sit side-by-side with the data.

Another challenge is how best to configure the systems themselves to maximize the ability to capture and use the data that can feed the models. At IBM, we’re seeing an increased demand for our easy-to-deploy appliance systems that can manage data for machine learning flexibly across private, public, or hybrid cloud.

And, of course, at the heart of the configuration challenge is the issue of data governance. Here the big questions are: How do we classify and unify the data from different support portals? How do we make sure that the data we’re collecting enables the specific client support decisions we want to make? And how do we create secure systems that still give our data science teams smooth access to the data they need to build and train their models?

Over the Horizon

Whether riding in autonomous cars, buying books, or fielding issues from clients, the great promise of machine learning is about anticipation. Above, I mentioned scenarios where our client care agents could help solve issues as soon as they emerge — but that begs the question: Can we solve issues before they emerge? Where can we use the power of anticipation to get ahead of client issues?

As our machine learning models become more sophisticated, the nature of anticipation is likely to evolve as well. It’s one thing to anticipate the need for engine service or the need for a security patch. It’s another to anticipate the deeper needs of people and organizations as they evolve through time. “Two years after clients ask us about x, do they tend to ask us about y?” How might we observe how clients learn and grow? And how might we tailor our support to their growing knowledge — and to the drift of their curiosity?

Inevitably, I think that same curiosity will extend to wanting to understand how we make our decisions to support them. This brings us back to visualization. We can imagine offering tools to our client care agents that let them describe a client’s issue in natural language and which respond with visual representations of the options for help that we can offer those clients — including the weights and probabilities the system is calculating (similar to Watson’s way of visualizing potential answers on Jeopardy).

In the spirit of self service, we can then imagine giving clients direct access to these tools and letting them describe the issues they’re encountering in their own words. The visualizations could give the client clear feedback that would allow them to adjust their descriptions or drill deeper. And their responses would inevitably improve the models themselves.

Unstuck, Unfrustrated, Unconfused

It won’t come as a surprise to anyone reading this post that — even with all the hype — machine learning really is poised to enact profound changes in the way we live, work, and think. Certainly the potential benefits can sometimes feel mingled with real hazards, and as always we need to proceed with care. But to the extent that machine learning allows all of us to spend less of our time stuck, frustrated, or confused, it means more time spent productively engaged in the deeper work that we’re eager to accomplish. To my mind, that’s the real promise of machine learning — for client success and beyond.

Whether you’re new to machine learning or an old hand, I encourage you to check out IBM’s Data Science Experience and our Machine Learning Hub. And I invite you to follow me and this publication on Medium for further updates.