Driving Business Impact with Generative AI: A Strategic Approach

We sat down with CIBC’s Head of Enterprise AI Platforms and Solutions, Chris Patterson, to discuss the challenges and opportunities with Generative AI (GenAI).

Q: What is GenAI and how does it differ from traditional artificial intelligence (AI)?

GenAI represents a cutting-edge advancement in AI, distinguished by its capability to create new content from input instructions. Unlike traditional AI, which primarily focuses on pattern recognition and data analysis, GenAI employs sophisticated machine learning techniques to generate text, images, code, and even music. This technology has the potential to revolutionize various industries by automating creative processes, enhancing productivity, and driving innovation.

For example, tools like OpenAI's large language models (LLMs) can write reports, summarize vast amounts of content, extract information from unstructured materials, and transform unstructured data into structured content. This showcases the immense versatility and power of GenAI. However, this technology also introduces concerns about privacy, misinformation, ethical implications, and regulatory obligations.

Q: What’s CIBC’s approach to Generative AI and how is it different from other emerging technologies?

CIBC has a long history of innovation! We have Canada’s #1 rated mobile banking app and have a history of embracing emerging technologies. For example, CIBC was the first large bank in Canada to deploy 24-hour automated cash dispensers, automated telephone banking, mobile banking and mobile eDeposits. We view GenAI advancements as an opportunity to transform how we work and interact with our clients.

We’re already using AI in key functions across our bank. Existing implementations include sophisticated risk, fraud, and information security models that can detect fraud and enable our team to help prevent losses for our clients. What makes this moment in time feel unique though is just the sudden, exponential pace of AI advancement. I vividly recall my computer science teacher in 1999 explaining his efforts to build an application that could parse and understand natural language. He believed it was an unattainable goal, and creating a chatbot that could pass the Turing test seemed impossible. Yet, seemingly out of nowhere, ChatGPT emerged and changed everything. We now have AI systems that can interpret intent from a statement and respond with highly convincing answers, putting the Turing test within reach. Within a year, we saw the development of detailed images and soon after, full-motion videos. AI can now produce music at an advanced level, and there seems to be no end in sight for its progress.

So yes, the speed of AI feels different, but our approach is consistent. We continue to explore how we can thoughtfully, responsibly leverage these quickly-evolving tools to enhance the banking experience for our clients, and enhance productivity for our teams – like we always have.

Q: What unique challenges have you faced trying to implement solutions using these large language models?

CIBC approaches emerging technology thoughtfully and with a client-focus, prioritizing that risks and regulatory requirements are thoroughly addressed. Our journey to being #1 in mobile banking is a testament to this; we were first to market and it has been years of meticulous design, development, and deployment to create Canada’s leading mobile banking platform.

GenAI has sparked public interest in a different way than mobile banking. As a result, we’re seeing widespread eagerness from our clients and teams to leverage this technology to simplify daily tasks. The unique challenge has been that eagerness, and that this technology is so new that we needed a more clear and unified understanding of how to effectively implement and use these models.

With so many teams across our bank and the many creative ways to use this technology, we needed a singular view on how people were exploring GenAI – what tools they were looking at and how they wanted to use them. At the same time, we needed to make sure we were collaborating where it made sense to make our implementations more efficient. To manage this, we established a new AI Oversight Committee to centralize and oversee all AI-related activities, ensuring that all risk control groups had visibility into ongoing projects, shared resources and knowledge, and to make sure we were all informed and aligned rather than working in siloes.

Our first significant deployment of GenAI was in our Contact Centres. We introduced a Knowledge Central chatbot (basically the directory where all our process documentation and policies are stored) to answer common client inquiries, with the goal of reducing the time and effort our team members spend escalating calls. Initially, this seemed a straightforward application of a retrieval-augment-generate (RAG) system. However, we learned early on that not all our data was fit for the needs of the models we were using, due to formatting or various factors, and addressing that was definitely a unique challenge. In a business where response time and accuracy are critical, it meant scaling back the content to a more manageable volume for the pilot release. Another interesting challenge was finding some LLMs struggled to deliver specific responses with 100% accuracy when required. This technology limitation meant we had to scale back the chatbot’s scope to exclude use cases involving strict verbatim instructions.

Overall, it’s been an incredibly positive learning experience and has already informed our approaches to other use cases across the bank. The chatbot is showing promising results that are continually improving, team engagement with the system is high, and most importantly we are seeing this enhance our clients’ banking experience as a result.

Q: Can you elaborate on the various pilot programs you have been running?

Beyond our Knowledge Central pilot, we have been running several other large-scale Generative AI pilot programs:

• GitHub Copilot for developers: GitHub Copilot is a tool designed specifically for software development teams and was one of the first vendor systems to leverage LLMs. We launched our initial pilot in 2023 and are still in a pilot phase, with the latest version including a chat feature that acts as a pair programmer. It’s been well received and we’re seeing a tangible increase in productivity among developers. With continued use and familiarity, we anticipate further time savings on average and plan to deploy this software at scale in the near future.
• Office Copilot for knowledge workers: We have also been conducting a pilot of Office Copilot and have seen project managers in particular, experiencing the most immediate benefits from using Teams Copilot, which enhanced their note-taking and post-meeting activities. Similar to its GitHub counterpart, we’ve already seen tangible benefits to overall productivity improvement with this software, although we’re still exploring what types of tasks and roles are best suited to maximize these tools as they’re currently offered.

CIBC-AI: we are also implementing our own productivity tool which functions similarly to ChatGPT but with additional features. The goal was to create a tool that provides direct access to a language model and supports more CIBC task-specific assistance. Early feedback has been positive. CIBC AI enables more advanced use cases, and unlike a third-party tool, we can track every possible data point produced by the pilot participants, allowing for deeper analysis of usage, use cases, costs, and quantitative metrics.

While many systems are still in the pilot phase, we anticipate several notable deployments across the enterprise in the latter part of 2024 and into 2025. These pilots are crucial in helping us understand the potential benefits and challenges of integrating AI tools into our workflows, while at the same time ensuring we make informed decisions for future implementations.

Q: What specific factors are important for deciding between vendor solutions and bespoke (custom-built) solutions for AI technologies?

This is an interesting question as we have considered several key factors making these decisions across our bank.

• Integration with existing workflows: how well the software integrates with the existing workflows of employees. What are the complexities to be considered in this integration?
• Ease of use and intuitive prompting: how intuitive the software is to use. A powerful tool won’t be useful for people if they can’t figure out how to use it.
• Specific use cases and vendor expertise: compelling vendor solutions have emerged from firms with deep knowledge of a specific use case or business, using AI to enhance their existing software rather than building new software around an LLM. Having an AI expert involved in these evaluations is critical to distinguish between superficial and genuinely valuable solutions.
• Managing risk is a key consideration: We need to ensure that we are protecting all our stakeholders as we deploy new AI solutions.  Some suppliers have thought through the risk management side of their solution and can make contractual commitments backed by auditable artifacts. Many have not. CIBC strives to ensure that all AI is managed to the same standard, whether developed internally or from a third party.
• Cost is always a factor: Vendor solutions often come with licensing fees and ongoing subscription costs, which need to be weighed against the potential productivity gains. On the other hand, bespoke solutions may involve significant upfront development costs but can be tailored precisely to the organization's needs, potentially offering better long-term value. Additionally, the cost of maintaining and updating custom-built solutions should be considered, as well as the potential need for specialized talent to manage these systems.

There are so many considerations on a case-by-case basis as we continue to assess and learn in these early days of Generative AI, but this gives you clearer picture of the main things we’re trying to balance.