Centriam’s Data Science team recently authored a new white paper on the massive opportunities for modern machine learning algorithms to improve customer experience. This paper argued that newer methodologies such as gradient boosting, deep learning, and multi-arm bandits dramatically outperform regression, text analytics, A/B testing, and other methods commonly used today. In line with Centriam’s findings, McKinsey recently reported improvements of up to 62% using deep learning and other modern methods compared to existing predictive analytic methods. Imagine if you could increase your save rate, upsell attachment, or survey completion rate by 62%!
However, many organizations are failing to realize those potential gains. There are many reasons for this, including lack of scored data, absence of deployment tools, and a dearth of data science talent. But just as important are the potential “human roadblocks.” Today I am focusing on one of them — algorithm aversion — and how this behavior in your co-workers (and maybe even yourself) can be impeding the benefits of machine learning in your organization.
Defining algorithmic aversion
Algorithm aversion is simply the desire to use human judgment after you have been shown an algorithm that outperforms humans on the same task. It has been demonstrated repeatedly in laboratory settings (see references at the bottom of this blog) and several data science leaders I recently talked to about this phenomenon all had similar real-life experiences where algorithm aversion negatively impacted their work.
Algorithms are used a lot to improve customer experience programs, including selecting which customers will benefit the most from a limited time offer, predicting which members are most likely to leave in the next 60 days, or identifying which members are most likely to reach an elite status level. Regardless of what is being predicted or classified, algorithm aversion can occur. And this aversion costs real money as people resist, subvert, or simply ignore the business process modernization being pursued.
So how do you pivot your colleagues from aversion to acceptance?
Discover four machine learning methods that can boost customer experience - Download the free ebook!
The same research that demonstrated the prevalence and severity of algorithm aversion, also suggests some ways it can be held in check.
- Provide people influence over the prediction. When people are allowed to modify predictive output, they are both more likely to use algorithms and have more confidence in the algorithm’s predictions. These benefits are realized even if people can only make small changes — as low as two percentage points. By allowing small changes, you still achieve most of the performance benefit of the algorithm, while greatly increasing human compliance.
- Focus adjustment on the unknowns. The best time to adjust an algorithmic prediction is when you have access to information not available to the algorithm. For example, imagine an algorithm that uses past purchases, service history, and demographics to make predictions. The algorithm is almost certainly better than I am at taking all those factors into account. But let’s say I am on the phone with a customer and learn about how satisfied they are with my brand and how much effort they exerted in their last transaction. Those two pieces of information are not available to the algorithm. If I choose to adjust the prediction, that is where I should focus. I should not worry about gender, income, or what they last purchased, as the algorithm can almost certainly weigh those factors better than I can in my head.
One call center we work with is trying something that combines these two points. Selecting the “best” save offer is done by an algorithm and presented to the specialist, but the specialist can override the offer based on information gathered during the call. Some refer to this as “the other AI” – Augmented Intelligence – where the human is augmenting the intelligence of the algorithm.
At Centriam, we believe there is a tremendous opportunity to use algorithms to improve customer experience. We also believe it is important to take the human element into account. Sometimes giving up a little algorithmic accuracy to gain organizational adoption is a worthy tradeoff in a great customer experience investment.
Two research papers, covering a total of seven different studies, were repeatedly referenced for this blog. If you want to better understand the “state of the science” behind algorithm aversion, I recommend you start here:
Berkeley J. Dietvorst, Joseph P. Simmons, Cade Massey (2018) Overcoming Algorithm Aversion: People Will Use Imperfect Algorithms If They Can (Even Slightly) Modify Them. Management Science 64(3):1155-1170.
Dietvorst, B., Simmons, J. P., & Massey, C. (2015). Algorithm Aversion: People Erroneously Avoid Algorithms after Seeing Them Err. Journal of Experimental Psychology: General, 144 (1), 114-126.