About eight years ago we were working with a large North American bank on the development of their first customer-facing chatbot. The aspiration for the chatbot was that it would answer basic account questions and ultimately provide financial advice. While the Natural Language processing elements of the solution would be easy to implement in theory, we had two major obstacles to overcome. First, the customer account (and specifically transaction data) was in different formats and legacy systems. We didn’t have a unified customer profile. Second, the AI engine we had planned to use was unable to track context from one conversation to the next. The dilemma is one that many banks still struggle with: riches of customer information, and relatively efficient handling of customers in their channels, but once a customer hops over to another channel the experience falls apart. The data is often not in a form that can be leveraged by AI algorithms and one channel does not know what the customer just did in another, due to legacy silos. However, we have come a long way. We now have all types of data in the cloud, which provides banks with a whole range of new capabilities.

The recent rise of ChatGPT has highlighted not only how sophisticated chatbots have become (thanks to generative AI), but also the way in which such services can augment and tailor the customer journey at various points along the human-digital spectrum. The GPT models clearly have significant potential to become part of a data-driven, customer conversation that can be balanced with high touch, human engagement. The quality, speed and depth of ChatGPT’s responses have surprised many people and have got industry commentators wondering what this means for the future of customer engagement and the role and nature of search engines.

However, it’s unlikely we’ll be using publicly available Open AI models for our clients. For starters, the cutoff on knowledge for the underlying GPT model was September 2021. We will create instances of a data model (like OpenAI GPT-4) in Azure OpenAI, load client data onto the model and fine tune it, then ensure it is secure, isolated and protected. We then support the client in training the model on its own data. (GPT-4 has demonstrated that such models can be trained much faster.) This will allow us to identify instances of fraud, for example, in real-time. The time to value will be much quicker as the client fine tunes the data and there is reinforcement learning from humans. This – and advances in neural networks - will improve the use cases as they have not been working that well over the last 2-3 years, with bots giving wrong answers. Azure OpenAI (available in the US and Europe) provides up to 99% answer accuracy compared to Q&A Marker’s 85% ‘out of the box’ solution.

• Prioritize the potential. There are many use cases close to the customer, where banks can apply GPT-4 in their business, including operations, EX and CX. Who wouldn’t want a chatbot to make the most of account-opening incentives to switch your savings? Focus on those areas where you can deliver value quickly. Here are a few examples of how we are thinking of applying GPT-4 to create a ChatGPT-like experience for our clients:

• Be aware of the limits. There are still issues around bias and the consistency of AI responses. AI tools are good where 90% accuracy is needed. If you need more, be careful. This is why humans still fly planes and drive cars. The solution is only as good as the data it is trained on. It is not capable of replacing human creativity, intuition, and empathy, but it will augment human activity and help perform tasks more efficiently.
• Modernize! Get your data ready. Data is typically spread across banks in different silos in a variety of formats and not easily found by those who need it most. Those banks who have already completed projects in areas such as RPA or machine learning, and hold their data in the cloud, are well placed to leverage their data with AI services. Where it becomes much more interesting is when GPT-4 and Dall-E 2 (effectively, its visual AI equivalent) are brought together in areas such as advertising, web design and communications.

Given the rapidly changing market dynamics in this area, banks should prioritize those use cases that can deliver value quickly, work within the limits of such technology and prepare to leverage generative AI by moving to the cloud and organizing their data effectively across the business. As another colleague of mine noted recently: “We are seeing an ‘AI-First’ world emerge.”

Incidentally, the North American bank did launch the chatbot service. While it garnered a fair amount of early usage, it largely redirected customers to their website. We now have a great chance to set a new standard leveraging Generative AI. The bar is currently low and the opportunity is high!

To realise the value of generative AI for your bank, find out more about Avanade’s assessment options.