I’m afraid it can do that. And that. And that. AI is a truly transformative technology. The public recognises this. ChatGPT had 1.8bn UK visits in the first eight months of 2025, up from 368m in the same period of 2024. And ChatGPT is just one AI tool. There has never been a new technology so quickly and so widely adopted.
Everyone is scrambling to find the best way to utilise it in their own areas of business. Pensions are no exception. We’re still learning the use cases, but there are a number of key areas where it is already making a difference.
The first is finding patterns in structured data. There is nothing new about computers uncovering such patterns. As long ago as 1992, data analysts identified that supermarket sales of beer and nappies might be correlated. This surprising finding was explained by hassled young dads rushing to the supermarket for an essential, and feeling the need to treat themselves. Computing power now means more wide-ranging correlations can be found.
In pensions, AI can be used to interrogate administration data. I’ve been told how administrators are using it to identify what members really need help with, so these administration services can be improved. This should make them more personalised and more effective. It’s worth noting, though, that I’ve heard no mention yet of replacing administrators with agentic AI. Just because you have a hammer doesn’t mean that every problem involves nails.
AI is good at interrogating structured data, but where it really comes into its own is working on much less structured data.
Pension schemes routinely have missing beneficiaries. With more than £31bn in unclaimed entitlements and more than 3m inactive or lost pension pots, reuniting pension members with their full retirement benefits is one of the biggest challenges facing the industry today. There are already highly capable tracing agencies, but there is an increasingly large residue even after they have done their work.
Using a combination of AI tools and human ingenuity, Zedra has partnered with Heka to achieve striking results in locating the previously unlocatable. Data is the foundation of every pension’s exercise. Improving it will improve so many outcomes.
A lot of the focus on AI has been on how it might transform human knowledge. Perhaps it will. But the lowest hanging fruit on the tree of knowledge are those where we already have the techniques but have been unable to use them economically until now. Some of the regular actuarial reporting can be produced more cheaply. AI can now do a lot of the heavy lifting that junior actuarial staff would have had to do in the past at a cost that is uneconomic for small schemes.
As a result, AI should be able to commoditise this and other similar services. Of course, this would require consultancies to cannibalise some of their own fees. Will they be willing to accept this? Will they decide they cannot afford to let their rivals do this? Time will tell.
Finally, AI is already fundamentally altering the relationship between administrators and members, almost without administrators knowing it. That’s because AI is being used not by the pensions industry, but by the consumers of its output. Members are plugging the announcements and communications that they are being sent into AI, translating them from gobbledygook into plain English.
It’s a shame it’s necessary for members to do this, but I report rather than condone. If it helps members understand their benefits, so much the better. There is a risk that this will make communicators lazier. That’s something we’ll need to watch.
We are seeing members use AI when they’re putting together their correspondence with scheme administrators. This is a good thing if it helps members make their points more effectively, so that administrators can address their concerns more quickly.
This is, however, a double-edged sword. We are also seeing members plug administrators’ responses directly into AI for a further response, without any intervening thought on the members’ part. This means that members can disengage from accepting valid points that administrators might be making by outsourcing their thinking to a machine instructed to be truculent. This will make disputes more intractable unless administrators can stage an intervention between the member and the AI.
There are many, many things that AI can do for us, but there are some things that it can’t. It can explain things for us, but it can’t understand things for us. I’m sorry, I’m afraid it can’t do that.