What is AI Good For?
“We need an AI strategy. What should it be?” is a question I hear more and more at the moment. It’s a slightly odd question; here’s a solution, find me its problem. I have a hammer. What can I hit with it?
Often what is touted as AI, isn’t. It’s a good, old automation algorithm; predictable and pre-determined. Or it’s the data-analysis that informs the parameters of an algorithm. I heard someone describing fuzzy logic as AI matching in a dashboard discussion the other day. It isn’t. It’s some extra conditional statements in the code. It’s a perfectly respectable approach, which doesn’t need the glamour of AI to make it more compelling.
So, what is AI then? Well, the common description is when we use computers to perform tasks that we associate with needing human intelligence, like vacuuming the room, or switching on the lights when asked, or driving a car. Which seems a little vague and unhelpful to me. I think it’s more helpful to think about this in the context of Machine Learning, which is a subset of AI, and is concerned with getting models and solutions by looking for patterns in massive sets of data.
In particular, when we apply it to language, we can summarise complex bodies of knowledge, understand them, and apply that understanding to answer questions. One approach (Retrieval-Augmented Generation, RAG) enables you to define which bodies of knowledge you use, such as your scheme documentation. This can give some very useful superpowers, and reduce the risk of getting nonsense based on rubbish in the wider internet.
So, AI can be a powerful approach to solving many problems. There are some examples I’ve seen that I thought were cool.
Data validation. I’ve used the anomaly detection techniques to look for problems in data. It’s a similar approach to the ones banks use to detect fraud. Essentially, you’re scanning your admin data to look for things that seem odd, patterns that are different to most of the data. It’s particularly effective with investment data, which is complex and voluminous and it can flag up problems long before you’re aware of them with traditional techniques.
Back office administration. I’ve seen some great examples of how AI can turbo-charge back office administration. Recording meetings, transcribing them, summarising them and writing minutes. I’ve also seen it used to answer technical questions from administrators about benefit rules, by analysing the scheme’s documentation and other resources.
Identifying vulnerable customers. I’ve been discussing how we can use AI to review the recordings of customer service interactions to check that a representative hasn’t missed signs of a vulnerable customer.
In all of these cases, AI is used as an additional tool to support humans. It flags things, gives you a head start, speeds things up. But it doesn’t replace human judgement. And, so far, it doesn’t give members front-line, unmoderated support. I think trustees would be nervous of an AI bot cheerfully telling a member that the solution to an adequate retirement income is to manage their life expectancy with a one-way ticket to Switzerland.
So it helps to understand AI, and see it for what it is, a useful technology that can make us more effective and efficient. But the key truth about using technology still applies. Be clear about the problem you’re trying to solve before you decide on its solution.