How AI can revolutionize small commercial underwriting

Insurers hear a lot about AI and machine learning these days. Amidst the hype, it can be hard to parse what technologies are field-ready and what’s aspirational.

For small commercial underwriters, one of the most exciting and powerful real-world use cases for AI/ML is leveraging it to analyze and extract insights from unstructured data.

As the name suggests, unstructured data is information that's not neatly organized into pre-defined fields. This could be photos, online reviews of businesses, text documents, web pages, videos and audio files. Unstructured data may contain immensely valuable insights, but it’s historically been either inaccessible to underwriters or very difficult and time-consuming to retrieve and analyze. 

How AI changes the game
Computer vision algorithms derived from machine learning can identify objects in photos and can scan images posted to social media to identify potential risk exposures. 

By identifying those exposures automatically during the underwriting process, insurers can gain a better understanding of the applicant’s operations without having to ask intrusive questions or rely on information provided on an application. In fact, when this data is provided with a rich suite of traditional underwriting data, they may not need an application at all. 

Look at restaurants, a very common small commercial risk. Patrons may post multiple images in a social media review that reveal the presence of flaming shots, bouncers or a DJ booth—all relevant to an underwriter. Computer vision algorithms trained by insurance experts to spot the relevant objects can help identify more granular information than what may be relayed in general industry classifications (NAICS/SIC). This can eliminate the often error-prone process of manually classifying risk exposures based on information received on an application—and, again, may help eliminate the need for the application in the first place.

Clean and relevant
Insurers need meaningful data, and AI/ML models can turn complicated, structured and unstructured data into actionable insights. 

But not any data will do. 

To build useful models, you need current, accurate and well-maintained information that’s vetted not just by data scientists, but by insurance experts. After all, if you build a model with poor quality inputs, you’re likely to get poor quality outputs. Data quality is a multi-faceted concept, but for the purposes of unstructured data analysis, one important metric is the sourcing. 

Some unstructured data can be collected by simply using software to copy-and-paste web content en-masse and then analyzing the results. While this process casts a wide net, it could be scooping up loads of irrelevant information—like images and text that are used in advertisements that accompany the unstructured data you’re actually interested in. It’s also contingent on the vagaries of what information these websites make available for copying. What’s here today, could be gone tomorrow, leaving underwriters in the lurch.

A more reliable method is to secure access to cleaner, relevant data for AI/ML model building through partnerships with social media websites. This way, image analytics models are built on more reliable data that can deliver relevant information to insurers.

A piece of the puzzle
As valuable as AI/ML-derived data is, it’s important to recognize that it’s a supplement to, not a replacement for, other sources of predictive underwriting intelligence.

The universe of insight embedded in unstructured data provides a view into risk, but you can widen your lens by pairing it with traditional sources of risk insight, such as business firmographics, loss histories, licensing, violations and credit histories, plus property and location data. In tandem, these sources can provide a holistic view of risk.

Using AI data to accelerate your operations
Today, small commercial insurers can leverage AI/ML data and analytics as part of a holistic solution to transition from manual underwriting workflows that lean on lengthy applications and web research to one where quotes are issued and most policies are bound automatically and (nearly) instantly.

The journey has three broad phases.

Step 1: Prefill: Leveraging an array of data sources, including unstructured data sourced from computer vision algorithms, insurers can prefill application data on small commercial risks using just a business name and address. Human underwriters can then review this information against underwriting guidelines without having to chase down data through web searches or phone calls. 

Step 2: Selective automation: Based on risk appetite, certain industry classes can be identified for automated underwriting. In this environment, application data is prefilled and then automatically analyzed against insurer underwriting guidelines to determine acceptance or whether additional information is required. 

Step 3: Full-blown automation: As insurers learn from step two, it’s a short leap to step three, which is to fold additional businesses into the automated workflow. Even in a fully automated environment, there are some risk exposures that may trigger manual reviews of submissions. But by leveraging the efficiency gains delivered by application prefill and the automated underwriting of select industry classes, insurers can set themselves up to drive automation across a much wider array of risks than they ever thought possible.

Alan Turing, one of the pioneers of computer science, famously predicted that by the end of the 20th century, we’d “be able to speak of machines thinking without expecting to be contradicted.” 

Today’s 21st-century machines may not be “thinking” by human standards, but thanks to AI, they are learning. And no one is likely to contradict you if you assert that these machines will play an even larger role in insurance underwriting in the years to come.