AI isn't just speeding up tasks - it's also helping insurance companies make better decisions in everything from processing claims and applications to risk assessment
As AI penetrates the insurance market, Neal Silbert, GM for insurance at insurtech DataRobot, tells Peter Littlejohns about the impact automated machine learning will have on predictive analytics in the industry
Artificial intelligence is already saving time and money for insurance providers – and now they’re hoping its use in predictive analytics will revolutionise the way they assess risk.
While the application of AI has until now mostly been limited to accelerating tasks, Neal Silbert has identified how it provides assistance in making decisions too.
As general manager of insurance at Boston tech firm DataRobot, he is in a prime position to observe established insurers automating everything from processing claims and applications to risk assessment.
He says: “Process automation can be called AI, but unless the software is using AI to make a decision, I don’t think it is.
“It’s like a spreadsheet formula with the ability to move the data from one programme to another, it can make some very important decisions, but it isn’t necessarily using AI to make those decisions.
“Machine learning is a way of looking at the data that’s available and automating the creation of statistical formulas that will then predict the best way to do something.”
AI will make predictive analytics in insurance more accurate
Actuaries are a pivotal part of any insurance company, as it’s their work in assessing risk that forms the firm’s entire pricing structure.
Given their importance, it’s unlikely we’ll see actuaries being replaced anytime soon.
But in order to create the most accurate models, they will have to take advantage of AI, believes Neal.
He says: “Machine learning does not replace actuaries, it allows them to spend more time focusing on the difficult problems, and more time on the parts of a problem where they can make a difference.”
The current models used by actuaries are “20th century technology,” claims Neal, who also draws a distinction between normal machine learning and the automated machine learning technology invented by DataRobot for a wide range of industries that also include banking, manufacturing and retail.
“Normal machine learning would be a process where you input your data, run it through one or more modelling techniques and you’re given an output – a score or prediction,” he adds.
“The problem with this method is that if the result isn’t perfect – and it never is perfect – you have to go back and change your model, so people would spend anywhere between 60% and 80% of their time trying different variations of their data, which could also mean restructuring it so it’s on the same scale.
“Automated machine learning does a lot of this restructuring for you, allowing you to spend more time adjusting your model so you can have maybe 93% to 95% confidence in it, as opposed to 80%.”
The ability to automate data restructuring means that automated machine learning can be applied to test data against “millions” of different models at once, allowing an actuary to assess the best model to use for a specific problem.
Machine learning will make predictive analytics in insurance more consumer-focused
Management consultancy firm McKinsey recently predicted that manual underwriting will be obsolete by 2030, but Neal calls this a lofty aspiration.
“Some underwriting can be automated,” he says. “Other types of underwriting will be very difficult to automate because the data isn’t comprehensive enough to make an automated decision.
“It’s an aspiration in some circles of what could be done, but it’s a very aggressive prediction.”
Instead, he says that in some cases, AI is better used as a guide to make decisions, rather than a tool that makes a decision itself.
“AI can help to identify customers that are low risk and require very little or no intervention,” adds Neal.
“An automated underwriting process can deal with them fast, allowing more time to be spent on the cases without enough data for straight-through processing.”
In the case of a fast-tracked application, a client’s claim will be measured against the data collected on other clients to see if they fit a risk profile appropriate to the business.
If they do fit a profile, the application will be fast-tracked or underwritten automatically without the need for human intervention.
Neal says this increased efficiency, coupled with better insights driven by predictive analytics, will result in a better customer experience.
He says: “If we can make the process faster, an underwriter can be more responsive to situations where customers with a high level of risk need some extra handling.
“Maybe you can learn enough about them to get a better idea of their risk profile, which might mean you can still offer them a policy.”
Non-technical insurance workers can contribute to predictive analytics
The global shortage of data scientists is nothing new – in the US alone, LinkedIn found 150,000 are needed – but perhaps predictive analytics software could help alleviate the problem.
If an insurance company can’t hire enough data scientists to look for insights within their data, automated machine learning software can allow non-technical personnel to test hypotheses without a data scientist present.
Neal says that removing the technical barrier to using machine learning technology allows people within a business with different expertise to contribute to improving predictive models.
“We embed the AI software with the data science techniques that we found are fairly standardised and could be repeated. That allows non-data scientists to make very reliable and accurate models.”
“The knowledge that claims personnel have about claims is often greater than the data scientists themselves, so they can bring new insights into making predictive models, making them more accurate.
“The best insights you get from data are when data scientists and other insurance specialists can collaborate on a problem, and automated machine learning brings those other specialists closer to the level the data scientists are on.”
Limits of automated machine learning
The limits of automated machine learning are found in situations where there’s a lack of information and data.
Neal acknowledges this and says a creative or well-informed analyst might try to use substitute data to complete a hypothesis, but in some circumstances the data is simply not there.
He adds: “There are some types of insurance that are highly specialised and have very few or no claims to use as data.
“It’s very hard to find a pattern with a small amount of claims – ideally we want to see hundreds, thousands, maybe tens of thousands of claims to use as data.
“You can get some insights from a small data set, but if you have almost none, machine learning isn’t going to be the right technique.
“There are some traditional actuarial techniques that can work pretty well with very small numbers, and that’s the best you’re going to be able to do.”