[spencerchubb]

I work in insurance. Most states have a Department of Insurance who approves price changes. Insurance prices are calculated using simple features, such as "If car is from year 2016, then multiply price by x"

However, the numbers are not publicized because they don't want competitors to have that info.

Another fun fact, a lot of people wonder, "Why doesn't an AI startup just disrupt the insurance industry?" It's because the Departments of Insurance have to understand the price formulas. Neural networks are infamously hard to interpret, so we would have to reform regulations before we can use neural nets.

[robocat]

> Why doesn't an AI startup just disrupt the insurance industry?

Also insurance requires that the insurer discriminate a good driver from a poor driver. However they should not discriminate against a protected status.

Good luck setting up a system that can only discriminate using some signals and not others.

It is a good mental exercise trying to think of ways to set up a startup that can discriminate without being obvious about it.

[lotsofpulp]

> Good luck setting up a system that can only discriminate using some signals and not others.

This is the current system.

[thwarted]

And an AI based system wouldn't correct that, it would only make it even more opaque.

[jandrese]

The requirement isn't to avoid discrimination, it is to not get caught. That somewhat opaque "AI" layer is great for that.

[CydeWeys]

And if it's actually systematically charging e.g. a specific minority more than other people, it will get caught in a hurry and end up being hugely costly for the company.

This kind of stuff is easy to catch. A single person typing some different parameters into an insurance quote webpage can catch it.

[jandrese]

It's one thing for people to know what you are doing, it's another to prove it in a court of law, or even to get law enforcement interested in the first place.

[BlueTemplar]

Until you are forced to explain exactly how your helper algorithm came to that conclusion, explaining the logic and calculations step by step in easy to understand language... (With the "I can't" not being an acceptable answer.)

[tomrod]

Nah, there are some nice metrics to capture disparate impact between categories. Check out Microsoft's fairlearn library some time.

[thwarted]

It's because the Departments of Insurance have to understand the price formulas. Neural networks are infamously hard to interpret, so we would have to reform regulations before we can use neural nets.

I'd like to know that the methodology is understandable and defensable and correctable, and having "an AI startup" disrupt with neural networks isn't going to do that. At best it will just be another excuse used to justify the state of the industry, "the computer said it so it must be right". Reforming regulations so it can be made even less transparent is not the way to go.

[nerdponx]

What's interesting though is that while pricing is strictly regulated, underwriting is significantly less regulated, at least in P&C commercial insurance. Insurance companies have been exploring the use of ML and AI for that task since at least 2017, when I got a job doing precisely that.

Also, things like machine learning for image recognition in claim photos, satellite data, etc. has also been in use for at least the same amount of time.

I believe the Lemonade renters insurance product also does some kind of "AI" claims processing. I wouldn't know what that looks like, my focus when I was in insurance was solely in underwriting.

[phonon]

> the numbers are not publicized because they don't want competitors to have that info.

What do you mean? Pricing is all publically available (sometimes not all the details that feed into the model that create the pricing tables.) You can search SERFF by carrier, by line, by State, and read the actuarial filings.

Most of the time, pricing is refined by looking at prior year(s) losses, and adjusting. You go to the State, explain how much you've been losing, and they review. All that correspondence is public as well.

[wolverine876]

I'm not sure what that means: How does that explain the widespread and large price increases, and the GP's experience?

[tomrod]

> Why doesn't an AI startup just disrupt the insurance industry?

More than this, it's a fundamental misunderstanding of what AI is and what it can do to ask this sort of question.

Best thing I've seen come on the market was Root, based in Columbus and founded by a former Finance director from Nationwide insurance. It used your cell phone to send telemetry signal to classify your driving behavior.