‘You’re right’ is the costliest thing AI can say about money
MoneySense writer Vickram Agarwal shares surprising ways responsibly using AI has helped him make day-to-day money decisions.
Advertisement
MoneySense writer Vickram Agarwal shares surprising ways responsibly using AI has helped him make day-to-day money decisions.
A few months ago, I wrote about trust, money and AI, and offered some thoughts on how we can use it without outsourcing our financial future to it.
I use AI a fair bit in my day to day, and I work with several startups building in the space. I sit somewhere between skepticism and evangelism, which has led me to focus on responsible use. It feels like the appropriate middle ground, especially when it comes to money.
I prefer not to think of it as the answer to everything, but I also don’t indulge the fear of terrible things happening as a consequence. I see it as a fantastic resource that creates more space to think for myself rather than the opposite, which is where a lot of the concerns lie.
Over the past several months, I’ve been using my GPT in some surprisingly practical money situations. The goal wasn’t to have it make a financial decision for me, but to better understand what was happening, ask more relevant questions and avoid a potentially expensive financial outcome.
I am not handy at all.
When I found a leak in our home, my reaction after the initial panic was fear of how much it was going to cost.
Instead of picking up the phone and calling someone to inspect the situation, I took some pictures and uploaded them to my GPT. I wasn’t sure what I was asking it to do. I just figured I might get a little more insight into what I was dealing with.
With a little back and forth, it helped me understand what I was looking at and what I needed to do next: talk to my neighbours, get their insurance information, contact my own insurer, connect with the property management company and meticulously document everything. A lot of this may seem obvious, but it isn’t if you’ve never had to deal with it before.
AI’s utility was most apparent in helping me make sense of everyone else’s version of the event. Everyone involved was speaking from their own perspective, and there were times when those perspectives didn’t line up. That misalignment is where things can start becoming expensive. Having a tool to help me synthesize everything I was hearing made it easier to figure out all the moving parts and the plan of action that followed.
The fix ended up involving multiple parties, a little frustration and a lot less anxiety than it could have, but most importantly, it cost me zero dollars.
If you know me, you know I love live music, and concerts are an important part of my life experience. I’ll happily pay for a great concert, but there’s a line between paying for an experience you value and getting fleeced trying to make it happen.
Dynamic pricing, scalpers, resale marketplaces and multi-stage presales have made buying tickets an expensive ordeal.
So the night before I log into a presale, I give my GPT everything I know about the venue, the upcoming event and my budget for the show. I then ask it to help me work through the likely seating and pricing based on historical data, the prevailing online buzz and other venues across the globe.
It builds me a rough layout of the venue, predicts pricing for different sections and ranks them for me based on value and alignment with my budget.
It isn’t perfect. But it gets me close enough that when the queue opens, I’m not spending fifteen minutes trying to figure out whether the $240 seats are actually better than the $80 ones. I know where I’m looking and what I’m willing to spend before I make the purchase.
I’m not asking AI whether I should go to the concert. I’ve already answered that question for myself and set the guardrails around it. What I’m doing is using it to inform my process, so I can stay within budget and keep myself from making a bad decision while I’m caught up in the rush of trying to get tickets.
For months, there was a strange crackling noise coming from my car speakers. I’d done some basic troubleshooting by switching cables, changing the source and testing Bluetooth connectivity, and I had a sense of what might be going on.
A few days before I took the car in for its regular service appointment, I reached out to the dealership to talk through the issue. They wanted almost $200 just to diagnose the problem, and the part that might need replacing, based on my assessment, was less than $100.
I understood why they couldn’t just take my word for the diagnosis and swap the part, thereby assuming the liability of it not working. But I also didn’t want to spend close to $300 to find out whether something under $100 would fix it.
So I asked my GPT what else I should be thinking about. It asked for some basic information about the car and told me the issue was common for that model, and that there was an open voluntary recall and a possible software update that could fix it. All of it would cost me nothing.
That was information I didn’t previously have, and those few hundred dollars stayed in my pocket.
A 2026 Ipsos survey for TD found that 41% of Canadians believe AI can help them make better financial decisions, but 71% still say they trust human intelligence more than AI.
Two in five see the potential, and most of us still lean on human judgment when it counts. That is comforting in theory. In practice, “use it, but don’t trust it completely” is more of a feeling than a strategy.
Here’s where I got tested.
I was working through a significant financial decision with a professional, and our numbers weren’t lining up.
My thinking was clear, but I had AI review their analysis before I responded. It raised some flags about their approach and helped me draft a response, which was well thought out and better than anything I could have come up with by myself, or so I thought.
The professional came back with a more detailed breakdown, and this time around my GPT sided with them. I didn’t agree, so I continued to push back. Interestingly, AI then built a rationale and found stats to support my bias. I wasn’t looking for confirmation or to be right. I wanted to arrive at a decision that was well thought out and based on data, but that isn’t what was happening.
The professional came back yet again and showed that one of my arguments rested on a premise that was wrong. It turns out that premise was my GPT strengthening the rightness of my argument rather than telling me flat out that I was incorrect.
Researchers have named this algorithmic sycophancy, the tendency of AI models to agree with users and validate their position, even when it’s wrong.
That’s when I realised how important it is to know the difference between using AI because I know I’m right and using AI because I want to be right. I was fortunate to be working with an expert who had guardrails and a framework, because the alternative could have been very expensive.
None of these situations involved AI telling me what to do with my money, which is a line I’m not comfortable crossing. For the most part, AI helped me see something I may have missed, or added perspective I would otherwise have had no access to. I decided what mattered, but as my last story shows, even arm’s length engagement can be a slippery slope.
It is hard to build a framework around something that changes without warning, and harder still to offer generic advice when the circumstances that can lead you down a negative path are so nuanced.
What I do know now is that I can’t let myself slip into autopilot. When I feel truly passionate about something, that is when I need to turn my critical thinking up a notch, and the best thing to do is probably to stop listening to my GPT and pick up the phone.
Share this article Share on Facebook Share on Twitter Share on Linkedin Share on Reddit Share on Email