Our Approach to agentic AI: Why the human is the driver

We’ve spent the last two years building CIBC AI (CAI), our internal AI platform, from the ground up — the use cases that shaped it, the proprietary data strategy underneath it, the AI fluency scoring we use to measure adoption, and the Champion Network that helps it spread. The thread running through all of it is a simple idea: owning your own platform and your own data gives you something you can’t buy off the shelf. That foundation is exactly what makes the next chapter possible. So let’s talk about agentic AI.

Walk the floor at any AI conference right now and you’ll hear the same argument: what actually counts as an “agent,” and where the line sits between automation, an assistant and something genuinely autonomous. We’re going to skip that debate. It has been running for two years, and it hasn’t made a single deployment any easier. The question that matters isn’t what an agent is. It’s what it takes to put one to work.

Making some type of agent is actually pretty easy with today’s technology. You can stand one up before lunch. Deploying it into production is another matter entirely, once you account for scale, governance and cost. The running joke is that everyone has an agent in a pilot and almost no one has one truly running in production. Closing that gap is the whole job.

Start with scale. The industry has rushed to deploy autonomous agents across organizations, with mixed results. Take software delivery. A multitude of solutions have emerged to “agent-ify” that work, for developers, business analysts, project managers, and so on. They work well when the roles are generic and common across the industry. But how do they scale when your organization runs on hundreds of roles and job functions that are unique to it?

Generic agents fail at the edges, because the edges are where your people’s craft lives: the judgment, the context and the institutional knowledge that make a role yours and not the industry’s. That’s why we built the personal agents and the marketplace. We now have a two-year-old dataset containing tens of thousands of agent definitions, usage metrics, and sharing and adoption statistics. Our team can use that to focus on the most popular and most useful information when we design our agentic products. You won’t get this level of intelligence and data unless you build your own software.

We’ll be evolving these agent definitions into skills and plugin packs. Imagine a world where you’re a new product owner joining the organization, and you get access to a product with a series of plugins designed specifically to help you work, containing all the best intelligence and knowledge from every product owner across the bank. We can level the playing field, scale up fluency across the board, and rapidly deploy new skills, knowledge, tools and capabilities to everyone in those common roles. This is how craft knowledge gets democratized — the experienced practitioner and the colleague who just started operate from the same shared foundation. It’s a very powerful concept. But how do you deploy it effectively?

We’re not convinced that we’re ready for agents of this nature to run fully autonomously, without human intervention. Without a proper control plane, these agents are operating blind, and that doesn’t work in a regulated environment. But this isn’t only about regulation — it’s a deliberate cultural choice and, we’d argue, a competitive advantage. Keeping our team members in the driver’s seat produces better business outcomes: human judgment shapes the output, accountability stays clear, and the work is something the bank can actually stand behind.

To that effect, we’ve evolved CAI into CAI 2.0, which introduces a new agentic harness. It gives team members highly sophisticated capabilities, but it leaves them firmly in the driver’s seat. This is in production pilot now, and we’re already seeing some truly astonishing results. Deep, structured research can be done at scale. Meaningful analysis can be run against that material to produce structured spreadsheets and research reports. Those reports can be turned into presentations in seconds — work that used to take weeks.

This level of capability is expensive. We’ve already seen the era of the free lunch ending. Most products have moved toward consumption billing, and the cost to run AI has outpaced the fixed-price model. CAI 2.0 is no different.

In this version, to help manage costs, we don’t let the user choose the language model. We decide for them, based on the task the AI is working on and our team’s expertise in knowing which models are best suited for which type of activity. Our model garden will need to scale up, and we’re already exploring how to best use open-source models to find the right balance between quality and efficiency.

It leads right back to change management. We are building products that help team members operate agents, review their work and build new skills as the technology evolves. The skill we are scaling across the bank is not prompting. It is supervision.

That supervision layer also gives us a way to measure adoption, quality, confidence and impact as we shape the workforce of the future.

This matters especially for developers. As AI changes how software, data and digital products are built, CIBC’s application developer framework creates room for technical talent to grow through both people leadership and advanced individual contributor paths. Read about the application developer framework (PDF, 120 KB) Opens a new window..

Not every strong technologist needs to become a people manager. Deep specialization, mentorship and engineering leadership are all essential as AI becomes part of the everyday development toolkit.

We are supporting that shift through practical upskilling. More than 450 CIBC developers have participated in AI training workshops in recent years, building skills in areas such as AI development and multi-agent systems.

While much of the industry races toward full autonomy, we are building the governance, skills and cultural muscle first. The autonomy we eventually grant must be something our people, clients and regulators can trust. In the near future, every team member becomes a manager. Just not only of people.