This interview is with Vin Mitty, PhD, Sr. Director of Data Science and AI, LegalShield.
For Featured readers, how would you introduce your current role and the types of AI and data problems you tackle in legal services?
I work as a Senior Director of Data Science and AI at LegalShield. My role connects data, product development, and assists people with legal needs in accessing affordable legal services. We serve millions of members, so our focus is on providing cost-effective solutions and solving real problems for our customers.
The AI and data projects I lead usually fit into a few main areas:
- Understanding and predicting behavior on a large scale, such as identifying who might leave, who needs help early, or where legal service demand will rise.
- Decision support, where we create models that help teams set priorities and act at the right time instead of just reacting.
- Building trust and encouraging adoption. Since legal is a cautious and sensitive field, we don’t pursue technology (AI or otherwise) just for the sake of it.
In the end, our goal isn’t just to add more AI; we want fewer surprises, better results for our members, and systems that people trust and use.
What experiences or decisions most shaped your path into leading Data and AI in a regulated industry?
I have worked in a few regulated industries: banking, government, and now, legal services.
The first realization was that a technically perfect model can still be a total failure. I spent years watching incredibly smart teams build sophisticated models that were technically flawless, but no one trusted them, no one used them, and they didn’t change a single decision. I understood that being right doesn’t mean anything if no one trusts us. To actually move the needle, your work has to be explainable and defensible.
The second experience was a conscious choice to step outside the “Data/AI Bubble.” I started spending more time with legal, compliance, and customer-facing teams than I did with other data scientists. Sitting in those rooms changed my perspective. You start to see why the “caution” exists.
Finally, I decided to treat trust as a design requirement, not an afterthought. In a regulated environment, you don’t get trust for free. You have to build it into the model, the process, and every piece of communication, especially with probabilistic technology like AI, where the answers to the same questions can be a bit different every time. I see accuracy and consistency as core requirements in our industry, not just nice to have.
When setting AI strategy, how do you choose the first high‑leverage use case based on what has worked in your own launches?
I don’t begin with AI. I start by looking at where the business is already struggling.
The first valuable use case is often a daily decision that people aren’t confident about, such as prioritizing, managing retention risk, allocating resources, or dealing with churn. If teams are debating, guessing, or just going with their instincts, that’s a sign.
I also ensure that the first use case is small and low-risk. I look for something that helps people make decisions, not something that takes over their jobs. In regulated settings, trust is more important than complexity. A model that people understand and actually use is better than a perfect one they tend to avoid.
I always ask a simple question: if this works, who will do something differently on Monday? If there’s no clear answer, it’s probably not the best place to start.
The launches that succeeded for me led to quick changes in behavior, had someone clearly in charge, and used data that people already trusted. After that, you can take on bigger and riskier challenges.
Can you walk us through a specific implementation where LLMs turned unstructured customer feedback into decisions that stuck with non‑technical teams?
We built a Voice of Customer system that made it possible to use unstructured feedback on a large scale.
We pulled together call transcripts, surveys, app reviews, and third-party reviews into one pipeline. We then used LLMs to summarize and classify that feedback into clear customer issues and themes. The key was normalizing everything, so a call transcript and an app review could be compared on the same footing.
Non-technical teams are finding this useful because of how we are sharing the results. Executives receive easy-to-read trends and changes in volume. Product, CX, and operations teams receive monthly summaries that use real customer language and match their areas of responsibility. For more detail, we write short narrative reports instead of making people use dashboards.
Things are changing as teams have stopped arguing over individual anecdotes and started focusing on the real customer signals. Decisions about onboarding, messaging, and support are now based on patterns they can see and trust, not just on ‘AI output.’ The technology is fading into the background, and the customer voice is becoming a real part of decision-making.
When a model’s recommendation conflicts with expert judgment, how do you make the call under uncertainty and keep trust intact?
We inherently understand that models are only as good as the data they are fed and cannot handle all the nuances of the real world. Forecasting is a good example. We are now building automated sales forecasting models, which will help our analysts save hours every day from crunching numbers. However, we understand that real-world events are not built into the model—such as a new, large employee benefits client signing on, major natural disasters, or economic turmoil. These are not things our models can account for.
So, we use the model-based forecasts as a starting point and allow analysts to apply their judgment, adjust the numbers, and make the final call. Keeping a human in the loop is a critical step in the real-world implementations of AI/ML models.
Thanks for sharing your knowledge and expertise. Is there anything else you'd like to add?
Each tech boom brings excitement, hype, and fear. We saw this with the web, big data, the cloud, and now with AI. The same pattern keeps coming back. This time, the potential benefits are real, but there are some conditions.
Recent MIT-related research on GenAI pilots shows that most projects do not deliver measurable results. The biggest challenges are integration, trust, and discipline—not the model itself.