Building the big picture: how we train AI to assess property damage

Homes are trickier for AI to assess than cars. With autos, there is a lot of standardization of models and parts. For instance, most cars have two or four doors, but properties don’t have a standard number. Also there is far more variation within each object that makes up a property. There are dozens of different door and window types, and they may be made of wood, metal or different types of plastic. So while homes seem a logical next step, they’re a considerable challenge for AI.

Turning complex tasks into simple steps

That said, the method we use to make homes digestible for the AI is the same method we use for recognising auto damage. You break a complex problem down into a set of simpler steps which are easier to deal with. This simplification is done by consulting with insurance experts who are experienced in assessing properties and the specifics of property damage.

The result is that this complex picture becomes a series of narrow, circumscribed tasks that current AI technology can deal with. On top of these AI tasks are sets of rules and traditional software which act as the glue that holds it together. 

Training AI to recognise property damage

Imagine a fence damaged by a storm. You take pictures of the fence on your smartphone and upload them using an app. At Tractable, the entire fence is visually broken down into component parts such as panels, poles and concrete block foundations. 

These component parts are then assessed individually by the AI. It looks at each part to determine whether the damage is due to the storm or normal wear and tear. It asks whether each part can be repaired or needs to be replaced. 

Assessing images and extracting information is exactly the kind of task that AI excels at. Tractable’s system also looks up the cost of repair or replacement based on the property location – which calculates the sum transferred to the policyholder’s account.

AI that continuously learns and adapts

Right now, the AI deals with relatively modest tasks, like broken windows. But in future it should be able to deal with more comprehensive damage, including homes flattened by a typhoon. Applied AI is trained on historic cases that have been assessed by humans – but it learns continuously and, where necessary, receives input, oversight and correction from human experts. Over time, the AI gets better at the narrow tasks it’s been assigned and will also expand its capabilities within those tasks.

Our launch in property claims shows how Applied AI can be quickly adapted to new domains and start learning new visual appraisal tasks, by systematically segmenting and codifying expert knowledge. This ability, based on Tractable’s Applied AI technology, allows us to also share knowledge between new domains and enter entirely new areas of expertise in a fraction of the time. 

For people whose homes have been damaged, AI is helping to create more efficient and accurate insurance services with the potential to improve those services each time a claim is settled. In a world where property damage due to natural disasters is increasing, this solution is helping many people to recover their homes and their lives faster.