Paper Cuts, Prediction and the Future of Insurance

In defiance of all expectations, something exciting is happening in insurance. Something with the potential to drive the industry to a crossroads: fundamentally transform or cease to exist.

To understand, let’s imagine that you just started a business.

Paper Cut Insurance

1st Iteration

You offer insurance for paper cuts. All you know, from past experience, is that paper cuts affect 1 in 5 people and treatment for a cut costs $30,000. Five people purchase your insurance. The result looks something like this:

The cost of a potential treatment is split equally across the group into $6,000 chunks. When someone ultimately slices their fingertip, everybody’s payments will cover the $30,000. You charge 20% on top of this to keep the lights on and pay yourself a salary. You also make some extra money by investing the $30,000 before you have to pay it out.

This is the simplest way to think about insurance. The world, however, is a bit more complex.

2nd Iteration

You read a new medical study. It reveals that people with office jobs are 50% more likely to get a paper cut! When another set of 5 people purchase illness insurance, 2 of them are office workers. This time, you split things a bit differently:

The customers who are 50% more likely to get nicked pay 50% more. You collect the same total amount of money to cover costs, invest for return, and pay for treatment as you did previously - it’s just distributed to better reflect the risk each customer carries.

This is closer to how insurance companies operate today. Actuaries do their best to predict the probability that a claim will be made. When they identify that certain groups of people are more likely to make a claim, the cost of insurance goes up for that group. Smokers are more likely to get lung cancer - they pay more for life insurance. Teenage boys total more cars - they pay more for car insurance. It’s effectively personalization, breaking a big group into smaller groups to make more precise predictions.

Insurance companies personalize because it makes good business sense. Charging less for lower risk groups helps companies attract more low risk customers and it keeps prices competitive with other insurers. Charging more for higher risk groups ensures that, as riskier customers sign up, they’re paying enough to cover the cost of increasingly likely payouts.

So, insurance companies take the information available to them, do their best to predict the likelihood someone will make a claim, and use what they find to inform personalized pricing. Let’s take that to the extreme.

3rd Iteration

You now develop advanced algorithms to aid your insurance business. You feed these algorithms with detailed profiles of millions of people who have been cut and who have not. The result: 100% accuracy in predicting whether someone will need treatment for a paper cut. When the next 5 customers show up, allocating the cost of treatment is trivial:

Four of your customers are no longer customers - there’s no sense in hedging against nothing. The fifth is now bearing the full cost of treatment plus your cost of operating - they’re better off going uninsured and paying for the treatment outright.

Assuming they do just that, your business model breaks down - no customers to serve, no money to invest, and any costs you might still have are a pure loss.

The Verdict

Insurance is about sharing risk as a group - providing security against improbable, but financially devastating events by distributing the cost. Personalization is about making those groups smaller. Take that to the extreme and you’re no longer sharing. The benefit of insurance is eliminated.

For one person, the simplified relationship between the price of insurance and the predicted likelihood that they will make a claim looks something like this:

The price of insurance increases as the predicted probability of an incident increases. Because incidents are real things that either happen or don’t, better prediction will drive the probability closer to either 0% or 100% - the two extremes we saw in our last paper cut example.

So, while we’ve seen that perfect prediction makes insurance pointless (Why hedge against uncertainty when you’re certain?), perfection is unlikely. The pressing question is “What happens when we get close?”

As the probability of an incident approaches 0%, the price goes down to almost nothing meaning insurance companies make less money off each person. This works as long as the number of customers buying the product goes up correspondingly - it’s the low margin, high volume strategy insurers apply to low risk customers today. Unfortunately, relying on sustained volume increases is dubious.

Somewhere in the process of incident probability approaching 0%, cheap insurance crosses the line from being a good deal to being superfluous - it might feel prudent to be insured for a 1 in 5,000 chance of incident, but 1 in 5 million borders on superstition. What’s more, the line where consumers decide insurance isn’t worthwhile is getting easier to cross as the next generation proves increasingly less interested in buying insurance at all. Insurance stands to become a niche, but accessible luxury for the shrinking population who prefer an extra whiff of caution to good chocolate or VIP movie tickets.

The case is more bleak for those whose probability of an incident approaches 100%, people who are very likely to need treatment or support. You don’t need perfect prediction to price people out of the market - you don’t even need to be close. For most people, paying for 10% of a $1,000,000 medical treatment is just as inaccessible as 100%. For those who can afford it, the fact that insurance companies need money to operate and make a profit means you still don’t need perfect prediction to be better off going without.

So, perfect prediction makes insurance pointless and really, really good prediction has insurance companies charging a shrinking group of people less and less for protection against threats that are barely real while boxing out those who desperately need support.

It’s a dire picture. Fortunately, it’s not one we have to worry about unless prediction gets really, really good.

Prediction is getting really, really good

Artificial intelligence (AI) technology is improving. Google’s AI AlphaGo beat the world’s best human Go player. Its successor AlphaZero beat the world’s best software for Go, chess, and shogi. Self-driving cars are a real thing you can buy.  

As AI gets better, so does prediction. That’s because, distilled to its simplest economics, all AI does is make prediction cheaper, faster, and more accurate. These can be real-time predictions (Is that man suffering a heart attack based on his symptoms?) or, in the case of insurance companies, predictions of the future based on the present (Will that man suffer a heart attack based on his health habits?). AI technology is making really, really good prediction a very real possibility.

Further, the data that AI relies on is getting better, accelerating the pace of prediction improvement. Life insurance predictions improve when they’re informed by a real-time view of activity levels from every Fitbit and Apple Watch. Home insurance predictions improve when security systems, thermostats, and smoke alarms are linked to a coordinated database. Car insurance predictions improve when informed by real-time speed logged digitally by the vehicle or just calculated from sensors in the driver’s phone.

The speed of data creation, variety of data sources, and volume of data accessible continue to skyrocket. Everyday people generate troves of information through connected devices, credit card purchases, internet search histories, and more. All of it stands to better inform prediction.

While some data sources, like DNA, have been marked off limits by governments, policy often struggles to address complex technical issues. AI has already established a reputation for outfoxing government limitations on insurance data usage, thwarting anti-discrimination laws by using addresses as a proxy for income and occupations as a proxy for gender. Advances in prediction will not be legislated to a stop.

These AI advancements aren’t something that might impact insurance - they’re being put to use today.

Insurance companies are getting on board

Major insurance players are doubling down on predictive technology. Unitedhealth Group, the largest insurance company in the world, is already boasting about “AI with an ROI” and making progress in their ability to predict heart conditions. AXA, Europe’s largest insurer, is using AI to read engineering reports and assess property risk. Manulife has launched a new life insurance product with the expectation that AI will be doing most of the underwriting. The list goes on.

Start-ups have also hit the scene. These ‘InsurTechs’ are dedicated to enhancing insurance with technology. AI and better data usage are major themes. Oscar: AI-enabled health insurance. Lemonade: AI-enabled home insurance. Chisel: an AI solution that takes unstructured data and (Getting a bit wild) makes it useable for more even more AI applications. These aren’t just a blip - average annual InsurTech investments for the past 5 years are roughly 6x what they were in the 5 years before.

So, while insurance companies have not yet reached game-changing prediction capabilities, evidence suggests they are doing everything to get there as fast as they can.

The Opportunity for Transformation

It is possible that, as prediction gets better, insurance companies will take advantage of information asymmetry to cling to their current model -  just because an insurance company knows that you’ll never get a heart attack doesn’t mean that you know. An insurer could use really, really good prediction to accurately identify low risk customers and keep their prices artificially high. This seems unlikely for a few reasons.

First, competition creates clear incentives to drive down artificially high prices. While temporary collusion between big incumbents might be possible, the multitude of insurance companies and active startup scene would make it unsustainable.

Second, maintaining the illusion of hedging risk where none exists would be an outright scam. A scam that would, in the most cynical case, be vulnerable to whistle blowing and, in a more optimistic case, be shunned by insurance leaders as a legitimate strategy.

The most compelling reason that suppression is unlikely, however, is that powerful prediction is too valuable to suppress. Not to inform the distribution of risk, but as a tool for reducing it.

Therein lies the transformation.

An implicit assumption in the discussion so far is that the future states being accurately predicted are fixed. In some cases, like inherited genetic disorders, this is true and the insurance model breaks down as described. But in most cases, the statistically predicted future is not immutable - there is a chance to intervene.

A university graduate gets their first job in a new city. While recreational sports kept them active in school, the new job is sedentary and they don’t have any friends in town pushing them to sign up for soccer. Four months later, they’ve put on weight, eat most lunches at the nearby McDonald’s, and buy a 20 oz bottle of Coke from the office vending machine every afternoon.

They get a phone call from the company that provides health insurance through their work.

The agent’s explanation is kind but firm. Predictions suggest that, if the graduate’s last 4 months of diet and exercise habits become the norm, the risk they will develop Type 2 Diabetes in their lifetime will exceed 75%. Fortunately, as part of their health care package, they qualify for coverage of $1,500 in personal training sessions and $750 of healthy meal delivery services. If they confirm they’d like to take advantage of these benefits, claims will be automatically reimbursed based on the insurer’s access to credit card transaction data. Intervening now will constrain Type 2 Diabetes risk levels to the standard 5%.

Or, more simply, if you run an insurance company and know who is going to get a paper cut, you are uniquely positioned to sell them a glove. Focus on prevention rather than treatment.

Insurance companies are well suited to play this role. They are set up as major aggregators of personal data, they have troves of data and advanced prediction models already, and their current business model incentivizes them to continue improving the accuracy of that prediction. Others who traditionally act as stewards of well-being - doctors, hospitals, driving instructors - lack the access, ability, and incentives to make predictions anywhere near the scale possible for insurance companies.

So the theory, in a nutshell, is this:

Prediction is improving. As it does, insurance companies’ current business models will be challenged. This will correspond with an opportunity to provide value by preventing incidents rather than funding their treatment. The best path forward will be to embrace this opportunity and transform.

A theory, though, is just a prediction. In the meantime, it will be interesting to watch and see what comes true. We can’t be certain of what the future holds.

At least, not yet.