The Center for Data Innovation spoke with Adrien Cohen, president and co-founder of Tractable, a visual recognition platform based in London that applies AI to accident and disaster recovery. Cohen discussed how using Tractable in the event of an accident can significantly accelerate processes such as damage claims.
Eline Chivot: How did you get started using AI for accident and disaster recovery?
Adrien Cohen: Tractable originally started as a research project for developing AI to perform visual checks on photos of construction setups, such as welds. Construction workers could be asked to take photos of their setup, and have an AI analyze it in real time to validate that a weld could be safely performed. The reason for this: What AI can do, which humans can’t, is to run thousands of visual photo checks across different events, all in real time, so that bad setups can be detected and prevented, without slowing down workers.
After working for a short time with welds, we soon found a better use case: Car damage. With accidents and disasters, cars get damaged most often, and parts are easy to replace. In addition, the £385 billion (€420 billion) global ecosystem for auto claims—involving repairers, part sellers, paint sellers, towing companies, auto recyclers, and car hire companies—contains many inefficiencies that AI can help with. By analyzing claims and observing outcomes, we can help insurers focus on the best players in the sector, and help them win more business.
It’s the perfect task for AI to help with: Heavy on image classification, where AI can surpass human performance.
Chivot: How does your technology use deep learning and image classification to automate visual damage appraisal?
Cohen: Assessing damage is a visual task. With the right approach, this can be carried out by a machine at the same level as a human expert—or even better.
Our solution uses deep learning for computer vision, and patented machine learning techniques. Our AI is trained on many millions of photos from opt-in customers and partners, and the algorithms learn from experience by looking at many different examples.
Once photos of damage are uploaded, the tool is therefore able to recognize what parts have been damaged, and assess how badly they have been affected. It can then recommend decisions, based on its confidence level. The system’s confidence level will depend on part visibility, photo conditions, and the extent of damage severity. This means the AI can also interact with the person taking photos to request specific additional photos on the spot.
Chivot: What are the challenges your technology can address? How is Tractable changing the claims process, the work of engineers and assessment experts, and outcomes for people?
Cohen: Tractable enables our customers—some of the largest insurance companies in the world—to capture photos at the point of accident, and to receive an estimate of the damage very quickly. Our belief is that when accidents and disasters hit, the response could be 10 times faster, thanks to AI. We are already delivering this as a solution to the automotive insurance industry, enabling it to process many more claims, more quickly.
For example, instead of it taking days or even weeks to assess a damage claim after a car accident, for insurers using Tractable it now takes minutes. Similarly, it is inefficient for body shops to receive vehicles that are total losses. AI total loss appraisals now take seconds, ensuring the right vehicles make it to the salvage yards instead of the shops.
Overall, our technology speeds up claim settlement, improves accuracy, reduces turnaround time, and delivers a revamped customer claims experience for the industry—improving the entire process for the end-user.
Chivot: An image of a damaged car does not always have all the information necessary for the claims process or the highest level of accuracy. Could AI be used to solve this as well?
Cohen: There are of course some claims where damage can’t be assessed visually, and this is often where a human expert is best employed—rather than on the many claims that are simple to assess and process. However, for most cases, our AI is usually as accurate as a human expert or better, and it’s continuously improving as it learns with more images and applications.
It’s like an auto repairer who has repaired over one million collisions, and consistently repairs to standard each time. We also do have examples of incidents where, from a visual assessment, a human couldn’t be able to tell the extent of the damage involved, but our AI can.
Chivot: How do you see computer vision disrupting other sectors than the insurance industry in the future? What may be some of the obstacles to its application and use by companies and users?
Cohen: There are two main challenges for applying computer vision successfully. One is finding a use case that’s commercially justifiable. It’s expensive to develop these solutions, not least as you need a huge number of images to train the AI on. So anyone looking to apply computer vision needs to be sure that there’s enough reward for doing so. Even the world’s biggest tech companies are struggling with this.
The second is accuracy. Companies regularly demonstrate that they can get their AI models to recognise about 95 percent of images in a given set. But that’s still too many errors to be truly useful, or be a real improvement on human performance—you need to get up to 99.8 or 99.9 percent to really make a difference. We’re at that point because we’re applying computer vision to a very specific use case with a very large data set. The challenge for others will be to find something similar, with enough data to effectively train a model to the required standard.