That's just not true. Source: my sister is a commercial underwriter at a large US insurance company. And while they do have software that gives a suggestion for a premium, it does not know enough to be accurate, so she has to always adjust it.
This is definitely an industry where more automation could easily be done, but the big insurers are a conservative, risk adverse group.
I don't dispute your retort because the post you are replying to was worded to sound very absolute, but...
AI doesn't have to replace 100% of all humans to have a huge impact on unemployment. If you introduce a system allowing 4 people to do the job 5 used to do you're setting the stage for 20% unemployment, which is a huge deal at scale, and also the 4-to-5 ratio is very conservative for a lot of modern automation projects.
Are there many fields where AI/robots will be doing 100% of the work in the near future? No... next to none, I'd think... but there are LOTS of fields where they will be doing a huge amount of the work while being supervised by a relatively skeleton crew of humans sanity checking their work.
This is true, in the UK at least. Insurance is the most trailing-edge industry I've worked in... incumbent firms are full of servers under desks, proprietary bloatware with no APIs, spreadsheets-as-databases. In my experience tech could make large swathes of the insurance sector redundant without even needing to resort to AI.
Insurance doesn't suffer the same degree of competition as other parts of the economy... it has a triple-walled garden of hefty regulation, significant capital requirements, and the chicken & egg problem that you already need to have relationships and experience in the insurance sector to do business there... or spend time and money buying them in. Even the banks white-label their insurance products from insurers.
Not to marginalize your sister's job, but couldn't that essentially be training for an "AI system" (NN or otherwise), and over time it will reach similar conclusions with increasing accuracy? Maybe they're doing that at her work already behind the scenes.
An AI system could probably get close, but the act of capturing all the variables might cost the insurance companies more then the current system, which involves a lot of 'gut feelings' about what is important or not for a policy.
There is also licensing involved, in my sister's case she had to earn a CPCU before she could do here job on her own.
>which involves a lot of 'gut feelings' about what is important or not for a policy.
Edit: I see you answered this same question below. Whoops.
Can you give some examples?
I thought it's pretty well understood that the "gut feelings" of experts have been and will continue to be outperformed by algorithms for these sorts of tasks. My imagination is failing trying to come up with something data-based that an agent would see and a computer couldn't.
Thanks for the info, and yeah licensing was something I hadn't even considered but could be a big liability for a company relying on machines (side note: wonder how the self driving cars handle this?)
My naive view from being a customer of various insurance companies is that there's a series of lookup tables and charts that does things like "male age 18-25, no history of smoking, here's your rate", with some room for variance based on other factors (some linear or log scale for family history or blood test results, for example). On the surface, those all seem like logical computer functions to do, and as an extension things that would work well training a NN off of after starting with the basic lookup tables. Can you go into more detail on what I'm overestimating?
Agreed that 1% could cost or make the most... now do you trust a human or computer to make the final decision? My guess: Humans will end up costing more.
Market stuff, like how the companies entire book is doing (what losses they've had so far that year), regional balancing, past payouts to the insured, raising or lowering the overall risk of the book, and just plain sales (sometimes you have to take on more risk then you'd like to meet a sales quota, but you can balance that by making other lower risk clients pay more).
It's also complicated because it's hard to model buildings correctly. My sister likes to tell a story about a total loss she had for a standard masonry building with sprinklers. From her point of view, this is the perfect type of building. They're hard to catch on fire, and if there is a fire, the building puts it out. At worse, your damage is limited to the one room or section where the fire was, since the internal masonry walls keep it from spreading. This building burnt down because the solar panels on the roof caught fire, which spread across all the panel over the entire roof. The sprinklers never got a chance to go off because the roof collapsed. It's certainly possible to model this one case, but the problem is there are a million one off cases like this, and we don't know about them until there's a loss. Right now, human intuition from the underwriters and inspectors is what they use to cover try to cover the gap.
I'm not in insurance but I assume that it is the nature of pooled risk that there will be a million one-off cases (or N cases where N is the number of members in the pool). Given that, is it possible or even desirable to make changes based on what happened to this specific insured business? Would the cost of that change, be it increased premiums or an outright refusal to insure, outweigh the cost of a smaller pool? I guess that's something for the AI to decide.
Why would it matter on the how the entire book is doing? Why can't the AI system take all of this into account assuming it has access to said data? Further, why should the AI system be used to run the entire company like you suggest. That is, it doesn't need to take into account every aspect of the business to replace a large portion of the workforce.
> This building burnt down because the solar panels on the roof caught fire, which spread across all the panel over the entire roof. The sprinklers never got a chance to go off because the roof collapsed.
Type 3 construction that burned to the ground? It doesn't take a human to realize or intuit that even type 1 buildings burn eventually (ask a firefighter!). If there is a way to truly model every possible variable when insuring a building, I will put money on humans doing a worse job than expert AI systems. You're asking the AI to not only predict but _know_ the future, and not asking that of the human... seems silly.
A friend told me about s Russia bank that tried to use true AI system. After initial precise results on the rate of defaults after one year they have to scrap it and went back to using logical regression with heavy human reviews. The biggest problem was that when AI errored, the errors were huge with no way to know what caused them.
Having seen how these expert systems are created, replacement is way overdue. For example a team of 10 "senior mortgage professional" spend weeks of figuring out what difference it makes if the co-applicant's age is 60 or 62. They argue, and arrive to a conclusion that it increases the risk by 0.01 (whatever).
Event the simplest logistic regression training + evaluation will provide value to most insurance, mortgage or other money-related decision/"expert" systems.
This is a very ambitious attempt to replace complex and messy human powered business workflows with an AI system. Most current business uses of "AI" tackle tightly defined problems, like assigning invoices to a certain group, filtering out spam etc. Even if the implementation is successful, they will still have to retain quite a few people to deal with exceptions generated by the automated process.
