Artificial Intelligence promises to introduce tremendous efficiency savings and generate greater business insights, but there are pitfalls as well as opportunities. Xceedance group chairman and CEO Arun Balakrishnan looks at how the insurance industry can work with regulators to use AI ethically and safely.

There is a huge buzz around artificial intelligence in the insurance industry even though it has been around for a few years. So why is it now such a hot topic and why is our sector, notoriously slow to adopt new technology, suddenly keen to jump aboard?

Banks and other financial markets were quicker to automate through AI, using it mostly for transactions. It was relatively straightforward for machine learning to take on these number-based tasks, but requirements in the insurance industry are more complex, such as gaining an understanding of policy wordings in natural and legal language.

AI offered us limited advantages, but when large language models such as Chat GPT – launched at the end of 2022 – were introduced, AI became relevant to our industry. These models work by associating a word with those that come before and after it, allowing AI to understand sentences in a meaningful way.

AI can now be deployed in the insurance world to trawl through huge amounts of data to gain insights, understand claims notes and write documents. Insurers have come to realise that they can have a junior underwriting assistant working 24 hours a day without complaining – it just won’t get you a cup of coffee.

Insurance professionals also saw how Chat GPT was being enthusiastically adopted by the public and realised if their teenager was using AI to write essays, then insurance professionals could surely learn how to use it. Consumers have adopted this technology more quickly than businesses and now the executives and boards of insurance companies don’t want to be left behind.

This recent eagerness to learn about and embrace machine learning and generative AI has prompted regulators to address the potential misuse of AI in underwriting and claims adjudication decisions.

Large language models were trained on content built up on the internet over the last 30 years – content full of hidden biases. An example of how AI picked up on our biases is a series of recent images created by AI of corporate roles using the faces of animals – the CEO is a lion; the COO is a wolf; the sales manager a fox. AI merely took our own perceptions and reflected them back at us, and this is the type of ethical danger that regulators need to worry about as AI is rolled out in insurance.

If underwriting decisions or claims adjudication decisions are based on AI alone, and made in a rush without proper oversight, there is potential for discrimination and privacy breaches.

As I see it, there are two ways to regulate. The first is policing everyone, with models being put through a discrimination test before they are used to make decisions. But this is inefficient, hard to enforce and is not scalable. The other way is self-governance; holding companies accountable for the consequences of their decisions, which is the approach Colorado and some other states have taken.

In May 2024, Colorado became the first state to have comprehensive AI legislation, stating that developers and deployers must exercise “reasonable care to protect consumers from any known or reasonably foreseen risks of algorithmic discrimination”. Developers have to share certain information with deployers, including the types of data used to train the system.

The Colorado Act, which comes into effect in February 2025, places the onus on insurers on how they use AI. It is a promising development as there has been too much of a rush for the industry to develop and deploy AI before guidelines are in place. It’s likely that other states will follow Colorado’s example, and many developed economies, such as the EU and Australia, are discussing AI regulations.

There’s no need to rush into deploying AI in every part of the lifecycle of a policy. While it is straightforward to use AI to make back office administrative functions more efficient, in other areas such as risk selection and pricing, and claims adjudication a more cautious approach is required. If AI models are used to support decision making in these areas, they should run in parallel with human decision making.

We haven’t scratched the surface of what AI can accomplish. When the internet entered general use in 1993-4, we couldn’t envisage the business models that would follow or the transformation of how businesses are run. There’s no reason for companies to fear they will be left behind in the rush to adopt AI when we don’t know where this technology will lead us.

Meanwhile, if consumers, businesses and governments are to retain confidence in the trustworthiness of the insurance industry, it is essential that guardrails are developed to ensure AI is introduced in a safe and ethical way.