Artificial intelligence has opened up a wealth of opportunities for both insurers and insurtechs, but turning AI’s vast potential into scalable, successful technology solutions is a major challenge.
Tractable is a great example of a company that has managed to do it, with its AI solution for assessing vehicle damage now being used worldwide to process thousands of claims every day.
Co-founder and President Adrien Cohen joins Matthew to discuss Tractable’s global growth, the challenges of working with AI, and what the company has planned for property insurance.
Talking points include:
- Why insurance is a great use case for AI
- Data challenges and training AI systems
- Why insurers struggle to develop solutions in-house
- Spotting opportunities and expanding into different markets
- Convincing customers about new technology
Sign up to the InsTech London newsletter for a fresh view on the world every Wednesday morning.
If you like what you’re hearing, please leave us a review on whichever platform you use, or contact Matthew Grant on LinkedIn.
Continuing Professional Development – Learning Objectives
InsTech London is accredited by The Chartered Insurance Institute (CII). By listening to an InsTech London podcast, or reading the accompanying transcript, you can claim up to 0.5 CPD hours towards the CII member CPD scheme.
- Claim 0.5 hours for listening to Episode 134 of the InsTech London Podcast
Damage assessment with AI – fast, scalable, global – Episode 134 highlights
Matthew: You co-founded Tractable in 2015 with Alex Dalyac and Razvan Ranca. What was the motivation for starting the company?
Adrien: There was a historic moment in artificial intelligence (AI) in 2015 where computers surpassed human performance in image classification. Alex and Razvan were doing machine learning studies at Cambridge University and we realised there was an opportunity to take the AI technology from the lab and bring it into the real world.
We looked at several use cases where the technology would make the biggest difference. In many ways, that’s the worst way to build a company, because it means starting with technology and looking for a problem, rather than starting by trying to solve a pain point.
We realised that accidents and disasters would be a great place for AI because every time there’s an accident, something gets damaged. The damage has to be visually assessed to proceed with the recovery, which is why we decided to build Tractable.
Matthew: What exactly is the AI doing better than humans?
Adrien: Any motor insurance customer who has made a claim knows how long and tedious the process can be. On average, a car accident claim takes more than 30 days to resolve, and the biggest bottleneck is visual expertise. What is the damage to the car? How much will it cost to repair?
Getting back to normal requires multiple appraisals of the asset, so we thought that if AI could learn this visual task, could we respond instantly? Could we use a smartphone to acquire photos of the damage and then use AI to make sense of the photos? If we could do that, the process would be faster and more efficient, which is what we’re doing now.
Matthew: Did it take a long time to build the technology before you could go out and sell it?
Adrien: The main limitation in AI today is the volume and granularity of data needed to train the system. When we first started, the number one priority was to get enough data to train the AI to understand the task.
The beauty of the insurance use case was that insurers had been documenting claims, including photos of damaged vehicles, for decades. There were millions of examples sitting unused on servers. We got access to data from our global customer base to train the AI so we could serve them better.
Matthew: You still had to find your first customer, so how did that work back in 2015?
Adrien: We’re in a great position now, with customers in 13 countries, but that first step is the hardest. New companies need to do things step-by-step. Start with a small training set that shows that the team can perform certain tasks and can grow and process more claims. There’s a little bit of showing and selling the vision involved and being open and transparent about the limitations and capacity of the system.
Try to find the visionary customer, the partner who wants to go on the journey. Not every insurance company wants to be a pioneer and those are not the first customers to have. In our case, we were fortunate to have Covea in France and Ageas in the UK championing our technology.
Matthew: How has Tractable managed to scale so quickly?
This kind of technology really scales globally. Companies are rarely able to go into so many markets so quickly, but AI and the cloud are a very interesting combination for scaling globally because the task is comparable.
A damaged car in Japan, or France, or Canada, requires the same kind of actions and the same kind of assessment. With this technology, a company can rapidly become a global leader.
Matthew: Japan has also become a major market for you and you’re working with Tokio Marine and MS&AD. Why has Japan become such a big focus?
Adrien: Japan has been a fascinating market for us for a couple of reasons. Number one, there is a big appetite for innovation. The appetite for adopting new technology and doing things differently was stronger there than in other markets.
Number two is timing. The market leaders there were trying to develop a similar solution but hadn’t managed to do it. When they saw what we could do, they realised they could deploy the solution without having to build it in-house.
That situation is the same in Western Europe and North America. Those areas have a higher frequency of claims but there aren’t enough people to deal with the volume. Technology helps support humans to work faster. If a typhoon or earthquake damages thousands of homes and cars at the same time, insurers need technology to be able to deal with the situation.
Matthew: The insurance industry still has a prevalence of legacy technology but there’s a particular challenge in Japan that most of us would have thought no longer exists. Can you explain?
Adrien: A big challenge we’ve had to overcome in Japan is information sent within the ecosystem. Repair companies and insurers are still using fax machines to send through estimates and information. It means we need to make sense of the fax where the information is not as clear, and we need to understand katakana (1) characters.
We needed to build additional modules and layers to make sense of the input before we could use the AI to perform a task. It’s not what people would expect, but 80% of the claims there are still being done through faxes.
Matthew: There are many third-party companies now offering AI applications, so what distinguishes Tractable from the rest of the market?
Adrien: Creating what we’ve built requires two things. It needs a lot of data and domain knowledge and it needs the right deep learning capabilities. Our researchers, machine learning talent and AI are the backbone of the software we’ve deployed.
The amount of data needed to train an AI system is in the hundreds of millions of data points, and a single insurer in one market just doesn’t have enough data. They also need to have the best researchers and it’s not easy to attract, manage and retain that kind of talent. The best researchers want to work for big tech and solve problems worldwide, so talent is limited and scarce.
That’s why it’s so difficult for insurers to develop systems in-house and why it makes much more sense to find the right partner to work with.
Matthew: One of the challenges with AI is what happens when the system doesn’t know the answer, or it can’t provide the answer with confidence. Is the AI smart enough to ask for help?
Adrien: Assessing damage to a car is a very subjective task. What the AI brings to the table is consistency and the ability to come up with a fair and objective assessment. That assessment needs a confidence score from the AI, which highlights if it wasn’t able to perform the task, or that it can’t give an accurate answer. That was one of the first things we developed at the beginning. The insurer can then default back to a manual process when it’s needed.
Matthew: How do you convince customers that what you’re giving them is better than what they can do today?
Adrien: A big part of every engagement is comparing what the AI is saying about a case with what a consensus of experts is saying. To systematically assess where humans and AI are agreeing, where there are disagreements, and where those disagreements come from.
That’s how to build trust and confidence and it’s a big part of all of our engagements. Today, our AI is used by the world’s biggest insurers over the three continents and there are tens of thousands of claims being processed every day.
Matthew: What information are insurers getting from Tractable?
Adrien: The main questions we answer are what is the damage and how much is it going to cost? Insurers receive a photo of the damaged car and the AI assesses the damage and provides an estimate. The assessment will include which parts have been damaged and the kind of labour needed to repair it.
With that information, an insurer can speed up its claims management and make decisions on the spot. It can tell a customer on the phone what repairs are needed and when they will start because it has the information straight away.
Matthew: You’re getting a lot of data now across 13 different countries. Do you see any different characteristics around the world for how people drive, or how accidents occur in different countries?
Adrien: The main value in what we do comes from focusing on the repair cost. 40% of the insurance premium is driven by the repair cost of the car. We’re uniquely positioned to give a sense of where repair cost inflation comes from, the different repair standards in each country, and how they can be improved.
The biggest benefit for our customers and the industry is to get a benchmark of best practices to see where the easiest savings are if car repairs are more efficient.
Matthew: What about videos versus photographs? Presumably, you can get more information from a video of the damage?
Adrien: We’re still mostly relying on photos from drivers, body shops and experts to produce an analysis. Video is going to help with fraud because there’s more information that can be analysed than in a few photos.
We are currently developing and piloting a solution that combines video and AI to evaluate damage more accurately by filming a simple lap of a car. That’s where the industry is going.
Matthew: Does that mean you can help guide the customer on where to point their phone, rather than them having to figure it out?
Adrien: The old process would be to take a video, send it and just wait. The video would then be sent to the cloud and then all the analysis would happen in the cloud.
What’s exciting now is the concept of embedding everything in the phone so the analysis can be done locally. We can then guide the customer in real-time and flag when the information is not clear. By guiding the customer, we will acquire much higher quality information, which means the analysis is of a higher quality as well.
Matthew: What are the barriers to adoption like if somebody wants to switch from human assessments to the Tractable solution?
Adrien: We went from zero to 25 customers in less than three years and the main reason for that is the integration is very simple. Insurers don’t have to redesign their claims system.
Our solution is just a layer on top of the claims system and the only thing we need is to acquire the photo at the right point in time. Once we have that, we can send results back to the decision-makers to take action. The integration process is very, very light.
Matthew: Does that allow you to talk directly with business users, rather than having to go through an innovation team?
Adrien: We do both. We want to improve and accelerate or streamline existing processes, and for that we work directly with the claims team or the chief claims officer.
We also have a big focus on the digital experience of the policyholder and that’s led more by innovation. We need systems that put technology directly into the hands of the driver, and for that we want to work with both claims and innovation teams.
Matthew: You’ve raised $55 million in external funding and are now looking at using some of the techniques you’ve developed for motor for property. Can you give us more details about that?
Adrien: A big part of Tractable’s mission and vision is to use AI whenever accidents and disasters happen. We’ve been very focused on motor as the first step, but this year we’re expanding into different types of damage and into property.
A typhoon or an earthquake damages thousands of properties at the same time and the speed and scalability of AI can make a difference. We have a property customer signed up and we’re piloting the solution as we speak.
Residential property will be first, but there won’t be a big difference in what we do for commercial. The value drivers might be different but the goal, understanding the extent of the damage quickly using visual information, is comparable.
Matthew: Speed of response is particularly important for property. Once someone has damaged their car it probably won’t get worse, but damage to the roof of a property opens the possibility of further damage inside.
Adrien: The customer wants an easy and quick response in both cases. Houses will deteriorate so insurers have to act fast, but mobility is also crucial for people so they can go back to work. For example, Admiral is using our solution in Spain to settle claims on the same day so customers can get back to their lives faster. That makes a big difference.
Matthew: We had Alex on stage at one of our events in 2015 when this was just an idea, and we’re delighted that Tractable is now a corporate member of InsTech London. Why did you decide to join us?
Adrien: It’s very clear that InsTech London brings together the best of the insurance and technology sectors. Being a member is going to be very helpful in facilitating both conversations and collaboration.
We’ve not been as involved in the UK sector because we initially focused on international expansion in North America in particular. We’re now accelerating and investing a lot in the UK market, so it’s the right time to get involved in what InsTech London is doing.
(1) Katakana refers to the characters used in Japanese script that are literal translations of foreign words.