Customer personas have been around for many years, as a useful design tool to understand customers and create better experiences. They typically rely on a body of customer research, and on having skilled designers capable of identifying the patterns and traits which are most meaningful to the situation at hand.
So can we create a persona using AI? We’ve been experimenting with IBM’s Watson product recently, feeding in customer interviews and contextual inquiry outputs to see what it can do for us. And we’ve learned something about how we create personas as well about AI.
We’ve used 2 of the Watson components; Personality insights and Tone analyser. Personality insights was the main focus. This is what we did and learned:
- We recorded a range of customer research interviews and inquiries, with their permission of course.
- We decided against feeding in the raw audio, though that’s an option in future and would pick up more sentiment. Why? A good chunk of the audio was our own researchers’ voices, and filtering that out was too hard for our initial testing. So we had the customers’ words transcribed, then fed in the text.
- The 2 main outputs are a summary of key traits, both as words and on personality scales. Then there’s a more detailed ‘wheel’ of a wider range of traits (see image).
- Some of the outputs aren’t useful. Statements like “Unlikely to enjoy country music” might be useful in some personas, but none that we’ve ever done!
- However other traits are much more interesting, e.g. is this person logical or emotional? Are they gregarious, excitement seeking or cautious?
- Spotting the dichotomies, i.e. the important differences between different customers is a manual process if you’re using the basic visual output – you have to compare the differences in the image below. But linking via APIs means you can use the data outputs and it’s much faster to spot those traits where the widest differences occur.
- We then filtered for independent variables (there’s no point having personas defined on scales which are all correlated to each other) to find the one or ones which really demonstrated the important personality differences.
- And lastly, we haven’t found a way to easily introduce context to the AI models. In other words, if you’re researching say customer attitudes & needs related to bank mortgage products, then Watson isn’t going to pick out anything specific to that context without more configuration and training than we’ve yet been prepared to attempt for a one off project.