The Rise of The Robots: AI large language models and the future of insurance policy wordings

Grant Pilkington, Aon

“A man’s got to know his limitations.”

In the 1973 film Magnum Force, the protagonist, Harry Callahan, says “A man’s got to know his limitations.” My understanding of AI is limited, so I thought it was time to follow Inspector Callahan’s advice, fill in my knowledge gaps and investigate the hype about AI large language models (LLM) in the context of insurance policy wordings.

And hype there certainly is. Hardly a day goes by without insurance predictions about magazines AI offering disrupting and revolutionizing the industry and its potential for automation and innovation. Devising optimal insurance products is my raison d’etre, so I was intrigued. Can LLMs be harnessed to accurately analyse and develop policy wordings? This article considers some of the opportunities and risks and whether insurance professionals should fear the rise of the robots

Testing the Chatbots

To investigate, I conducted a ‘trial run’ to assess whether publicly available LLMs can in fact draft an insurance contract. I gave an open-source LLM some parameters and asked it to draft a commercial property wording. Within a few seconds, the LLM produced a document which certainly had the appearance of an insurance contract. It had that title, but it wasn’t an insurance contract – at least, not one I would recommend to a client. A closer careful inspection showed that:

• The coverage provided under the AI-devised clauses was inadequate and did not provide the comprehensive ‘all-risks’ protection required under modern commercial property policies.

• Essential policy terms and coverages were missing.

• The operative policy language was elaborate and sprinkled with old-fashioned legalese, e.g. whereas, thereof, thereto, the same, aforesaid, herein, hereinbefore, hereinafter etc. To me, these have no place in a modern insurance contract.

• Several clauses were not as clear or certain as they ought to be, carrying high risk of a dispute as to the contract’s application in practice.

Considering the concerns in the output, as an insurance contract, the document was not something that an underwriter or broker should want to use.

I should say upfront that the trial run I conducted tested the draftsmanship of a single publicly available LLM – it was me dabbling and not a scientific exercise, so the results are not necessarily representative. No doubt I could have fed the LLM with better commands or tested a wider range of models.

However, overall, the trial run underlined that if the current technology is used to prepare insurance contracts, it is essential for the output to be thoroughly checked by a professional with the appropriate qualifications and experience before it is used in practice.

The need for professional oversight seems supported by two legal casesi widely reported in mid-2023. The cases provided a salutary lesson regarding the limitations of LLMs and the importance of verifying their output, particularly where the AI is delivering a professional product in high stakes scenarios. (See Box 1).

‘Garbage in, garbage out’

AI is only as good as the data coming in. In computer science, ‘garbage in, garbage out’ (GIGO) is the concept that, in any system, the quality of output is determined by the quality of the input. A strength of LLMs is that they access a huge online dataset from across the internet. The corollary is an important weakness, because there is plenty of biased, inaccurate, and false information available online and the current ability of open source LLMs to filter it out seems questionable.

The problem of ‘traditional’ contract language

Output from ChatGPT depends on feeding the language model with available relevant data, so the contracts produced are likely to include traditional contract language. Unfortunately, traditional contract language is awash with dysfunction: chaotic verb structures, overly long and information packed sentences, circular definitions and redundancy.

In most organisations, contract drafting is unscientific, inevitably conducted by consulting and duplicating earlier contracts of questionable quality or relevance, perhaps with a bit of well intentioned (but often misguided) ingenuity to reflect the new transaction. It produces a pathology of ‘copy and paste’ drafting, tainted by inertia, where people copy and rely on existing contracts, assuming that traditional contract language has worked before and will therefore work again.

In the ‘hurly burly’ world of commercial risks, where the stakes are often high enough for the parties to pick a fight over nuances, this is pretty flimsy wisdom.

Customization of manuscript policy wordings

Customization is crucial in ‘manuscript’ commercial insurance contracts where policy wordings are bespoke, as each business has different requirements and risks. Relying solely on a LLM might overlook the unique risk profile of a specific business, producing a ‘one-size-fits all’ policy wording that fails to adequately protect the policyholder.

Commercial clients run complex cross-border operations and experience complicated losses and claims. The issues can be knotty or novel or arise piecemeal. Policy wordings should routinely undergo revision to keep pace with solution for major insurers. From a London and clients’ evolving risk profile and may need adjusting depending on the loss record or unexpected uninsured exposures that were not anticipated but are sometimes a fact of life. Important factors sit outside the ‘four corners’ of the insurance policy

Certain terms and conditions will also need updating to account for new legislation, case-law and regulatory changes. Given where open source technology is currently at, relying exclusively on LLMs to ensure good product governance in real-time seems risky.

It is possible that AI could help insurance professionals review lengthy contracts quickly and identify whether certain ‘boilerplate’ or onerous clauses are included in them. However, that’s not really ‘AI’; it’s largely an advanced search function which is already available in contract review tools which many lawyers use. In that situation, the ‘AI’ – such that it is – is not undertaking the professional analysis or advising. Rather, it’s essentially an information gathering tool.

Conceivably, an AI bot could assist with or produce standardized cover, perhaps based on model ISO coverage forms and endorsements commonly used in North America markets, or Institute forms for marine business classes.

However, for large complex risks, standardization is generally not a marketable solution for major insurers. From a London and Lloyd’s market perspective, brand differentiation stands on a reputation for deep experience, product innovation and creative solutions for tricky risks. At present, LLMs seem unable to generate suitable bespoke wordings without human intervention.

Current conclusions

While LLMs have transformative potential for the insurance industry, current limitations and potential pitfalls seem likely to slow their adoption in the development of policy wordings. Insurance professionals should ensure appropriate oversight and not slavishly follow the output of available technology

For now, at least, the role of products counsel and wordings specialists remains vital for complex and higher-value risks. The immediate application of AI is probably to finetune what already exists or identify inconsistencies, rather than replace human expertise entirely.

That said, the algorithms and technology are rapidly improving and prototypes in the labs of top providers – as yet, unreleased – are very likely more powerful and interesting than what is public. If the technology improves enough to analyse the available data and integrate relevant laws and regulations, it could be leveraged and scaled. And then the robots can really come out to play.