How insurers use AI in auditing to fix property data errors

• Understanding property data mismatches helps catch issues
• AI-powered synthetic data can respond to risks and perils
• Processes used to produce property information must be auditable

Where do you live? Your address provides one answer; your map coordinates provide another. And if they don't point to the same location, insurers must find a way to determine which source is accurate, insurtech and property-mapping experts said.

Understanding patterns in property data means figuring out if data patterns are real, or reflect problems with the data itself, said Amy Rose, chief technology officer at Overture Maps Foundation, an interoperable map data infrastructure organization.

"It's being able to say there's a systemic issue with coordinates or address points being flipped. The address says it's 123 Main Street, but where it is on the map, if you plot it by coordinates, is wrong," Rose said. 

As long as these issues and the data are communicated, the end user — human or machine — can perform the proper validations, Rose said. "The real power is in auditability." 

Insurance companies are applying computer vision, synthetic data, data sourcing rules and AI to determine accuracy and consistency of property data. 

"Insurance companies use a variety of techniques to reconcile differing information about property risks, terrain, data, and other information like historic claims," said Franklin Manchester, principal global insurance advisor at SAS. "The most frequent examples are computer vision and synthetic data generation to smooth over some of those differences."

Synthetic data, itself an AI capability, lets insurers augment data sets to respond to trends such as emerging risks and secondary peril activity that worsens climate disasters, according to Manchester. 

Certain data-sourcing rules can help address discrepancies, according to Ben Brammer, head of product at Agentech. Examples of these rules include "source of truth," addressing which data source supercedes others if there is a conflict, or a recency rule in which the most recent data supercedes others. 

"The key here is to have this documented or codified so the decisions are consistent and transparent," Brammer said.

AI can also address discrepancies by normalizing and aligning unstructured data, or making rule-based decisions and retrieving data confirmations, Brammer explained. AI can be built to normalize unstructured data by interpreting documents, data and reports. 

"This can be as minimal as understanding that a pipe jack and a pipe boot typically refer to the same thing, or can be as obtuse as inferring that a two-story property could be confirmed by the presence of a cornice return," he said.

To make such decisions with AI, rules can be set to capture data through APIs or web searches. "If there is a scenario where a clear decision is unable to be made and an alternate data source is available, AI can be invoked to retrieve the 'tie-breaker' data," Brammer said. 

Non-spatial data is easier to audit, particularly when data is combined, according to Rose. Even a seemingly straightforward intersection of spatial data can significantly transform it, she added. As an example, Rose described deriving building heights by combining pieces of data from different types of sources. 

"The derived height is probably not at the actual spatial resolution of the building itself; it's very likely a coarser, generalized value. If they just publish heights without all the information about how they processed it, then there's nothing to tell the user that the height value may not be appropriate when exact heights are needed." she said. "That auditability trail is one of the most important parts, particularly when you're dealing with spatial data."