This is how AI fails in (Re)Insurance
A Pre-mortem
Start Disclaimer: The views and opinions expressed in this blog are entirely my own and do not necessarily reflect the views of my current or any previous employer. The information shared is for informational and discussion purposes only. Any reliance on the information provided in this blog is at your own risk. I do not make any representations or warranties regarding the accuracy, completeness, or suitability of the information.
This blog may also contain links to other websites or resources. I am not responsible for the content on those external sites or any changes that may occur after the publication of my posts.
End Disclaimer
AI Johnny Appleseed
I was speaking on an AI panel recently and I found myself saying things like “AI in reinsurance is happening” and that it was “already here”. Look at me- a veritable AI Johnny Appleseed for (re)insurance. Why do I continue to pitch AI, and particularly, why now? AI and machine learning applications are at an inflection point due to a confluence of compute and model architectures, as evinced through the innovation and visibility of generative AI in 2023. AI (let’s call it “AI to sell it, Machine Learning to build it”) is going to be ubiquitous, and there will be tremendous benefits to the (re)insurance industry, from automating the manual boring parts, to enabling full quant trading strategies for some of the more commoditized underwriting lines.
In some ways, the industry needs to reach a critical mass adoption point so that everyone gets to see the ancillary benefits that happen from the collective identifying, cleaning, and aggregating of the data over time. These processes in turn lead to better collective decision making. Down the line, these improvements can lead to problems similar to those that occurred in the Quant meltdown of Aug 2007, but that's a long way off, and those types of issues, e.g. risk model convergence and correlation assumptions, would signal a much more mature, AI and machine learning, augmented (re)insurance market. For now we need everyone working together to achieve enough power to leave the runway before we worry about falling from 30 thousand feet.
AI in (re)insurance seems so inevitable to me that I’m doing a pre-mortem on what can go wrong. I present to you an incomplete “Reverse Bucket List” of things companies would need to scratch off for the adoption of AI in the (re)insurance industry to be a non-starter. I want the (re)insurance industry to prosper and thrive through AI and machine learning adoption. Let’s look at a hypothetical future where that doesn’t happen.
Bosses say No
Bosses are the ultimate forcing functions. Let’s say you’re Jessie, a highly placed executive at a (re)insurance company. You’ve made it this far. You’ve survived multiple cycles in the markets, seen a lot of changes. Someone from within your company (or outside of it) comes to you with a plan- AI and machine learning can help in a lot of ways- scale your decisions and cut the time it takes to make them, reduce your expenses through automation, even make fairly precise predictions about the future.
Person with the AI plan: “It’s going to take this many people, this much money, and this many years to build out our AI roadmap”
Jessie: “That's a lot of resources to devote to this. Besides, we have been profitable for many years. Why would we change things now?”
Record scratch, the idea is grounded- over before it even started. But don’t despair. I’m sure there’s still a lot that can be done with Excel, VBA, and Power BI (spoiler- there isn’t. That lemon’s been juiced)
Dedicated AI Teams don't get Hired
If you are lucky enough to be at a (re)insurance company where management understands the importance and relevance of AI and ML, you’re still not out of the woods. An AI/ML team should be different and distinct from your IT team. They should report to a different person, preferably one of the underwriting heads. Right now in (re)insurance, the majority of underwriters, actuaries, claims people, and down the line don’t know what an AI team does or how it could help them. So why would anyone want, or know, to hire them? A good AI team will cut a clearing in the forest with a direct sightline to higher profits and lower costs. As Henry Ford said, “If I had asked my customers what they wanted they would have said a faster horse.” Steve Jobs said it similarly, “People don't know what they want until you show it to them.”
The Head of your AI Team is a Theoretician
What’s a bad “head of AI” at a (re)insurance company in this context? Someone who’s never had skin in the trading game before. (Re)insurance underwriting is hard. Someone who comes from a theoretical background and has never had a model explode on contact with the real world is going to have a tough time in the industry. The best risk managers have traded before. The best AI head will have too. They will know how to articulate monetizable ideas so that you don't float in the “let’s do AI at our company” miasma. This is crucial. A good AI leader will understand asymmetry and fat tails and know how much worse it feels after a trading loss than it does after a win(Tversky/Kahneman)- in short, someone who bears the scars of practical experience. The first (re)insurance Chief Algorithmic Underwriting Officer is working today, just probably not in the (re)insurance industry.
Your Systems Can’t talk to one another
So you got the green light from your boss, even hired a dedicated AI team led by someone who’s seen some things. Where to start? Does your company have separate systems for everything? Do you have a Frankensystem cobbled together from all the corporate actions and mergers and acquisitions? Are your policy, claims, accounting, or actuarial systems written in a coding language where all those coders have retired?- think Powerbuilder (Ada, Pascal, Objective C- I had to look those up). The basic machine learning equation to solve for, broadly speaking, is “policy features = frequency/severity claims features” (an oversimplification but that's the gist). This equation is hard to conjoin if systems silo and wall off their data ‘Cask of Amontillado’ style. Startups and Insurtechs have an advantage here- the ability to start tabula rasa. For those with legacy systems- start over and run in parallel with old systems. Avoid tacking on to the existing Frankensystem- in effect creating future vestigial system appendages. To quote Henry Ford again, "if you need a machine, you pay for it, whether you get it or not."
Your Data isn’t Cleaned
So you got the green light from your boss, even hired a dedicated AI team led by someone who’s seen some things, and now your systems can talk to one another. Happy day! Wait. Well, it’s going to be hard to run that XGBoost model if your data is dirty- will be even harder if you don’t know where your data is. The data science trope that you spend “80% of the time cleaning the data to spend 20% time applying machine learning algorithms” is real. Our beloved trope made it through its awkward “I used to work in finance for a bank doing stats until the Great Recession, but now I'm called a ‘data scientist’ ” adolescence and matured into a trope because it's true- and 80% is probably lowballing it. You want to “do AI” at your company? Might need to have all those words like “insured”, “insured name”, and “insured company” all tie back to the same thing. Tough thing about any domain specific and siloed data set- they really need the domain expert to look over them. You can automate the cleaning to a point, but it takes a human to go through them to sanity check the numbers. The Data Science and Machine Learning team is not going to know if that order of magnitude limit is a real outlier, or priced in a different currency someone didn’t bother to annotate.
(Bonus- You can also adapt and fine tune large language models with reinforcement learning from human feedback (RLHF) to do this. That’s another type of machine learning project that will become more commonplace in the industry over time but, again, not without human intervention and not without…good data and all the previous dominos in place.)
Google hasn’t “solved for (re)insurance” yet
Everything is ready. The machine learning and AI table is set. Time for AI success. Unfortunately, you arrive at your desk one Monday morning to find that insurance has been solved. There’s probably some scraps left over, but now there’s fierce competition for what remains. FOMO, particularly existential FOMO, is strong. But why hasn’t this happened yet? Why hasn’t Google or Amazon or some other company “solved” for (re)insurance using their boundless AI resources and forced the (re)insurance industry to play catch up? Is the return profile just not interesting enough to go through the trouble of running everyone over? What about startups, smaller companies? Where are the insurtechs who were going to solve it? There’s a lot of reasons for “why not right now” and why an AI solution hasn’t already happened. Insurtech companies (some who may skew more tech than insurance) still need a balance sheet. Big tech companies don’t want to risk turning away from their core competencies. An industry wide, reactionary embrace of AI happens, if and when another company figures out a way first. The FOMO would be just too powerful. Google hasn’t done it…yet.
Don’t Slow Down
AI and machine learning are here. AI in (re)insurance is happening. The technology and its utilization will be ubiquitous. Get things approved, hire your team and hire the right person to lead it. Connect your systems. Clean your data. Connect the dots. Google and Amazon have other things on their plate now. Keep going. The future is bright. Don’t slow down.