Saving 80% in 90 Seconds? | Hacker News Reader

Saving 80% in 90 Seconds?

blog.lemonade.com

58 pointsRadim56 comments

Why is the title here mention AI if the post itself has absolutely no mention of it? This is basically blatant advertisement for lemonade. Geico, Metlife, and others offer very competitive rates for renters insurance (<$80/year) in major metro areas and have done so for several years

It's actually there in the penultimate paragraph: "So how can Lemonade be 80% cheaper? You guessed it: by building an insurance company, from the ground up, powered by A.I. and behavioral economics."

I have similar questions after seeing the example monthly rates they give for the competition. I've never in my life paid that much for renter's insurance, it's always been around ~100/year. Progressive and Geico (reselling Travelers), mainly.

So I just kinda assume it doesn't apply to me, or are best-case edge scenarios where they judge risk of certain customers dramatically differently. But it certainly can't be "legacy insurers have enormous overheads, making a $60 product ($5/month) impossible" as a unique advantage, since I pay $75 a year today.

(And it does cover property damage liability, not simply personal property, something they seem to indicate their own does not: "Renters insurance covers personal property, rather than real-estate." Maybe they mean renters policies usually carry less property coverage, like $100K instead of full value of the property, I don't know, but it's just another weird thing in a weird-reading sales pitch that doesn't seem to apply to the world I've been living in.)

Is your policy a standalone policy or bundled with another policy, where the marginal overhead of the renters policy is lower?

Also, the $100k liability coverage you're referring to is more likely than not personal liability coverage, not property damage liability. The intent of that is to cover any claims resulting from e.g. one of your guests getting hurt on the property. Not to cover structural claims themselves. Most of the rentals I've lived in require that coverage, as it makes their premiums cheaper since their liability coverage acts as an umbrella policy that only triggers after that $100k is exhausted from your policy. But any structural claims that don't include personal property or liability issues would still go straight to their policy, not yours.

You are not covered for property damage with a renter's insurance policy. The $100,000 coverage is for liability (e.g. a guest slips and falls and she's you.)

You are also covered by a smaller amount (e.g. $10,000) for your personal property.

"and she's you" -> and she sues you

I was a sucker who converted.

In my case, as a homeowner, they were nearly 20% more expensive than my current carrier homeowners and umbrella liability. So either their underwriting model is crap, their data sucks (I literaly live next door to a paid firehouse and got 3* for fire protection).

They also don't ask you detailed questions unless you explicitly do so -- in my case that was nearly a $100 difference.

Last time I bought renters insurance, in 2006, it was something like $35.

Their interview approach to data gathering is slick though!

They just changed their title to "Saving 80% in 90 Seconds?" after getting a bunch of upvotes from the previous AI-powered insurance title. Talk about manipulative and disingenuous behavior. I flagged the post.

"Prototaph" by Keith Laumer is a short story with this premise. All the computers refuse to insure a man, so he goes on a quest to figure out why. It's a fun read and incredibly prescient, published in 1966!

It's available online legally:

http://hell.pl/szymon/Baen/The%20Baltic%20War/The%20Lighter%...

Isn't that the wrong way around? An arbitrary life insurance on this man should be available for a price of 0, since there is zero risk of ever having to pay out.

I suppose it's almost like a null return, something that can't be computed when the only expected outputs are positive non-zero numbers.

What a great read! The language is a bit bumpy at times, but it was published over 50 years ago.

What does 'BB' mean?

There's an English idiom "sweating bullets" [0] which refers to a state of panic or dread, possibly with the image that droplets of sweat are being forced out at speed.

I believe the author is replacing "bullets" with "BBs" [1] to fit the youthful 21-year-old protagonist. BBs are a much tamer kind of spherical projectile popularized by a variety of teen-toys. The actual origin of "BB" goes back to serious guns again, as a defined size of metal buckshot pellets.

[0] https://en.wiktionary.org/wiki/sweat_bullets

[1] https://en.wikipedia.org/wiki/BB_gun

Ahh, excellent explanation, thank you!

I assumed it meant he was sweating bullets. Like little shots from a bb gun

They've sold less than $100,000 worth of renter's insurance, after spending millions of dollars pre and post launch. Their reinsurance costs are almost as much as their premiums, and they haven't dealt with any large claims yet. Frankly, they have no idea what their overhead and acquisition costs will be if and when they sell a material number of policies.

As far as I know, they use standard ISO loss tables with a 50% claims cost expectation. Even if that pans out, you need to sell a heck of a lot of $60 policies to even cover what they've spent to date.

There are a lot of smart people working in competitive insurance markets. To think AI or Lemonade can drastically reduce the cost of an intangible product is naive. Most of the cost in insurance is insuring the risk opposed to admin.

I have to agree with this comment. Insurance companies are highly sophisticated math machines. They are not sitting around with some majorly outdated process. This is a highly competitive market.

The most likely scenerio is that new enterants will do one of 3 things: overprice insurance, underprice insurance, something illegal like discriminating

I am curious if they are going to make an AI hedge fund. This is how insurance companies make money. If they can create money without having to pay traders and analysts, it just might work.

Insurance companies are _not_ hedge funds. They are also typically quite low margin, volume business. Further, in more common products such as life ins policies, the risk models are pretty heavily regulated on state/federal level.

Insurance companies make money on fees and on "using" the cash you're paying (upfront due to amortized payments) for your policies by making low- to mid- risk investments in various - typically low-liquidity - assets.

(Source: I create risk management systems for insurance companies)

Fair enough. What I meant is that they will shovel funds into the markets in order to actually make money. The best way to do this is to have a higher risk approach plus no human traders.

That sure sounded great to the people running AIG in 2006!

We've come a long way in 10 years. This company was recently discussed on HN https://www.bloomberg.com/news/articles/2017-02-06/silicon-v...

I had state farm (auto) for a few years and at some point their algorithm just decided to screw me more than doubling my premium. My agent couldn't explain it so I just had to switch providers. Even this is basically a submarine for lemonade it'd be interesting to know how many insurers are using Machine learning and if not why not.

I've always wondered if there was a market for a startup like Credit Karma for car/renters/homeowner's insurance. Instead of checking for cheaper rates at policy expiration, the broker startup would check constantly (daily? weekly?) across as many insurers as possible, and automatically have a new policy written for you if it was able to find another policy matching your existing coverage at a lower price.

Why would the rates change so often? I thought they are based on some risk profiles that don't change that quickly.

They did you a favor, State Farm sucks at claim payment.

Different insurers have different underwriting models. When I was a dumb kid, Progressive was the best deal, they sucked once I got married.

If you have a good record/credit, you should get Amica, USAA or Erie, period.

A lot of commenters are speculating as to AI analyzing the risk, which, as they also point out, isn't necessarily a new idea. However, if I interpret the article correctly:

>Indeed it has so been achieved. It is available 24/7, courtesy of a delightful bot who will fashion you a $5 policy in 90 seconds and with zero paperwork.

Between this and the comments about acquisition costs, it seems like they're saving a lot compared to traditional (legacy) insurance companies by cutting out customer service reps and agents.

I'm a happy Lemonade customer. I signed up because a) the policy was dirt cheap, and b) the policy agreement was clear and also quite generous. It was not AI-generated: it's just a PDF like every other insurance policy agreement. And the policy was only so cheap because they let you take out a policy with a lower coverage limit than you would normally be able to find from a traditional insurer. Moreover, I didn't get pushy sales calls after getting a quote (Liberty Mutual was particularly offensive in this regard).

I was skeptical, then my phone got stolen. I reached right out to their "emergency hotline" and told them what happened. After some initial confusion I got to work directly with a customer service rep who basically bent over backwards to ensure that I was satisfied. It didn't seem like a terribly sustainable way to do things, and I wouldn't expect that kind of service once they start to grow, but it worked out great for me and I sure as heck got my money's worth in coverage.

I work in property insurance, and I'm curious as to how lemonade got their product past state regulators if it's truly some whiz-bang AI or machine learning stuff driving their model. Most states want to be able to review and understand filings, and a black box isn't conducive to that. Maybe they're using it for feature selection?

Edit: Found one of their filings on SERFF public filing access. Lookup WESA-130711665 on https://filingaccess.serff.com/sfa/home/NY if you're interested. 'Rate Support.pdf' has the meat.

It doesn't look like anything special, honestly. They admit most of it is copied from the largest insurers in the state, and they use a proprietary credit model from Transunion. This makes sense given that they don't have any data to train it on since they're, you know, new.

The AI stuff they are claiming is to reduce their LAE. (Claimbot type stuff.)

Their rates (and forms) are nothing special--ISO rates + Transunion score. The only thing unusual in their pricing is they allow you to choose low coverage amounts + don't charge installment or credit card fees. So you can buy a decent policy at $5 a month. I am skeptical that will ever be net profitable though....

But...aren't these various use cases for Neural Networks that impact regular people's lives in a significant way sooner or later going to have to show how they come up with a particular decision? And won't that be hard to do?

Yes. AI is the marketing name for complex algorithms that nobody really know how they work or why they do what they do.

This will ruin countless lives.

What? no. "AI" is the marketing name for "machine learning, now with 100% more chat bots". Most "machine learning" is not "complex algorithms". It's mostly just support vector machines, clustering algorithms, and some off-the-shelf neural networks whose behavior is well understood. Whether that's the case here isn't clear, but a project like DeepMind is not representative of AI in the wild.

If it does somehow manage to ruin countless lives, it won't be because of AI. I don't think AI did anything at all for me when I signed up; I didn't even use a chatbot, I just answered the same questions I answered on every other insurance quote, and happened to get a better price with sweeter terms than I got at any other insurer. It also didn't do anything for me when I filed a claim: they have a clearly-defined policy like any other insurer, and I worked directly with a (human) agent to get my claim resolved. The AI stuff is a marketing gimmick.

If it ruins any lives at all, it'll be because the company is selling insurance for a fixed, below-market rate. But Dan Ariely isn't stupid, and I assume they'll at least be able to honor their existing contracts in the event of some kind of widespread catastrophe.

You can look at the gradients into the neural network and see what features are changing the output and by how much. This approach is called LIME, see https://github.com/marcotcr/lime

Interesting. Seems like the insurance industry will change dramatically with the introduction of more data and AI capabilities. Imagine if an insurer ran a search of social check ins at bars against their auto insurance customers. Seems like some people might be paying $2K/mo for car insurance while others pay $40/mo... since with more analysis it will become a lot easier to determine when those big loses are coming and exactly who from rather than guessing/averaging like they do now.

Until they decide that as you don't have an active social profile you are a 'high risk' client :-)

That until won't last long. I know for a fact FB was negotiating last year with a few big insurance companies. Many individual metrics correlated to insurance risk are already available to FB customers, but the FB standard terms limit their use currently.

> since with more analysis it will become a lot easier to determine when those big loses are coming and exactly who from rather than guessing/averaging like they do now.

This is kind of going to break insurance.

When you create more risk pools, you increase the number of low risk people and decrease the number of high risk people who buy insurance. But it isn't clear that having more uninsured high risk people is a societal win -- especially if it means replacing insurance claims with bankruptcies.

Meanwhile "more [low risk] people with insurance" is not an unmitigated good, because people will buy things when the insurance is paying for it that they wouldn't when paying out of pocket, which significantly increases costs.

And it's easy not to have much sympathy for Those Dirty Drunks About To Get What They Deserve, but it's really the person the drunk hits who suffers when the drunk was driving without insurance because it was prohibitively expensive. Or the hospital who has to treat a bankrupt patient with no medical insurance.

And being high risk isn't always a result of some moral failing on your part. Doing genetic testing for medical conditions would certainly allow people without high risk markers to pay less for insurance, but what happens to the people who have them?

In the limit where the predictions become extremely accurate there is no point in even having insurance because your net lifetime insurance premiums will be so close to your net lifetime insurance claims that the only thing the insurance company is providing you with is overhead.

But if any insurance companies do this then it doesn't even matter whether all of them do, because the ones who do will get all the low risk customers and the other insurers will de facto end up with a high risk pool with high premiums. So we really need to decide if we actually want this as a matter of legislation.

Insurance is already regulated at the state level, although it varies from one state to the next. One possible solution to your scenario is risk pooling like North Carolina does with NCRF.

The purpose of insurance is to protect you from unpredictable or unlikely risks. If a risk is predictable and likely, what's the point?

If you want to distribute risks fairly, that's more a job of government subsidization. It is unfair that someone born with a genetic disease has to pay large medical bills, and I'm ok with the government subsidizing that.

On the other hand, I don't want to subsidize the risk of serial drunk drivers. It's totally fair they have to pay higher premiums or not be allowed to drive.

Ideally there could be a compromise. Like insurance companies agree to insure bad drivers if they take remedial driving lessons.

It's theoretically possible we could use insurance as a free market solution to government regulation. Instead of building codes, require building insurance. And the insurance companies will do their best to actually predict the risk of various disasters and what things actually work to reduce them.

> The purpose of insurance is to protect you from unpredictable or unlikely risks. If a risk is predictable and likely, what's the point?

That's the problem. The more accurately we can predict insurance claims the less reason there is for anybody to buy insurance because there is less uncertainty to insure against.

And one of the things existing insurance does unintentionally is to spread the cost of misfortune across more people. If the misfortune becomes more accurately predictable then the people not likely to suffer it are spared the cost but that only moves more of it to the people who do.

But you make an interesting point in the sense that it isn't enough to prevent insurance companies from using the information. If insurance customers can accurately learn their own risk, the same result will obtain regardless because the low risk customers will opt out of insurance until the premiums are high enough that the high risk customers do too.

> And the insurance companies will do their best to actually predict the risk of various disasters and what things actually work to reduce them.

Unfortunately that is completely the opposite of how insurance actually works. The insurance company has little control or involvement in such things because it isn't cost effective for them to micromanage everything. So what actually happens is that if you require insurance, people buy it and then take more risks because they have insurance.

The fundamental tenet of insurance is to buy predictability at the cost of inefficiency. If you independently improve predictability the only thing left is the inefficiency.

The death spiral. I get it and I agree with your analysis that this has the potential to "break insurance." I don't think the direction insurers are headed in with this is good for society or even their own industry but the companies who do it first will get a huge benefit in lower risk pools... so absent legislation to prevent it this its is happening.

Great marketing until you realise there are no links to their product or home page anywhere on this page.

I work in the insurance industry and for them to extrapolate the average customer acquisition cost for car insurance to a renters customer is actually pretty laughable.

Mods / Dang, the actual title is "Saving 80% in 90 Seconds?" "When Tech Makes Insurance 5x Cheaper",

The tldr is: "our acquisition costs are already 10x lower than legacy carriers."

Ok, we reverted the title. Submitted title was "A “killer business model”: AI in the insurance industry".

Submitters: please use the original title except when it is misleading or linkbait – that's in https://news.ycombinator.com/newsguidelines.html.

I'm sure no one in the insurance industry has thought of using AI....

this is actually pretty interesting, thanks for sharing. didn't realize that upfront CAC represented the lion's share of E[cost] in renter's insurance...

Doesn't Geico do the same? They also offer renter's insurance next to car insurance. Also what does using AI mean exactly?

So...stats (the same, I presume that a normal insurance company uses) and an online-only, agent-free model?

Doesnt work in SF ... anyone knows why?

Only filed and approved in NY so far.

What does AI have to do with anything?

AI is becoming a marketing buzzword. All software that do anything remotely clever or interesting is being presented as "AI" helping you solve business goals.

> So how can Lemonade be 80% cheaper?

Well, it's just sugar, lemon, and water...

Reply on news.ycombinator.com