A new story from AI in Business

AI in Business
|

Automatic TRANSCRIPT

Do I want to keep my existing system in place? Might I want to go with that less modern statistical techniques? Well, I want to get to go with something more established to more traditional. And this might be fine for me, because this much which used my risks for that. So it's quite important to think about the impact of risks and also how what this means really for my investment project. Got it. And to your point, plenty of times that will be simpler solutions where the upside of getting that extra 5% of performance on why is just not worth the maintenance cost and the potential downside of an algorithm steering somewhere versus maybe the rule based system that works for now. So it's always going to be a balance for many applications. Some have to be ML, some don't. So when it comes to quantifying what that downside would be, I think as a business, let's say I'm a big manufacturing firm and I'm about to spend untold millions to start predicting maintenance failures of my machines. I've got a bunch of drill machines press machines, whatever I got in my manufacturing warehouses and factories. I'm going to put on sensors. I'm going to start training algorithms. So I'm going to make a big investment here. And it's going to affect my business if it fails and I start producing flawed products that's horrendous. So big consequences this might be an insurable thing. How do you go about quantifying where the failure points are? Because part of it is, you know, a business person doesn't know, oh, here's all the ways AI could fail, right? They didn't go to school for this. So how do you walk leaders through maybe the considerations to think about for where the failure points are and how risky they are? Well, what's your method? Yes, sure. I mean, the first question is really does the AI work as expected? This really means to the predictions the AI makes. Are they really living up to my expectations in terms of being correct or being not too far off from the ground truth? I think that's really the first question in the first risk that's the risk of predictive performance. So once I know my AI solution works, there might be other risks I need to watch out for. For example, if I'm more in a consumer sensitive area, then discrimination fairness related questions might become a topic. If I'm just using machine learning model to do predictive maintenance on my machines, of course, fairness and discrimination is not really an issue. But if I ultimately use machine learning to do great, assessments and ultimately decisions about our.

Coming up next