A Digital Twin Should Be Allowed To Change Its Mind
Getting a customer right today is only part of the job.
If someone tells us they love a brand in April and stops buying it in October, I want to know what happened in between.
Maybe the price went up. Maybe a delivery went wrong. Maybe they still love it, but their life changed and they no longer need it.
That difference matters to anyone trying to understand customers. It matters even more when we build a digital twin: an AI representation intended to help us explore how someone might respond.
A convincing answer today is useful. I’m interested in what happens when tomorrow gives that person a reason to answer differently.
The same cancellation can mean different things
Take a fictional customer, Maya, who buys a monthly coffee subscription.
In April, Maya likes the coffee, thinks the price is reasonable, and enjoys having it delivered. A digital twin built around those answers might reasonably predict that she will renew.
By October, she has canceled.
Now imagine two possible explanations.
In one, Maya starts a job where good coffee is available every morning. She still likes the subscription. She just has more coffee than she needs.
In the other, Maya’s routine stays the same, but two deliveries arrive late. She runs out of coffee, buys another brand, and decides she cannot rely on the subscription anymore.
The purchase history looks similar. The next useful question is different.
For the first Maya, we might explore a smaller delivery or a longer gap between orders. For the second, we might ask what would make the service dependable enough to try again. Each starts from a more useful explanation than labeling her a former customer and offering a discount.
A preference, a circumstance, and a purchase
Get Lewsearch’s stories in your inbox
Join Medium for free to get updates from this writer.
I would want a digital twin to keep these three things separate.
Maya can like the coffee and stop buying it. She can dislike a delivery experience without disliking the product. She can also change her opinion of the brand entirely.
If we turn every purchase into a preference, we risk misunderstanding the person. If we treat an old preference as permanent, we risk making the same mistake from the other direction.
The interesting question becomes: what changed, and what stayed the same?
That gives us something more specific to investigate. We can ask whether a different offer addresses the actual problem. We can also recognize when we do not know enough to suggest one.
What the research gives us so far
There is real research behind the idea of simulated people carrying experiences forward. In their 2023 Generative Agents paper, Joon Sung Park and colleagues built agents that stored experiences, reflected on them, and used them to plan. Their evaluation examined believable behavior in a simulated environment. It did not establish that those agents could forecast a real customer’s changing brand preferences.
A later study by Park and colleagues, revised in June 2026, built agents from interviews, surveys, or both, and tested predictions on held-out responses. The authors compared survey accuracy with participants’ own two-week test-retest consistency. This preprint offers evidence about representing individuals across questions. It leaves a separate question for customer research: how well can a model anticipate changes over months as someone’s circumstances evolve?
Predicting those changes is the question I want to pursue.
Give change a fair test
For a customer study, I would start with people who agree to be followed over time. Record what they think now, what they actually do, and the relevant changes they choose to share.
Then ask the model to predict a later decision before seeing the answer. Track whether it recognizes people whose choices change, as well as people whose choices remain steady.
A model that predicts everyone will switch can look responsive while being wrong. One that predicts everyone will stay can look consistent while missing the people we most need to understand.
And a fluent explanation after a cancellation should not earn credit for predicting it beforehand.
For Maya, success would mean doing better at anticipating her next choice, with enough evidence to help us ask a sensible follow-up question. We would still need to ask her whether our explanation fits.
That is what interests me about following customers over time. We get to learn where preferences hold, where circumstances take over, and where someone genuinely changes their mind.
I want digital twins that can make room for all three.
