Two banks. Two decisions. One useful lesson for lending executives: modernizing the machinery does not mean handing everyone the controls.
CIBC announced its adoption of Thought Machine’s cloud-native Vault Core on October 29, beginning with a Simplii product. A day earlier, Global News reported that RBC had decided to block third-party personal AI agents from authenticated digital properties, including online banking.
These are separate initiatives, not a joint program. Our read is that they belong in the same executive conversation: how to increase the institution’s ability to act without losing control over who acts on its behalf.
CIBC is changing the product foundation
The Simplii Account is the first client offering built on Vault Core, with a pilot this fall and features such as folders and savings goals. CIBC says its own product and engineering teams will build on the platform, giving them greater control over what they create and how quickly they introduce it.
The scope matters as much as the ambition. The announcement starts with a defined offering; it does not establish a completed bank-wide core migration or independently measured efficiency gains.
For a lending technology leader, the useful question is not whether to copy CIBC’s vendor decision. It is about whether your architecture lets teams improve a bounded product or workflow without making every change depend on replacing everything.
RBC is drawing an access boundary
RBC told Global News it had decided to block third-party personal AI agents from direct access to authenticated RBC digital properties. That wording is narrower, and more useful, than “RBC bans AI.”
The same report says CIBC does not support personal AI agents directly opening or accessing accounts. A bank can pursue new infrastructure while restricting an outside tool’s access; those positions are not inherently contradictory.
Our recommendation is to separate capability from authority. A tool’s ability to navigate an interface says nothing about whether your institution has approved that actor, that action or that recovery process.
Policy is the moat when software can enforce it
An executive policy that says “use AI responsibly” is not an operating model. Translate the intention into controls that developers, underwriters, and reviewers can actually apply.
- Identity: Distinguish a customer, an employee, a service account, and a delegated software agent.
- Authority: Separate permission to read information from permission to change records or initiate a transaction.
- Evidence: Record the instruction, policy version, source material, result and approval.
- Recovery: Define who can stop execution, reverse an eligible action, and resolve an exception.
These are proposed design priorities, not a description of either bank’s implementation. The practical test is whether your controls produce evidence during normal operations rather than require a reconstruction after something breaks.
Modernize one workflow without creating a new blind spot
Start with an operational bottleneck whose boundaries you can explain in one meeting. Missing-page checks, document classification, or exception routing are more useful pilot definitions than “enterprise AI transformation.”
Build on existing infrastructure where it remains fit for purpose; replace components when the business case supports it. Require the pilot to preserve authoritative records, use approved interfaces, and stop safely when inputs or permissions change.
Use the digital-twinning lens carefully. An institution-specific model of policies, roles and workflow states can help test proposed actions against your operating rules; it is not a reason to pool customer records indiscriminately.
Set a baseline before deployment. Measure cycle time, manual intervention, exception accuracy and the completeness of action histories; then ask risk and operations to review the same evidence.
The board does not need another debate about whether innovation and control can coexist. It needs a program in which neither one is an excuse for neglecting the other.
FAQs
Is CIBC replacing its entire core banking system?
The announcement does not say that. It describes adopting Vault Core to build products and experiences, starting with the Simplii Account pilot.
Has RBC banned all AI use?
No. The reported decision covers third-party personal AI agents’ direct access to authenticated RBC digital properties, including online banking; it should not be expanded into a claim about all internal AI or automation.
Are the two banks working together on this?
The cited CIBC platform announcement and RBC access decision do not establish a joint program. Combining them is our editorial analysis of modernization and control, not a claim of partnership.
What should a mid-market lender do first?
Choose one workflow, identify its accountable owner, and define the tool’s allowed actions before procurement. Establish a baseline and require demonstrable auditability, exception handling, and rollback before expanding the pilot.
Does a digital twin mean pooling borrower data?
Not in the design approach recommended here. Model your institution’s rules and workflow states; introduce any borrower information only with a defined purpose, appropriate permissions, and controls.