In February 2025 I wrote about responsible AI in financial planning and some of the challenges which firms must contend with in order to responsibly integrate AI into their business. We’re now 15 months on from that article and, as is the case where any emergent technology reaches the tipping point into mainstream use, the landscape of AI use in financial planning has changed.
The key factors I mentioned in my article, hallucinations, cheating, the hunger for data and energy and the exploitation of real people, all remain relevant and appropriate, however as AI continues to grow and develop, our approach must also develop.
I recently had the opportunity to present to the Insurance Institute of Perth and Dundee on the subject of Responsible AI use, in which I discussed the importance of five key pillars of responsible AI use:
– Safety, security and robustness
– Appropriate transparency and explainability
– Fairness
– Accountability and Governance
– Contestability and redress
These are all important but they feel clinical, cold and “compliancy” without the discussion which took place in the session around what each of these mean to real people.
Those who are at the forefront of AI involvement in financial planning may say that the latest models have done away with the hallucinations which were present in previous versions. I’ve certainly found that the more recent models hallucinate less, but that doesn’t mean that they don’t hallucinate at all. LLMs are built to sound confident, so if they have incorrect information they will tell you that it is fact, rather than an inference, unless you have specifically revoked their permission to infer.
Those of us who are getting to grips with AI and how it might be beneficial in financial planning are getting better at prompting, but even the best prompt can sometimes be largely disregarded by mainstream LLMs. I don’t mean ignoring the small details, I mean completely left at the traffic-lights stuff. The models will get better, and we will get better at using them. But that doesn’t mean that they won’t sometimes infer something, confidently, when they are supposed to be presenting just the facts.
In training we talk about the stages of competence, from unconscious incompetence to unconscious competence. A new trainee on the cusp of conscious competence may think they know the answer and be very confident in expressing that they know the answer, but due to their lack of experience they may not notice finer details which stand out to a well-trained mind. I find working with an AI can be helpful, asking it to summarise things for me or to check my work, but there are times when it confidently tells me something, repeatedly, and I have to ask it to check its response because through experience I know that it is incorrect.
I was recently “forced” to buy a new laptop. Not by a nice sales person who saw me and my wallet from a mile off, but because I am going away for a week and my five-year-old laptop finally gave up. Despite weighing the same as a small car, I loved that laptop. It got me through lockdown. The metal case was battered and bruised; one of the USB ports is broken, to the point where I keep catching it on things. I’d had the entire motherboard replaced at one point. It just worked. Most of the time. Except the trackpad. And the microphone. And the battery.
Whilst I would rather like a dinky little reconditioned one to write on (if anyone wants to donate one I’ll happily take it), I needed a new workhorse for presentations, working in client offices and so on. Something with a big enough screen to take my “office” with me, anywhere. (There’s a reason the Unburdened logo is a backpack after all…)
The one I purchased, from a nice chap at Costco, has a built in NPU (Neural Processing Unit) which supposedly supports AI based tasks. As I write this I have spent the last 45 minutes performing various AI related tasks on the machine to try and test this capability (short of generating images and videos which I try to avoid because of the vast amounts of energy they use and the issue of exploiting other peoples’ work). It would seem that no matter how hard I push it, I can’t actually get this extra little brain to kick in.
AI has become something exciting. Something people will pay money for. Extra money. If it has AI it must be better. Laptops with AI. Phones with AI. Financial Planning with AI. Just because something has AI capability, is generated by AI or is in some way connected to AI doesn’t make it better. It makes it a bit more complicated. With complication comes a risk of failure.
In a recent presentation at an industry conference, it was highlighted that whilst AI might make mistakes, so do people. We all make mistakes. But when people make mistakes we learn from them. We store that information to help us the next time we perform the task. Sometimes we make the same mistake again. If we have the correct training partners we may eventually learn to get it right. AI is doing this too, particularly if we are actively training a model on what “good” looks like, but we need to make sure that learning is genuine, not just on the surface.
Where models are trained on a limited data set, we have to be aware that they might make more mistakes, very confidently. If what they are working on is something about which they don’t have much data they might infer an answer without the necessary information. Making sure that the AI doesn’t make something up to fill the gaps is part of good prompting, but they still have a tendency to go their own way when they want to if we’re not careful.
I’m currently working on what might be quite an ambitious project which needs some form of machine learning capability to make it work. Not everything we do needs AI. Using AI requires careful thought about why we are using it, what is the purpose, what are we trying to achieve, how does AI fit in to the overall project, what are the potential disadvantages of using it and what might go wrong? There’s a space for AI on my project but I need to be sure that we’re doing this thoughtfully, paying due regard to the potential limitations and the possibility that it might cause more problems than it solves.
Until next time.
Alan