PENSIONS
Is there room for AI in the pensions guidance space?
Andy Coles MCIPPdip, deputy head of pensions and benefits, Dorset Healthcare University National Health Service (NHS) Foundation Trust, deliberates whether artificial intelligence (AI) could be used to provide guidance to members of the NHS Pension Scheme (NHSPS)
I ’ ve been teaching an AI tool all about the NHSPS. Not because I think it’s going to make my work life easier but because I want to see how close it’s getting to being able to replace me. This isn’t purely selfish concern about losing my job, I promise. The appetite for personalised information and guidance about the NHSPS has increased greatly over the last few years, as the scheme itself and the pensions landscape around it have grown more complicated. Providing that guidance (and never advice, of course) usually falls to individual NHS employers. It’s a very time-consuming duty and fulfilling that demand has consequences. A good pension scheme is a great retention tool but only if its members understand the benefits they get from it. Rapid rules changes, made worse by the McCloud judgment, have led to members opting out or disengaging from the scheme. Retention is a big issue in the NHS, but few employers can afford to employ large pension teams to spend time with individuals, helping them to understand their options and benefits. AI that could act like a pensions officer, answering members’ questions in plain language, would be a big step in meeting this growing need for guidance. The technology You’ve probably heard of, or encountered, large language model (LLM) AIs such
"A good pension scheme is a great retention tool but only if its members understand the benefits they get from it"
as ChatGPT and Google’s Bard before. But the AI world moves fast and since then, OpenAI (the creator of ChatGPT and arguably the clear leader in the field) has introduced the ability for paying subscribers to create their own custom chat AIs. These can be given specific instructions on how to behave, the tone to take and they can also be trained on the user’s own data. I’ve created one of these custom bots and fed it large amounts of publicly available information about the scheme – no confidential data of course, just guides, fact sheets and material from the NHSPS website. ChatGPT already knows a lot of this information from its base training. The vanilla version of ChatGPT can answer broad questions about the scheme with a fair amount of accuracy. The difference is that custom AIs can be told to prioritise and reference specific sources over the wide library of general data pulled from the internet. Once created, these instances of ChatGPT can be kept private (which is what I’ve done) or, in the latest development from OpenAI, shared on a kind of AI app store for other people to use. There’s already one publicly available NHSPS ChatGPT
on there. I tested that alongside my own version and found the results to be very similar. "Artificial intelligence that could act like a pensions officer, answering members’ questions in plain language, would be a big step in meeting this growing need for guidance" The current state The results are both a comfort and a concern. LLMs are excellent at replicating human language patterns. This fluency comes at the expense of expertise though, even for the custom bots. Trained on vast libraries of human speech and text, they’re excellent at predicting likely candidates for the next word or phrase in a sentence and
| Professional in Payroll, Pensions and Reward | April 2024 | Issue 99 60
Made with FlippingBook - Online magazine maker