CIBC's agentic AI pilot puts AI FinOps on the bank P&L

CIBC’s pilot matters not because it proves that banks can use AI... They obviously already can. It matters because it shows that the economics of cognitive banking are beginning to emerge.

CIBC’s original AI chat tool, CAI 1.0, costs about $3 per user per month. Its newer agentic workspace, CAI 2.0, costs between $4 and $250, depending on how heavily it is used.

That gap tells us more about the future of enterprise AI than another demonstration of what a model can do.

CAI 2.0 is being tested by about 250 employees across business banking, capital markets, legal, procurement and risk. The executive team is using it as well, including CEO Harry Culham. The system can build pitch decks, compare credit agreements and prepare regulatory reports. It is not just answering questions. It is taking on parts of a job.

CIBC has built a proprietary layer around OpenAI and Anthropic models. It routes each task to the cheapest model capable of doing the work. Anthropic’s Opus, the most expensive model in the mix, is restricted to a smaller group because the cost is higher.

That is how banks will have to think about agentic AI. There will not be one simple licence fee. There will be light users, heavy users and a small number of employees who can consume a large amount of compute.

Jaime Tatis, CIBC’s chief data and AI officer, said one user could consume $3,000 of tokens in a month. That sounds high until the work contributes to $1 million in revenue. The relevant question is not whether the token bill is large. It is whether the work is worth more than the bill.

This is where agentic AI differs from a conventional enterprise assistant. Microsoft Copilot and Google Gemini cost roughly $40 per user per month. CAI 1.0 costs about $3. CAI 2.0 depends on what the user asks the system to do.

A user asking a few questions is cheap. A user asking the system to analyse a group of agreements, prepare a report, check the report and start again after a change in the instructions is not.

CIBC plans to introduce usage limits and charge business units for what they consume. It did not need that level of control with CAI 1.0 because the cost was low enough to treat as overhead. With CAI 2.0, the bank has to manage model choice, access to expensive systems and the value of the work being performed.

That is AI FinOps, and it is not just an IT issue. It belongs with the people running the business.

The immediate use case is internal. CIBC is using AI to help its own employees work faster and take on more complex tasks. The customer opportunity follows from there.

Banks have been promising personalisation for years. In practice, much of it still means placing customers into better segments and recommending products based on those segments.

Agentic AI could make the service more specific.

A small-business customer could receive financing advice based on its cash position, supplier payments, tax obligations and existing credit facilities. A wealth customer could receive an explanation of a portfolio that reflects their knowledge and tolerance for risk. An employee could have the customer’s next step prepared before the conversation begins.

That is the idea behind cognitive banking: a bank that understands the customer’s situation and acts on it, rather than waiting for the customer to navigate a product menu or ask the right question.

CIBC is not there yet. This is a small employee pilot. The bank has not disclosed an average cost for CAI 2.0, any return on investment or the value created so far. The $250 figure applies to heavy users, not everyone in the pilot.

There is also a question about how easily this model can be copied by smaller banks without CIBC’s engineering resources.

The cost should come down. Models are improving, smaller models are becoming more capable and inference is getting cheaper. Tasks that require a frontier model today will increasingly be handled by cheaper models designed for particular types of work. Competition between providers will put more pressure on pricing.

That's going to be important because the economics determine who receives a personalised service.

For now, banks reserve their highest levels of attention for their most valuable customers. If the cost of running an agent keeps falling, more of that service can be delivered digitally and at scale. Hyper-personalisation becomes less of a premium feature and more of a normal expectation.

CIBC’s pilot is therefore a useful early benchmark. Agentic AI can create more value than a chatbot, but it can also create a much larger bill. The bank has to learn to control the cost before it can extend the service.

Today, the system is helping employees. Over time, the same capability could sit behind lending, wealth management, business banking and customer service.

The banks that benefit will be the ones that understand the economics as well as the technology: what each interaction costs, what it produces and when a more capable model is worth paying for.