The Puzzle of Scaling AI in Insurance

Building a performant model in insurance is one thing. Building, deploying, maintaining, and governing hundreds of models in production is another challenge entirely.

AI is already integral to insurance and has been the driving force behind much of the sector's innovation. However, AI adoption is a challenging process that brings with it certain obstacles and risks that any insurer must confront to be successful. 

Selim Cavanagh, Mind Foundry’s Director of Insurance, was interviewed by Jonathan Swift of the Insurance Post to discuss some of the challenges the insurance industry faces around adopting and scaling AI and how Mind Foundry has been working with insurers to overcome them. What follows is a transcription of that interview, slightly edited for this blog format.

(Watch Selim’s full interview with the Insurance Post here.)

Can you explain the origins of Mind Foundry and why the business chose to work within the highly regulated world of Insurance? 

Mind Foundry was founded in 2016 by AI professors Stephen Roberts and Michael Osborne as a spinout from the University of Oxford. They wanted to take the techniques they learned in academia into the business world. One of their fundamental concepts is transparency, or being able to understand exactly what's happening with your AI. They decided to take this, and other technologies important to Responsible AI, and apply them to what we call high-stakes applications where the decision of an AI either affects an individual's life or gets made at the scale of populations. These are massively important, and it has led to us operating today in several verticals, one of which is Insurance. As you said, it's a highly regulated industry, and transparency and human-AI collaboration are critical components of making that work fairly for consumers and regulators in different countries.

What do you see as the biggest challenges that Insurance companies face when adopting and scaling AI to actually provide benefits to the business and to also their customers? 

We went out to the industry in Q4 of 2023, and we asked that question. We're really trying to understand exactly what's happening. Things are moving so quickly with AI, and we didn't want to get that wrong. We asked leading insurers across the UK, Europe, and the US several questions. Firstly, "how mature are you in AI on a scale of 1-10?" And secondly, "what are your biggest challenges in deploying AI?"

For the AI maturity question, we had a median response of 4 out of 10 which tends to mean, “We've experimented, we've tried to scale it, and we managed to do it in a couple of areas but we've not created an industry-wide model or belief that we can get this right. But we want to, so we need help.” 

Then we asked them what they were most worried about. We were quite surprised, but 100% of respondents said “governance” was their number one issue. Governing AI to help them scale was most important. The second most important was "skills", and this reflects the real skills deficit we are seeing with AI in the market. How do you actually do this stuff, keep scaling, and keep building? Third on the list of worries was "integration" and the question of “how do we actually take these models and this AI we built and make it work in the real world with our existing systems and methods of deployment, distribution, etc.”

The final important factor for all models is model deterioration. Once deployed, these models stop working over time, and you have to keep going back to them, retraining them to make sure they're fresh. This is for a couple of reasons. One is obviously the ROI that you expected when you deployed it. The second one is the question of whether the model is still compliant, fair, and delivering what you expected. 

How does Mind Foundry help Insurers feel confident that they can explain to consumers and regulators how they're using AI in a transparent and fair way?

We've invested a lot of time, money, and expertise into ensuring that all the models we help insurers build and deploy are absolutely transparent, fair, and compliant with the regulations of the markets in which they operate. 

We build, deploy and manage responsible AI for a number of different business areas, including pricing, claims and fraud. For example, we deploy a service and/or system across models and work out exactly what they're doing and the key features in those models. Insurers in the UK typically use hundreds of models now for pricing alone. And of course there are models to counter fraud, handle claims, and address other critical areas. With all those models, the challenge is understanding each one individually as well as the cumulative effect of all of them on your business. You need to be able to do that in order to know if they're working and to explain that in a way that makes sense. 

That’s a key area where we help customers.  By exposing those kinds of insights to them first, we enable them to make informed decisions around bias, drift, and other factors that can impact the performance of their AI and create something that's ultimately profitable. This means they are more compliant and transparent and meet the needs of both the regulator and the consumer. We believe the direction of travel is absolutely more and more towards that, as more regulation comes into the market, whether in Europe, America, or the UK. 

The Mind Foundry solutions span pricing, claims management, fraud detection, and governance. Can you give a small insight into what you can bring to each of those?

On an individual level, we do a lot of work for Insurers in those areas specifically. We can build those models where they don't have access to enough resources, expertise, or skills. We've built models in pricing around retention, working with actuarial experts to deploy new pricing models both in the distribution piece and the burning cost models.

Claims is also a really interesting area. We’ve built models in areas like large loss prediction, where if you can work out what's happening early, where maybe 30 - 40% of your claims sit, then that's a really powerful piece. We've built models that analyse driving behaviour and work out how people drive, why they drive that way, and then we deploy applications to help them improve their driving in the future and reduce the number of claims.

We also have really interesting examples around liability detection. One key consideration in claims is what resources you deploy and when and how you deploy them to address how that claim will change over time. 

The important thing to remember is that when you bring these models together, and they start speaking to each other, sharing information and insight, they all become a lot more powerful. Our infrastructure and systems and the way we build AI at Mind Foundry allow all of that to happen.

What advice would you give to Insurance companies looking to build their own solutions themselves and first considering an off-the-shelf alternative? 

We've seen a big shift in that over time. Insurers have lots of clever people, very analytical people, a lot of data, and many systems and processes that can analyse that data and turn it into insight. So where they start is, “Let's build this stuff ourselves. We want to understand it before we go to market and start to buy solutions and find things that may do this more efficiently, quicker and more safely.”

We’re seeing it emerge from the other side now. We've gone through this experimental period, and we're now thinking about how to scale this stuff. How do we make it really efficient, work everywhere in our business, and deploy it everywhere in a very safe manner? 

The shift is beginning to happen towards outsourcing and focusing those people internally away from building these solutions and towards building new models, thinking of new use cases, and meeting the business's needs.

What can we expect from Mind Foundry over the next 12-24 months? 

Over the next 12 to 24 months, you'll see us embedded in more places in the insurance market, essentially helping insurers scale up their existing model ecosystem to deliver more models in more places and more efficiently.

You'll see us in governance everywhere, helping insurers ensure transparency, trust, and ethics in their product delivery to market. Finally, you'll see some interesting things emerging in quantum, where we're seeing a rapid change in that technology's cost and value metrics and are actively helping insurers who want to be first in line for that technology understand which use cases for quantum will provide the most value when that technology is more readily available.