Building a safe, scalable enterprise AI platform

First and foremost, we offer a compelling chat-based experience where users can upload documents, presentations, or basically any type of file and have the AI support them with a specific task or series of tasks. This is very similar to any standard chat-based LLM tool, and we offer the ability to change the language model at any time, so the user has more control over which model they use for which kind of task. The business problem here is the simplest and most universal one — the hours knowledge workers spend summarizing, drafting and reformatting — and we wanted to give that time back. Culturally, this was our front door: the place where tens of thousands of employees first met AI inside a governed environment and started to imagine what else it could do. Table stakes, but the table on which everything else is set.

As users have spent time with the product and become familiar with the capabilities of LLMs, they can start to create their own personal agents, which act as micro-automations for the regular, repeat tasks they perform day to day. This interface is essentially a teaching tool, and once taught, the AI can repeat any text-based workflow over and over again with a high degree of accuracy. The human is always in the loop here, and the AI acts more as an augmentation to the employee than as something independent. The business problem is repetitive cognitive load — the small, frequent tasks that quietly consume a workday — and the cultural shift is that employees stop being users and start being builders, reimagining their own workflow rather than waiting for someone else to do it. Basic Chat created the fluency; Personal Agents turned that fluency into reusable work.

Employees who have created personal agents are given the opportunity to share them with others. As we have many common roles in the company, it is natural that we’d want to foster a culture of sharing. Not everyone is at the same level, but we want to make sure we’re maximizing the productivity opportunities. To that effect, we created a governed-by-design approach to sharing agents across the organization, with great success. The business value is direct: one well-designed agent can lift the productivity of hundreds of colleagues doing the same job. Culturally, the Marketplace is what makes individual reinvention scalable — a team member’s reimagined workflow quietly becomes the new normal for everyone in that role. Personal Agents let us build once; the Marketplace let us spread.

Knowledge retrieval is another common use case that comes up often when discussing LLMs, and it was one of the more common questions that surfaced in the early days of our platform launch. To address the need, we created a generic solution that allows a SharePoint site administrator to create a chat experience over their data. There are many products out there that do this as well, but we’ve found that some sites are complex and can’t be handled in a generic fashion. The business problem is the time tax of finding the right answer inside a sprawling content estate; the cultural unlock is that Smart Libraries put the power to stand up a domain expert in the hands of the site owners closest to the content. One of the complex cases the generic solution couldn’t fully handle is addressed by the next use case.

This is a well-commoditized solution at this point, but in a tightly regulated company such as ours, allowing a third party access to potentially sensitive meeting data and audio is a non-starter. To that end, we created our own Teams-based meeting notetaker bot that can join a scheduled meeting and will share notes after the meeting wraps up. This has been streamlined into a meaningful interface to make the process as simple as possible for our users. The business value is straightforward — time back, better recall, fewer missed actions — and just as important, we now own a transcription and summarization capability that is safe for our environment and reusable elsewhere. Culturally, this is a team-wide habit shift: meetings get reimagined as collaborative work rather than note-taking exercises.

All of this in a single calendar year. With the right platform and engineering mindset, you can see what is possible, and where re-use of services becomes paramount to scaling well and scaling fast. The throughline across every one of these use cases is a deliberate trifecta: a real business problem worth solving, a technology platform that lets us solve it safely and repeatedly, and a culture that encourages team members to reimagine their own workflow and share the result. Take any one of those three away and the others stop compounding. Read about recent CIBC AI use cases (PDF, 195 KB) Opens a new window..

We want to move on to data, as this is probably one of the most important, yet most overlooked, aspects of deploying generative AI at scale across a large organization. As a platform owner, it has been amazing to roll out something of this scale and magnitude and watch it take on a life of its own across the organization. But what does success look like in this case? What are the Key Performance Indicators, and how should they be measured? Simply knowing your product is being used is not good enough. We need to know everything: the who, the what, the why, the when, the how, and most important of all, how much will this cost to run?

Every single one of our products comes with deep analytics as a day-one requirement, with no compromise. With everything running on a pure consumption model, running blind is not an option. But while cost is clearly an important factor, it is far from the only one that matters. Knowing who uses each product, and how, is also useful intelligence when searching for opportunities to design more specialized solutions. The generic solution works well under most situations, but when we have the opportunity to take an easy 10% productivity lift from the generic solution and customize it to drive 50%, that's an easy decision to make when the scale exists and we can measure for direct benefits.

To that effect, we have built a data product containing over two years of deep intelligence on our users, from pilot data to production data. We have used this data not only to calculate daily usage trends, cost tracking and feature uses, but also to derive a meaningful AI Fluency score for each employee in the organization who uses our software. Fluency matters because adoption is not the outcome — outcomes are the outcome. A team that is fluent does more than open the tool; they redesign work around it, and that is what eventually shows up in client experience, productivity and growth. 

Fluency is measured as follows.

• Novice: New to AI tools, with minimal experience and a basic understanding.
• Intermediate: Comfortable with AI basics and using the tools more regularly.
• Practitioner: A consistent, effective AI user applying solutions to real tasks.
• Expert (key target): Highly skilled, maximizing AI features and driving innovation.
• Professor: An AI leader and mentor, setting best practices and guiding others.

Not every employee moves at the same pace, nor do they adopt AI into their workday equally. One of the most interesting, yet equally frustrating, discoveries has been the complete lack of any discernible pattern to an employee's AI fluency journey. An employee’s level, role, age, gender or location have not revealed any common patterns. The only common pattern that exists is that across all lines of business, we see a very similar bell curve shape to fluency. Our number one user of the platform, measured simply by consumption, was a temporary co-op student. The number two is a senior director. We have professors across all lines of business and at all possible levels of the organization, which means one key thing: this technology has the opportunity to support literally anyone in the organization, all the way up to and including the CEO.

How we get everyone using it equally is the hardest part of our role by far. The technology is easy in comparison to changing people’s behaviour.

To that end, we have designed an AI People Change Champion Network, built specifically to scale learnings and knowledge across the organization. We have fostered a network of over 1,000 of our most advanced employees and turned that into an engine of change by empowering them to teach others in their respective businesses. The Champion Network is the cultural infrastructure that makes everything else repeatable — the part of the operating model that turns a reimagined workflow in one corner of the bank into a reimagined workflow in another, without the platform team having to sit in the middle of every conversation. This is how you scale, and it was powered entirely by data.

A platform that solves real business problems, a technology foundation that lets us build once and reuse many times, and a culture of team members reimagining their own work and teaching each other to do the same — that is what one year of compounding looks like for us. It is also the operating model we intend to keep running. The use cases will keep arriving, the fluency curve will keep moving, and the Champion Network will keep turning individual breakthroughs into shared practice. That is the platform we set out to build, and the one we are continuing to build every day.