CIBC weighs costs as select employees test more powerful AI tool

CIBC has also tested open-source AI models, including from Canadian startup Cohere, though it isn’t using them at enterprise scale yet. Photo by Peter J. Thompson/Postmedia

Article content

Article content

The Toronto-based lender first launched an internal AI tool called CAI just more than a year ago. The new version uses agentic AI, which can plan and execute complex, multistep tasks from start to finish. Beyond analyzing documents and generating blocks of text, such tools use more powerful reasoning to build finished products such as spreadsheets, pitch decks and fully formatted reports.

Article content

Article content

Article content

The advantages are compelling but don’t come cheap. The original CAI tool, which CIBC built in house, costs about $3 per user per month on average, compared with roughly $40 a month for some fixed-license enterprise AI assistants such as Microsoft Corp.’s Copilot or Google’s Gemini, according to the bank.

Article content

Article content

The cost climbs for more advanced agentic tools, including products from OpenAI and Anthropic PBC. CIBC users testing CAI 2.0 are racking up $4 to $250 monthly, depending on how heavily they use it.

Article content

“It’s too powerful to have 50,000 people just using it without control,” Jaime Tatis, CIBC’s chief data and artificial intelligence officer, said in an interview. “We’re going to have a very strategic approach to how we roll this out gradually.”

Article content

Large banks have moved quickly to put AI tools in employees’ hands, but many firms are now grappling with how to manage token expenses — the computing costs generated by widespread use of the technology. JPMorgan Chase & Co. chief financial officer Jeremy Barnum recently said the bank is urging employees to avoid deploying the most sophisticated models for basic tasks such as summarizing reports.

Article content

Article content

CIBC started its latest pilot with about 50 users and has expanded it to roughly 250, drawing from areas including business banking, capital markets, legal, procurement and risk. The entire executive team, including chief executive Harry Culham, is testing the new model, and a handful of wealth advisers are also experimenting with it.

Article content

An average of 20,000 people use the initial tool every day. The bank doesn’t track the prompts employees enter, which it says are encrypted, but it does monitor usage patterns, token spending and employee proficiency, ranking users from novice to “professor.” That data helped CIBC identify a network of about 1,000 internal AI champions, who have given feedback on new features and use cases.

Article content

“CAI 2.0 is very different, because it’s agentic,” Tatis said, noting that it could take half an hour of compute time to generate a presentation an employee would be able to show a manager. “So for that, you need a little bit more expertise and know-how on what you need to do, how you want to do it and understanding how the platform works.”

Article content

Managing Costs

Article content

As generative AI tools advanced rapidly, CIBC chose to build its own employee platform rather than relying entirely on an outside vendor. Along with product designers, a team that started with about five tech developers and grew to roughly 20 built CAI 1.0 over about a year and a half, according to the bank. The 2.0 version was built in about three months by Tatis’s team using the existing foundation.