The adage, “garbage in, garbage out,” is as true today as it was 20 years ago. As Thomas Redman writes in Harvard Business Review, “Poor data quality is enemy number one to the widespread, profitable use of machine learning.”
Just because you can teach computers to help normalize data does NOT mean they can do the cleanup from years of bad habits. Whether your team is supposed to be logging information in two systems but only does so in one, whether you want to connect two disparate systems or whether you need to simply make up for a lack of data, there is work to be done before AI can take over.
Think of where you’ve seen AI in action, like the craze for images created by AI. Sometimes they turn out well, yet sometimes they are comically bad. And nothing is funny when you’re trusting AI to help grow your business. Missing out on key pieces of information could lead to disaster.
When we trust AI to do the work: This is an AI-generated image of an insurance agent issuing a policy to a local business owner.
How does this translate to insurance?
Imagine you have one system that records all customer interactions from a marketing standpoint – interactions on social media, website engagement and email tracking. You have a different system that tells you how many times they logged into your self-service customer portal (and what they did there). And of course, the customer could also be interacting directly with their carrier without your knowledge, perhaps buying additional insurance or researching whether their current insurance is good enough for them.
Your main system may also be disconnected from the system that logs the number of policy change requests, questions about their policy, requests for certificate of insurance or auto ID cards and even when claims are made.
Your marketing engine may paint certain customers or prospects as highly engaged customers likely to buy, when, in fact, the customers who are truly likely to buy are barely being touched from an education or upsell/cross-sell opportunity standpoint.
See also: The Risks of AI and Machine Learning
Why does this happen?
For AI to be successful, you need a data set that is normalized and “taught” with certain objectives in place. And you must have ALL required data in the data set to test your hypothesis.
You must also recognize that “AI is inherently probabilistic,” as Forrester suggests in a recent article on Bridging the Trust Gap Between AI and Impact.
Your marketing team may assume that the most active customers/prospects on social media are the most likely to buy more insurance from your agency. However, many agencies have found that those that are the most “taken care of” during the policy change/question interaction are the ones most likely to buy additional insurance. Why? Insurance is still very personal, and in these cases the agent has created a bond of trust with that customer. Without a complete data set, you may get a distorted picture.
“But we are extracting our data and dumping it all into a data lake.”
Another tenet of AI best practices is iteration – systems need to be taught and learn through continuous feedback loops. Andrew Johnson shares more details if you’d like to dig deeper into how these work. If your data is disparate, even if you are extracting it, normalizing it and dumping it into a database, the feedback loop remains manual. At best, you are guided by the observations of your firm’s management team, not complex analysis of actual user behavior.
With quality data, AI can begin making well-informed suggestions. Going back to our example, a system might recommend upsell opportunities based on complete information, better understanding customer sentiment. This hypothesis can be tested and refined through multiple iterations of conditions based on real-life relationships with the customer.
So, can AI solve our underlying data problem? The unfortunate truth is that you can’t even begin to use predictive analytics without good data.
But there’s so much you can do… once the data is right. If garbage in is garbage out, just imagine what the output could be when you start with good data.
For an agency ready to take on the future of technology and allow AI to work for its business, you must make sure the initial data set is clean, complete and housed in the same environment where your users interact with your business. Then, you can really start making progress.