Scouting for Growth
00:18:56 · Sabine VanderLinden
[00:18:56] S1: So as Asian carry more of the workflow or the task we do every day end to end, I think the question stop being, what does the machine can do and become who is accountable for the work, right? And we know we need that human in the loop. We are in a regulated industry. So we need to know what may go wrong as well. Pricing coverage was a big topic yesterday. Declare true because we were talking about agentic distribution settlement. So this changed someone's life. This changed someone's rules, accountability, responsibility in a regulating framework. So where especially does a human have to stay, right, in that loop? And how do you tell a decision that can be delegated from one which should not be delegated at all?
00:19:46 · Bastiaan de Goei, Sabine VanderLinden
[00:19:46] S2: So I think that, first of all, what is important is that the... When staying with the underwriting example for a moment, the ultimate quotation is the underwriter's responsibility and that person is accountable for that decision. And so the tooling though needs to help the underwriter understand what they are saying yes to and what decisions were being made, what data was being used as input for those decisions. And I see it as something that we are investing in, in order to make that really easy. As an example, and this is a very small example, we... Earlier this year, we released ChatGPT for financial services. This is a tool that is focused on banks... Banking right now. But as a very tactical example, we noticed that analysts, banking analysts, when they were preparing investment presentations and using AI tools, were double checking a lot of the data points and that went into those presentations. [00:20:44] S1: 100%. [00:20:45] S2: And so... Because it needs to be precise. And just like in insurance, banking is a business based [00:20:51] S1: And on last, you know, I do the same. I still check everything because it's about reputation now.
00:20:51 · Sabine VanderLinden, Bastiaan de Goei
[00:20:51] S1: And on last, you know, I do the same. I still check everything because it's about reputation now. [00:20:56] S2: Exactly. It's about reputation and banking, a lot of financial relations are based on trust. And so... But if you double check every single data point that an AI tool provides you, then you've just lost a ton of efficiency. And so as a very practical example, we released as part of ChatGPT for financial services sources by paragraph and that are provided in the interface of the results where you can then just hover over that particular claim or data point and see exactly not just which document that particular claim came from, but also which paragraph with the exact, line where that data point came from. And so this actually practically starts saving an analyst like a ton of time as they are going through these types of workflows. Now expanding that, that specific example across like an entire workflow, an AI tool should be able to give you the steps that they took, the assumptions that it made and the data points that informed those assumptions. So that the underwriter or claims adjuster is empowered with that understanding and can make an informed and competent decision.
00:22:03 · Sabine VanderLinden
[00:22:03] S1: Another conversation I had yesterday, which I thought was interesting is, you know, how do we codify human knowledge? When you look at, for example, the underwriting examples, a lot of experts have knowledge and expertise, right? And therefore, what happens is we need to make sure the system can manage exceptions. Yeah. So when the knowledge is not codified, what would be your recommendation for an organization which starts using, let's say, Astra or Dart to make sure that things, processes we might not know because it is the way we used to do it, it started to be codified so that the system is much better, I'm gonna say, acquainted with the processes we are working around or the underwriting exceptions which we didn't know about before.
00:22:56 · Bastiaan de Goei
[00:22:56] S2: Yeah. So the best way to address this is to start. And so start with engaging with the, with a DOTS agent because it learns over time. And so, every single, DOTS agent is equipped with what they call like a scratch pad, where it essentially accumulates understanding and knowledge of your workflows and will increasingly be able to understand like what your human knowledge and exception behavior, is. Now, not all of those exceptions are going to happen overnight and it'll be exceptionally difficult to basically or laborious rather to like input, all of those exceptions and knowledge into an agent on day one. But if you start working with it, it's already pretty good and it will only get much better over time. And every time that one of those exceptions happens, it will take a note of that and will remember it for next time. And so over time those agents get better because they start understanding you better. And that type of knowledge that sits in people's heads increasingly will then become codified and can become part of the corporate knowledge.
00:24:02 · Sabine VanderLinden
[00:24:02] S1: So let's go into token GPU a little bit. You know, now the part nobody puts on a conference slide, well, this is not free. And the bill scales over time with usage, as we know. And recently, I had to work on a little business case for a group of executives where they said, Sabine, should we should we build our own LLM? Right. Right? And they said, really? And they said, you know, strategy, you know, core competency. So we did the metrics, and then we look at the cost of GPU, tokens, and humans. Right? It's not cheap investment. So how should an insurer think about the cost of a decision rather than the cost of a token? And what's the part of the organizations that are getting real advantage, right, unfair advantage from those who quietly burning budgets, burning tokens, and not getting any outcomes from their investment.