Amongst a wave of buzzy deep-learning startups in the early 2010s, Tractable has managed to grow steadily by addressing a very specific problem with the technology: helping insurance companies to automatically assess vehicle damage.

Founded by Alexandre Dalyac and Razvan Ranca in 2014 after graduating from the company builder programme at Entrepreneur First (EF), Adrien Cohen joined the company soon after as chief business officer, and set out to discover where to apply its computer vision expertise.

© Tractable
© Tractable

"If you take a step back, at the time there was a wave of deep learning startups," Cohen told Techworld last week, speaking from a conference room at Tractable's London office, which is an entire floor of a WeWork near Old Street station in east London. "There was a big number of companies with great technology, but no problem to solve. Most of these companies, they ended up being acquired."

One example is Magic Pony, the image recognition startup which also graduated from the 2014 EF cohort and was acquired by Twitter in 2015 for $150 million. Then there is Bloomsbury AI, another EF alumni which was purchased by Facebook in 2018, and MetaMind, the AI specialist acquired by Salesforce in 2014 and whose founder Richard Socher is now the chief scientist at the Californian cloud firm.

Read next: UK insurtech startups to watch

Tractable, on the other hand, had a narrower goal. It just didn't know what it was yet.

"We looked at pipeline inspection, we looked at medical imagery – which has been in a big application of deep learning – we looked at oil and gas, seismic interpretations; anything that was visual," Cohen explained. "We eventually ended up looking at car accidents, because we realised it was the perfect narrow task for AI."

That doesn't mean it was simple, however: "Predicting the way you're going to repair a car is extremely complex. You have external damage, internal damage, you have so many makes and models, the lighting, the car park, it's extremely complex.

"It had never been done. So by definition, clearly, the machine learning team was like: 'well, that looks like a very hard problem to solve'. But we started working on it, and we managed to solve it."

The stack

What the firm actually built over the ensuing four years was a sophisticated image recognition application that can assess the damage level to a vehicle just from a set of smartphone photographs.

This involved developing a set of neural network models on a common set of cloud tools and services, eventually running them on cloud GPU servers, predominantly with the vendor Amazon Web Services (AWS).

"You need the data, you need the domain knowledge to know how to use this data and train your system and then you need a set of machine learning techniques that are specific to the complexity of the challenge we're facing," Cohen said.

This involved building teams of machine learning experts alongside industry-insiders like motor engineers, body shop assessors and insurance appraisers to train these models to the point where an insurer would be confident in the output.

"Those teams for four years have been working hand-in-hand and it's a feedback loop," Cohen added.

By building in a cloud-first way, Tractable is able to easily integrate its solution into existing digital workflows, whether that's between the insurer and the body shops, or, increasingly, between the policy holder themselves and the insurer.

"What we do is we accelerate your process," Cohen explained. "You can use this technology, you can insert this AI, as an insurance company, whether between yourself and the policyholder ... or you can insert it between you and the body shops; and we're doing both."

For example, US insurer AllState introduced a QuickFoto Claim feature to its mobile app in 2013, allowing customers to submit their own photos of the damage after an accident. This sort of reduction of friction for policy holders is an opportunity ripe for Tractable to insert itself in the future, and creates valuable efficiencies for the insurers and their claims handlers.

"It's pushed to the cloud, in the cloud it is processed by the AI in real time, we send back the result via web UI that either you on your phone or the claim handler on his desk, sees," said Cohen, "so they see the decision from the information from the AI and can make a decision on that basis."

Customer adoption

The important factor here is the confidence level of the AI, and the threshold at which insurers become comfortable enough in the AI to push a claim forward without a human intervening.

The system Tractable has built – called the Tractable Image Classifier – will take a set of images and combine them into a single output, from which it will assign a confidence level from 0-100 as to how badly damaged a certain part of the car is and a recommended action, like replace, repair or intact.

"In every decision that we make, we know whether the AI is confident enough to make a decision or not," Cohen said. "That can be driven by not having enough pictures of the damage or the quality of the picture is too poor to make a decision, so you either send more information or it's not confident enough."

Each customer will have a different threshold when it comes to confidence levels but say, for example, that the score is more than 90 percent, a customer could decide to push all of these cases straight to the body shop, rather than employing a human assessor as a burdensome middle step in the process.

"Depending on the confidence score, it would either be confident enough to automate, or if you're not confident enough, then I would want someone to have a look. And if it's really complex, then maybe I want to send someone," Cohen said, describing the system as a 'triage engine' for insurers.

As well as the clear efficiency gains this software brings insurers, Cohen also talks about the value of greater customer satisfaction if claims are processed faster.

Tractable today works with leading insurers in nine countries, according to Cohen, including Ageas here in the UK and Covea in France, but has grand ambitions to "onboard the top 40 players in the world" by the end of 2020. The firm has already raised around $35 million in venture funding from the likes of Insight Venture Partners, and has 100 people across three locations to help it get there, so what is holding prospective clients back?

"They all want to see it and understand it and in one way or another they want to start deploying it," Cohen said. "Now, the main question they have, is: how does this work? Is it accurate?"

To help answer these questions Tractable goes in to engage in what it calls a 'calibration', where it compares the performance of its AI system with the insurer's own team.

"It's a slow process," he admitted, "where you're going to process a number of cases, and review with them the agreement between humans and the agreement between machines and the AI. You want to show them, gradually, that actually it is able to operate at the level of performance that they expect."

The roadmap

Aside from its ambitious client goals, Tractable also has its eye on other damage assessment use cases.

"We're training it today... we've already starting to look into other types of accidents and disasters, because it's the same customer, this insurance company that we're working with, they have the same data, they have the same objective," Cohen said.

The next use case specifically is escape of water cases.

"We're already doing some research, we already have gathered some data sets, where we're refining our technology to train an AI to make sense of the damage," Cohen said. "Essentially, it's still a visual task where a large percentage of the cases currently are done by people looking at it from a computer, there's no magic, when someone can make the decision by looking at it, then a computer can."