By Micaela Kaplan
Creating an ethical business and customer experience (CX) begins with understanding the interactions that impact the business every day. To do so, many leading organizations collect massive amounts of data and use quality assurance to ensure that agents are behaving appropriately on calls. It’s easy to assume that this attention to customer interactions would be enough to stop negative encounters. While it’s true that monitoring all calls does provide a more complete picture of interactions, call centers do not exist in isolation. They exist within an imperfect world, where misogyny, racism, stereotyping, and other forms of discrimination happen on a daily basis.
Everyone is impacted by these societal biases and ideas, and whether we like it or not, this will show up in the everyday interactions that we collect. The ability to collect and analyze interactions in a business provides data – an invaluable tool in the fight to create and maintain an ethical business and build a more ethical world. It is crucial to let the data guide us to the insights that drive businesses, even when those insights involve taking a hard look at some of the biases that inherently exist in your own business.
Acknowledge the bias
A hot topic in data science today is the idea of building ethical artificial intelligence. Google, Microsoft, and other industry leaders have begun creating guidelines for building and maintaining ethical AI. These processes have been driven by several high-profile ethics cases, and new AI ethics labs are popping up at universities worldwide to help evaluate, understand, and address these issues. As data scientists, we are learning to address the biases of models that provide insights for downstream tasks. Creating an ethical business means acknowledging all the data you’re provided and understanding the biases it may have. While no algorithm or model will be perfect, we can help minimize the harm the output of a model may cause.
The first step, as always, is acknowledging the problem, and we need to go deeper than blaming the problems on the AI-driven models that we use. Our models are trained on data from the real world and therefore reflect real-world biases and problems. Unlike what movies and TV tell us, these models are unable to think, reason, or learn for themselves. Models are nothing more than some complex math based on the data provided—a biased model is the result of biased data. It is imperative that models that are used to drive our analysis are looked at through a critical lens.
Models can point us to real and concerning dialogue in our data, such as “I’d rather speak to a white person” or “Can you redirect me to a male representative?”. Interactions such as these not only indicate dissatisfaction on the part of the customer, but also indicate situations that might cause real harm to the agent on the other side. Traditional dissatisfaction language, such as “frustrated” or “angry”, won’t be able to find these types of nuanced, rare interactions. Finding these interactions will allow for changes in the organization that can help protect your agents and provide a comfortable space to do their jobs – if the organization chooses to listen to its data.
Drive change starting on the inside
While we can work to protect agents from outside harm, creating an ethical business also means preventing harm inside your business. Taking the time to unpack, understand, and address the various biases within your system can help create fairer practices for all, impacting both employees and the brand itself.
The call center provides a great use case for this, as agents are often scored on a variety of measures. However, many agents are set up to fail from the beginning by circumstances out of their control, such as poor speech recognition due to an accent or needing to talk fast to fulfill daily call quotas. Let the data guide you to these problems as they appear, and then take the time to fully understand and address why they’re missing the mark. You can begin by asking yourself questions like:
- Which agents are consistently scoring badly? Are any company practices driving these scores?
- Are there any populations that you leave out when you build your ideal agent? Your worst agent?
- How is the definition of dissatisfaction (or other emotions) reflective of the people around it? Does capturing dissatisfaction inherently capture racism, misogyny, or other biases?
In asking these questions, we may be confronted with uncomfortable truths about how our organization runs or how we evaluate success. This does not mean that you run an inherently unethical business – we all live in an imperfect world, and we often do things that cause harm we don’t intend to cause. For example, are you noticing that many of your highest throughput agents are scoring low on some types of language? How might you be able to allow your agents to slow down their speech while talking with customers?
While this may seem counterintuitive to a numbers-driven field, a speaker who is recognized better by a machine might also be recognized better by a human, leading to more genuine interactions. Have you noticed that many of your agent types for scoring, good and bad, reflect a certain type of agent? It could be time to rethink your strategy so that you can capture the universal good and bad behaviors of your agents.
Don’t overhaul data models right away—question them first
To say the only way to build a more ethical business is to change business practices is false. That process takes time, money, and commitment—and won’t always solve every problem it attacks. Instead, we need to think critically about our everyday practices, and to be intentional and transparent in our actions. In doing so, we will be able to have data-driven, introspective conversations at all levels of an organization about what is and isn’t working and how you can improve your business.
Arming ourselves with data means more than just looking for patterns. We can build tools in our QA systems that allow us to track changes in patterns over time, providing us with a number-driven measure of how well our changes are working. After finding problems in the data, build measures that capture what you’re trying to change, and track the directionality of that measure over time. Whether the number is growing or shrinking provides even more data to help tackle bias in our world.
This data provides us the ability not only to find problems as they happen, but also to measure our internal growth and the success of the changes made to address the problems. As stated in the Ethical Explorer’s field guide, “When we design for convenience and engagement above all else, we only acknowledge a narrow view of the human experience.” Proactively working to minimize the harm caused by outside customers, internal policies, and the employees who implement and are impacted by them, businesses can work towards creating a more ethical environment for everyone involved, and in turn improve customer experiences, employee experiences, and ultimately the bottom line.
About the author
Micaela Kaplan is a data scientist at CallMiner. She received her MS in Computational Linguistics at Brandeis University after graduating with BAs in Linguistics and Computer Science. At CallMiner, she helps bring linguistic insights into machine learning research. Micaela’s role as Ethics in AI Lead for the Research team enables her to work toward a more ethical world, one project at a time.