BNP Paribas is moving agentic artificial intelligence further into its banking operations under a five-year Google Cloud agreement, while retaining restrictions around which data and workloads can enter public cloud infrastructure. The French banking group will expand its access to Google Cloud infrastructure, Gemini models, and Gemini Enterprise as it develops AI agents for work inside Corporate & Institutional Banking.

Initial uses include supporting the preparation of corporate credit memos, while the bank is also examining agentic applications across sales, trading, research, and structuring. Gemini models will be incorporated into LLM@CIB, BNP Paribas’s internal generative AI assistant, which is already available to more than 65,000 employees, making the agreement an extension of an existing AI programme rather than the beginning of one.

That distinction is important because the new systems are moving closer to workflows in which software retrieves information, reasons over it, and performs a sequence of actions rather than merely generating text for an employee. BNP Paribas is pairing that expansion with controls covering data classification, access, model choice, and infrastructure, reflecting the operational consequences of letting AI systems interact with regulated banking processes.

Although the technology industry increasingly describes autonomous agents as the next stage of enterprise automation, BNP Paribas is not treating every banking workload as appropriate for public cloud infrastructure. Some highly sensitive customer information and critical operations will remain outside it, while existing security and data-governance rules will continue to determine which systems can use external cloud and model services.

Agents move closer to banking workflows

Preparing a corporate credit memo gives a useful indication of where banks see a practical boundary between generative AI and more autonomous software. The task involves gathering information from different sources, following established processes, assembling analysis, and preparing material for human review, which creates more scope for an agent than a single question-and-answer interaction.

BNP Paribas says agents will be authenticated individually and given access only to resources needed for their assigned tasks, while connections between agents and the bank’s information systems will be monitored and controlled. Once AI software can call tools, query databases, move information between systems, or initiate parts of a workflow, identity management and audit trails become as important as the underlying model.

The bank already has a smaller example in production through Nickel, its payment-account business, where Gemini underpins Nickel Assist. Around 200 customer advisers use that system to retrieve information from procedures and operating guidelines, providing an established reference point before the technology moves into more demanding institutional-banking workflows.

As deployments expand, the practical question shifts from whether a model can generate a plausible answer to which actions it is authorised to perform and how those actions can be reconstructed later. Credit, trading, research, and structuring activities all sit inside environments where an incorrect output can affect decisions, records, customers, or regulatory obligations rather than simply creating a poor draft.

Multi-cloud remains part of the control model

BNP Paribas says the agreement will reinforce an existing multi-cloud and multi-model strategy rather than consolidating its AI estate around one supplier. The bank intends to select models, infrastructure, and control levels according to individual use cases, weighing security and quality alongside cost and expected value.

That approach reflects constraints that have accompanied financial-sector cloud adoption for years. Banks have moved increasing numbers of workloads onto hyperscale infrastructure, but supervisors have simultaneously focused more closely on operational resilience, outsourcing, supplier concentration, subcontracting, and the consequences of relying on a small number of technology providers for critical functions.

The EU’s Digital Operational Resilience Act has applied since January 2025, placing ICT risk management and third-party dependencies alongside resilience testing and incident management. Financial institutions are consequently expected to understand not only the software they use but the supplier chain beneath it, including how services can be recovered, replaced, or exited when necessary.

BNP Paribas has already framed infrastructure as a prerequisite for AI rather than a secondary technical question. That is consistent with another large European bank’s decision to put infrastructure and data modernisation beneath a multibillion-euro AI programme, reinforcing the extent to which the harder work sits below the visible assistant.

Deployment changes the operating model

BNP Paribas is also extending AI training across the group, ranging from general literacy to specialist development of AI agents. That organisational work accompanies the technical deployment because automating a multi-stage process changes how responsibilities are divided between people and machines, while exceptions require explicit escalation routes and controls must remain effective as models and surrounding applications change.

The bank operates across 64 countries and employs more than 180,000 people, including over 146,000 in Europe, so even a narrowly defined use case can involve substantial operating change when deployed across business units. Standardising access and governance while allowing different teams to choose appropriate models becomes a management problem as much as a machine-learning one.

Google Cloud gains a significant reference customer for Gemini Enterprise in one of Europe’s most heavily regulated industries, although BNP Paribas is deliberately stopping short of treating public cloud and external AI models as appropriate for every system. Sensitive information, critical operations, permissions, and supplier dependencies remain part of the architecture rather than obstacles assumed to disappear once the technology performs well.

As agents move into credit analysis and other institutional-banking workflows, the useful measure will therefore be the work they are permitted to complete safely rather than simply the number deployed. BNP Paribas has put identity, data classification, infrastructure choice, and human oversight around the programme before widening it, which is likely to prove more consequential than the availability of another model endpoint.