Embracing the Mundane in AI — the need for Specialised AI in Financial Services

Embracing the Mundane in AI — the Need for Specialised AI in Financial Services

Why we need more specialised AI which is more accurate for Financial Services tasks, not bigger and more capable headline AI systems.

Would you hire Einstein to staff your call centre? (Image created using Dall-E 3)

In the rapidly-evolving landscape of artificial intelligence (AI), it’s easy to get caught up in the whirlwind of excitement surrounding the latest advancements. The AI community is again buzzing with discussions and papers devoted often to building bigger and more complex language models (LLMs) and broad API toolsets. It paints a picture of a future where AI serves as a personal assistant, ready to tackle any challenge alongside us. This is useful as a personal co-pilot and I would like one! However, this pursuit of a broad, capable AI agent misses what I think businesses need, especially within the realms of insurance, pension administration, and banking.

Whether it’s answering customer inquiries about a product, guiding someone through a process, or providing the financial guidance, the need is to solve that narrow mundane, yet crucial, task. This is at the core of customer management and financial guidance.

The truth is, that ‘big’ AI systems are not well suited to these applications out of the box. For a start, these large, all-encompassing systems are expensive and often slow to run. More problematic though, is that these big and broadly capable AI agents are often very flaky on real-world tasks that the business actually cares about. Accuracy is often low and is always rather uncertain unless you have done a huge amount of work to pin down and test across huge example datasets. No tech or project person wants to do this testing.

A somewhat better approach is techniques such as RAG (Retrieval Augmented Generation) for AI in which the AI uses a compendium of content to help craft its answers. This definitely works better the firms that are running POCs are mostly using these techniques. But in our experience, RAG is useful but not sufficient. There is a temptation to add more content to the compendium and this tends to just make the answers less reliable. How you curate the compendium of content makes a huge difference in the quality and accuracy of answers. So we cannot get away from the fact that subject matter expertise and content knowledge matter when building these systems.

Our approach goes a step further than this: a network of specialist AI agents. Our platform is built on the principle that specialised components, each focusing on a narrow task, can achieve significantly higher accuracy than their generalist counterparts. These expert agents can handle specific inquiries or do specific jobs with precision. We can link them together seamlessly to create more comprehensive customer journeys.

Opting for a network of specialised AI agents approach offers several advantages. Most importantly, we can test more comprehensively and deliver a significantly higher accuracy on tasks that businesses care about. Specialised agents can also be much more transparent, moving away from the “black box” nature of larger AI systems. And this network of specialists means you can build or buy components and plug-in specialist AI agents into your broader agent pipeline where you don’t have the internal skills to build or maintain that piece.

The trade-off with this network of specialists approach is that generality is lost. The AI system is now much narrower in capability, but much deeper on accuracy within that capability space. This type of AI will not be able to answer anything and everything, as you may wish for a personal co-pilot. We believe this is actually good thing for real use cases where businesses want to plug in AI to help customers through specific journeys. To draw a parallel, in a call centre, the operations manager does not hire a team of Einsteins to staff the team. Such a choice would be both overkill and misaligned with the job to be done.

The main takeaway from all of this is to say that I believe that the future of useful AI in financial services is not about chasing the latest developments or models. Rather it is about focusing on the mundane tasks and doing the boring nitty-gritty work of creating specialist components that deliver a really high accuracy that businesses can trust.

Many in the AI space, in particular in research, may have quite different view on this.

Find out more — https://engagesmarter.ai


Embracing the Mundane in AI — the need for Specialised AI in Financial Services was originally published in Chatbots Life on Medium, where people are continuing the conversation by highlighting and responding to this story.