German company Sachcontrol hired a former physicist as its head of data science to create a machine learning model that verifies property repair invoices using computer vision technology
Insurance claims for property damage can be a prime time for opportunists to inflate the cost of repairs for numerous reasons. Dr Andrey Lutich tells Peter Littlejohns how his team of data scientists are using computer vision technology to make sure insurers aren’t ripped off.
When properties spring a leak, catch fire, or suffer any other mishap, it might be tempting to use the opportunity to not only repair, but improve the damaged area and beyond.
While this doesn’t strictly fall under the category of insurance fraud — the cumulative losses can amount to a hefty bill for the provider that pays out.
Dr Andrey Lutich began his career as a physicist in the field of optics, working with several high-profile research organisations on projects that incorporated AI.
Now, he leads a team of data scientists at insurtech company Sachcontrol to further develop its computer vision technology — a sub-genre of AI that enables computers to understand images at near-human level — that verifies the extent and cost of property repairs.
He says: “In some cases, it’s very hard for insurers to decide that the invoice given in a claim is appropriate to the damage that occurred.
“In cases where they can’t do this, they can use Sachcontrol on the back-end to help them make a decision.
“Sometimes we call it fraud detection, but it’s not like someone went and faked an invoice.
Dr Lutich explains that it’s natural for claimants to exaggerate the scale of repairs due to a general perception that insurers will pay out as little as possible, but some will also use the opportunity to replace old materials and increase the value of their properties.
Sachcontrol is active in Austria, Germany and Switzerland, where 90 insurers use its technology to verify whether claimants are trying to achieve inflated payouts for either of the reasons described above.
Computer vision technology needed a ‘catalogue’ of property repairs
Founded in 2001, Sachcontrol used to send out civil engineers with rich experience in building repairs to assess the damage reported in a claim to help insurers avoid paying out for unnecessary repairs.
For the past six years, the company has kept a digital record of these interactions for reporting purposes, rich with damage descriptions and images, all stored in a “catalogue-like” format with pricing information.
This meant when Dr Lutich came on board to develop an AI-driven system to automate the validation process, he already had structured data on property repairs to train the machine learning model needed to make it a reality.
How Sachcontrol verifies the cost of property repairs
The use of computer vision follows a similar trajectory wherever it’s used in the world of insurtech, with one of the most common use cases found in the motor cover sector.
In this case, pictures are used to discern parts of a car and give information about how much each will cost to repair or replace, giving an estimated claim cost.
Sachcontrol’s parent company Solera — which acquired the insurtech back in 2014 — is already active in this space using its Audatex platform.
Sachcontrol’s SachEye system plays a similar role for parts of properties, but, according to Dr Lutich, this is much more difficult because cars are relatively standardised with a limited number of part manufacturers — the opposite of properties.
Dr Lutich explains that the system uses deep learning and computer vision to break images up and identify all of the features within them, as well as whether they’re damaged or not.
“SachEye takes an image as the input, aggregates all of the data from our catalogue and generates a calculation of the repair price based on an average from our invoices.”
The machine learning model can do this because it has been trained using thousands of pictures, along with the required descriptions to identify what an image contains, what materials it consists of and what constitutes damage.
The algorithm at work also includes the difference in average repair prices per region, which Dr Lutich says can be quite stark in Germany especially.
Sachcontrol makes its API available to insurers so they can plug it into their own apps and provide cost estimates for repairs as an added benefit to customers. German insurance giant Uniqa has taken advantage of this feature and integrated SachEye within its smartphone app.
Sacheye limitations and future plans
Any kind of new technology put to work in the insurance space is bound to have limitations, and like every AI-led computer vision model, SachEye becomes more accurate as it receives new data.
According to Dr Lutich, SachEye has an average prediction error rate of 25%, but for simple claims like hail damage on a set of blinds or a window, the level of confidence predicted by the model is much higher.
Alongside tuning the model with more data, he hopes to export Sachcontrol’s technology further abroad, working on a current deployment with an insurer in Australia.
Dr Lutich and his team are in the process of integrating data on Australian houses for this reason, as although it can recognise simple features like a window, it’s not trained on the different materials the country uses to build dwellings.
“The Australian way of building houses is different from the way they’re built in Germany; Australians use a lot of wood which isn’t an option in Germany,” he adds.
Another route to expand Sachcontrol’s reach is through its work with Appian — the software company it uses to build its applications — which it hopes will result in a technology partnership between the two that gives Appian customers access to SachEye.