You only make insurance cheaper by charging risky people more.
Right now it is mostly laws that protect categories of people that keep insurance companies from charging people more.
What’s the plan here, use machine learning in a “hands off” way with a black box algorithm to apply pricing discrimination in a way that a human could not because of regulation?
I'm tangentially involved in the insurance space and I believe Lemonade is trying to use machine learning to process claims because:
- Processing claims with humans is expensive; every step that can be accomplished by a computer will probably be cheaper.
- A claim processed via ML will probably be handled fast. A fast response = happy customer, which helps with retention. This is a big one.
- A claim that is processed and closed quickly is harder to amend. Some customers slowly realize that adding items to a claim is free money. Others (legitimately) forgot items and want to add them. A quick claim is usually cheaper than one that might take a few days (or weeks) to process.
- Younger generations are more used to working with a web pages and will likely look at humans (e.g. agents) as old-fashioned.
The big carriers are both scared and dubious of Lemonade. If Lemonade can somehow make it work they could do serious damage to the carriers. But it's hard to see how they'll make the numbers work, as their current losses show. Most of the carriers are trying to implement something similar (which is where I'm slightly involved).
Depending on your perspective of "ML", the insurance industry already uses "ML" (i.e. very complicated decision trees) to process claims. Very few large insurance companies are non-automated in claims processing.
The places where the money hides, so to speak, include (1) handling complex cases [customers] (2) scaling a human's ability to process non-automatable settlements. (3) scaling internal support interactions with customers (4) introspection to claims data and support data. (5) graceful handling of prior authorizations.
These problems are not as attractive, but they are where insurance companies spend most of their money. It's still a tech problem, but it's not super fancy.
Existing carriers struggle to solve these problems, because they have historically grown by acquisition, and as such do not have the kinds of unified data systems required for the rapid development of applications that perform the required kinds of introspection. It's a space that's ripe for disrupting.
I’d say customer acquisition is the biggest cost and insurance companies are terrible at it because differentiation is almost impossible in a price driven by extreme price war.
It’s not uncommon that 30-50% of your travel insurance premiums are going to a broker or price comparison website. Talking about great value.
It’s not as sexy as ML in claims but one of the big innovations Lemonade has developed is looking like the anti-insurer and creating a huge PR machine around that. True or not, it worked.
I’ve observed this, but I don’t think that any carrier has successfully figured out a way to grow that doesn’t involve brokers (digital or physical). They have a death grip on the market, with a very high percentage of potential members held behind their gate.
Like you, I’m not convinced that the value they add to the chain justifies their expense, but they’re legally and economically entrenched.
My only hypothesis for their eventual dissolution is that unit commissions will get smaller and smaller over time, as more brokers use tech to manage bigger books of business with lower employee headcount requirements, and brokers become more indistinguishable from carriers.
Or maybe some consortium of carriers will get together to build THE comparison shopping site, like healthcare.gov, and offer some ridiculous bonus payment to you the member for shopping there.
Yep, this is what I meant. The "first generation" of these engines were written in COBOL, and are quite nasty by modern standards.
Most people in the tech industry don't apply the ML label to these kinds of software. However, there is a ton of knowledge invested in these systems: certainly more than any one currently-living human has in their head.
I've been contracting with a mega-big insurance company for a few years. From discussions with fellow contractors in siimlar-sized insurance companie, we all tend to reach the same conclusion..
Ιmagine a spaghetti code. Now think of an IT eco-system that has grown with half-baked interfaces, patches, MANY MANY MANY MANY spreadsheets, legacy systems (and I mean having MS Server versions 10+ years old), and nobody dares to change passwords because there will be pain.
Now imagine that a mega-big insurance is growing by M&A, and it becomes mega-big by absorbing other companies, (speghetti-zed) IT eco-systems, and they try to absorb processes and data with even more spaghetti-ish manners. One sneeze and the world is coming down...
Now consider a new, modern company that has the talent, and none of the inheret problems/risks. They may not be mega-big (yet), but they got the potential to slowly steal chunks of the market, only if/the grow clean, fresh, and organically.
In the UK it's a (growing?) trend to have your app on someone's phone, tracking motion from the phone's sensors (accelerometer, GPS/speed, etc.) and they can give you lower premiums. The possibilities are endless... as long as they are modern enough.
I'm not sure what's in it for me as a consumer. I did a quote with them a couple weeks ago, and even with all the "discounts", it came in at almost double the premium with worse coverage.
I just checked, and for my apartment they come in at about half of my current renter's insurance for similar coverage. Maybe it's highly dependent on location, or maybe I'm just way overpaying for renter's insurance. Unfortunately for them, switching away from the bank that already handles 95% of my finances isn't worth saving $100/year.
(Keep in mind that I am just an engineer and don't deal directly with policies.)
My understanding is that it is generally considered that renter's insurance is too high at most large carriers. There are historical reasons for this which I won't go into (and I don't understand all of them anyway). The carriers don't change this because:
1) Lowering renter's insurance premiums would require raising premiums elsewhere (e.g. homeowners). Raising rates causes customers to leave.
2) Renter's insurance is fairly cheap as it is and people don't price-shop all that much.
3) Shifting premiums brings regulatory scrutiny and you have to do it right (and legally). It's not worth the hassle.
I think Lemonade started largely with renter's insurance to take advantage of that gap. It's one of the things that put fear into the major carriers.
Before I started working in the business space I knew almost nothing about insurance (and didn't care). It's been fairly interesting and I have more empathy for the carriers than I used to; they aren't quite the blood-sucking maggots that some maintain.
>Maybe it's highly dependent on location, or maybe I'm just way overpaying for renter's insurance.
Or maybe Lemonade is like many other "tech" companies, and is selling their product below costs in the hopes of growing into profits. From a consumer standpoint, gambling with your insurance in such a manner is scary.
It’s unlikely the estimate varied much from what other insurers would offer you, assuming you were comparing the same building materials and coverage. If Lemonade were so expensive, their loss ratio from the S-1 would be much better and they’d have fewer customers.
(Anecdotally, they are offering me about 95% of what I pay Geico for similar coverage, and without the auto insurance discount.)
I don't work directly for insurance companies so I am not a good judge of what jobs will be around. I would expect that any job that deals manually with claims--line entry, price evaluation, even fraud detection--will decline, maybe a lot. However, I don't see them going away entirely--there are just too many anomalies and the data you're dealing with is too dirty. That's not sticking my neck out too much; you can say pretty much the same thing for any broad industry. :)
I assume that Lemonade is banking on doing away with almost all manual processing so they're ideas are different from mine.
I think we’ll see the insurance market transform in a similar way that we’ve seen the financial services market transform. Although, it will go in slow motion, since the gains in insurance aren’t as immediately-accruing as they were in equity and credit trading.
You’ll see the same kind of jobs that exist today, but they will be smaller in number, and way less paper-oriented.
Many. Insurance is hardly one type of job and many lines of insurance (e.g. for larger multi-national corps, for certain types of property, and etc.) aren't automated anywhere near the level of what you see in Lemonade's marketing/website/systems.
It also seems they may see this filing as a way to reinforce their marketing as a “good” company that directs funds where they say they do. I could imagine them telling their customers to buy their stock as a way to be involved with how they operate.
Regulators are starting to lean in on ML rules and requirements. They can't hire data science experts to keep up with competitive demand and salaries, so they are going to require companies to make their ML-based outcomes accessible/auditable. Claims, pricing, risk modeling, underwriting - we're in the early days of companies using ML for these tasks.
At most big old and public insurance companies, claims payable represents a significant chunk of expenses, but not even close to 100% (it's closer to 60-70%). The rest is, generally, "administration" (humans processing papers, and managing humans processing papers, in cushy offices).
This is where better technology can result in lower costs. It's a volume/unit-cost game. Their unit cost per person is maybe a few cents or a few dollars cheaper, but at huge volumes it makes a big difference.
In addition to increasing efficiency, another way that comes to mind is to lower risk for the population as a whole. For example, investing in safer building codes, local emergency services, mass transit, etc.
Yep! One way that insurance compmanies can achieve this is by providing members of their insured population access to services that reduce their individualized risk.
Pre-COVID, many health insurance companies (my industry) were gearing up to offer free Doctor on Call (a service that, if well-implemented from a tech PoV, has near-zero margin costs), because access to such a program reduces the risk of expensive claims later down the line.
I'm sure there are equivalents in the kind of insurance that Lemonade provides. For example, they might offer free or heavily subsidized home security installation in certain zip codes with a history of burglaries.
> At most big old and public insurance companies, claims payable represents a significant chunk of expenses, but not even close to 100% (it's closer to 60-70%).
Yep, this is a huge caveat that this thread missed. Thank you for raising it.
For health insurance, the rule is:
> Health Insurance companies must spend at least 80% of the money they take in from premiums on health care costs and quality improvement activities. The other 20% can go to administrative, overhead, and marketing costs.
As a health insurer, you can lower premiums while increasing spending on "Quality improvement", to provide a better experience at a lower rate, and increase your market share. This is one dimension of competition that is only beginning to be competitively explored.
If you can get quality improvement at lower marginal costs (which is ultimately a tech problem), you're a more competitive health insurance company.
Which creates a perverse incentive for insurers not to care about payouts (If you want to increase profits, you have to increase payouts) -- so long as they can compete on costs with other insurers.
Controlling medical expenses isn't easy. Neither is getting admin + overhead under 10%. You need both to get a reasonably profitable insurance product.
I think this summarizes why single payer health insurance makes so much sense. The insurance companies are taking about 20% of health care dollars and provide no value.
If you assume it costs the health care providers the same amount to file a claim as it takes to process it, then the total overhead of having the insurance industry jumps up to > 30%.
I’ve heard the US numbers are comparable to that, but it is nice to be able to derive it from first principles.
Interestingly, this analysis suggests that the cost of filing a claim should be paid by insurance, not the customer.
Currently the cost is an externality for the insurance company, so they can waste time with nonsensical revisions and rejections before finally paying out.
If the insurance company had to pay for the paperwork on both sides of the process, they’d have a strong incentive to streamline claims.
Insurance can be cheaper if the customers lower their average risk burden. Lemonade tries to explain this, but not clearly:
“We seek to encourage good behavior and build a long-term relationship based on mutual trust by endeavoring to decouple our financial incentives from variability in claims. In our model, we minimize any incentive to deny legitimate claims as we aim to give back, rather than pocket, leftover monies. After our customers purchase a policy, we ask them to designate a charitable cause for us to support with the residual premiums from their policy. Despite there being no contractual obligation requiring us to donate leftover premiums to nonprofits, when a customer embellishes a claim, such customer reduces the total amount available that can be contributed to nonprofits. As a result, we believe customers are less inclined to embellish claims as they would be hurting a nonprofit they care about, rather than an insurance company they do not.”
Not disagreeing that is part of the strategy, but also it is worth thinking how much overhead there is in the insurance industry. How many offices are there nationwide? How many of the jobs are essentially basic data ingestion? Approving of claims? How much is spent on advertising?
> Hard to put together VC-sized returns out of trimmed fat.
That depends on how large the industry is, what its cost structures look like and how price sensitive it is.
A small, persistent cost advantage can be enormous in the insurance industry.
You're also saying that in a thread about a company that just produced VC-sized returns and is IPO'ing. Your premise clearly doesn't follow, as most VCs invest early and will exit with an IPO like this. The primary question going forward with Lemonade is for public shareholders and whether the company can get a lot bigger in the future. The VC-sized returns were already generated for the early VCs.
As a fairly happy user, it's not clear to me how Lemonade solves these problems and ultimately ends up making money by trimming margins.
They still market heavily, and I presume they paid well for their (pretty good) branding.
When I applied there was a wait between applying and having coverage: an amount of time I suspect correlated to a human having to review materials.
My local market has weird insurance regulation, so I had to reach out for customer-service to ensure coverage which wasn't explicitly in the policy. The regulation and compliance department of any insurance company is likely rather large and expensive and rather difficult or risky to offload to AI.
I haven't had to make a claim, but I get the strong opinion that automated claim approvals are basically a function of claim amount and claim history. Anything over $X or for your Nth claim still involves a human at least as a gatekeeper. The AI may be real, but it seems strongly like the usual "fake it till you make it" AI that lots of companies claim to have but are actually 'mechanical turks'.
Does anyone know what an insurance agent makes? If I go to my local State Farm office to get a home owners policy, what is the cut that goes to the local office/agent?
It's generally quite low, but it depends on a large number of factors, and varies a lot market to market. It's generally a couple percentage points of your monthly premium, per month.
An agent generally needs a couple hundred paying policies to be in the black.
You don't. The point is to charge based on risk. There is no sense in which you can transfer gains from one set of customers to somewhere else. The profitability of any group of customers depends only on the price you charge them.
And btw, lots of insurers specialise in pricing high-risk customers. If another insurer comes in and tries to subsidise low-risk customers using high-risk customers, then a specialist insurer just comes in and undercuts them profitably.
Even a low-risk customer becomes a bad risk at the wrong price. It is all about the price.
> apply pricing discrimination in a way that a human could not because of regulation?
I've read several anecdotes of people hard coding hacks into black box algorithms which end up being discriminatory even when stuff like race is not a direct input. I do not think the law cares how discrimination is arrived at.
That's true, and AFAIK know due to this, insurances that are bound to those regulation don't touch black box ML with a 10 foot pole. When we were pitching ideas to an insurance company ~5 years ago they basically said "If it's not human-explainable we can't use it".
Or you could lower overhead by cutting costs, hiring less people, changing commission structures.
The nicest business meal I ever had was during a lunch meeting that I tagged along to with a large insurance carrier. They could probably cut out $30 steak lunches to lower costs too.
One interesting approach would be to actually charge less risky people less, and "safe" people more, as a way to equalize various groups in the society. Once this has enough traction, doing it the other way round could be even stigmatized.
Right now it is mostly laws that protect categories of people that keep insurance companies from charging people more.
What’s the plan here, use machine learning in a “hands off” way with a black box algorithm to apply pricing discrimination in a way that a human could not because of regulation?