BNP Paribas Signs 5-Year Deal with Google Cloud for AI Agents

BNP Paribas signed a five-year partnership with Google Cloud on Thursday (24), granting the French bank access to Google's public cloud infrastructure, Gemini Enterprise, and Gemini models. The initial target is the Corporate & Institutional Banking (CIB) division, where the bank intends to deploy AI agents to prepare corporate credit memos, according to the joint statement from both companies.

The agreement includes the integration of Gemini models into LLM@CIB, the internal generative AI assistant of the division, which, according to the statement, is already available to over 65,000 employees. After credit memos, the list of applications includes sales, trading, research, and structuring.

Public Cloud with a Defined Boundary

The most relevant detail for those designing architecture in banking came from the press conference, not the statement. Marc Camus, CIO of BNP Paribas Group, told reporters that the contract does not represent a mass migration. "Signing this agreement with Google does not mean we will get rid of our on-premise infrastructure," he stated, according to Reuters.

Camus provided a concrete example of what will remain off: medical information maintained by the group's insurance operations, classified as highly confidential. For the bank, Google's technology is appropriate for less sensitive tasks such as invoice processing and data management. Google, in turn, stated that BNP retains ownership of the data, which will not be used to train, tune, or refine the company's models.

The statement summarizes the rationale in one sentence: the bank’s approach "begins with the business need," choosing technology on a case-by-case basis within the institution's security and governance framework.

Multi-Model, Not Exclusive

BNP has not switched providers. According to Reuters, the bank will continue to use open-source models and maintain its partnership with French Mistral AI, distributing providers by application. In 2025, the group had already renewed a multi-year agreement with IBM Cloud for dedicated infrastructure. The resulting design consists of three layers: critical data on-premise or in dedicated cloud, European models where sovereignty matters, and American hyperscalers where scale and cost are considerations.

This arrangement responds to pressures that are primarily European. The DORA regulation, effective from January 2025, requires banks in the European Union to map and manage dependency on critical technology providers, and reliance on a single hyperscaler becomes a supervisory item. A bank that publicly declares which data do not go to the cloud is also communicating with the supervisor.

Where Credit Memos Are Currently Written

The choice of credit memos as the first use case is not arbitrary. It is a lengthy, standardized document subjected to multiple layers of review, produced by junior analysts and support teams. In major European banks, part of this preparation and review work is handled at service centers outside headquarters. BNP maintains such operations in Portugal and India, and it is in these hubs that the automation of credit drafts is likely to first bring productivity gains.

In the United States, the competition is similar, albeit with a different provider backing it. During the July earnings season, CNBC highlighted job cuts associated with AI at JPMorgan and identified the bank and Goldman Sachs as the primary beneficiaries of the technology on Wall Street. BNP’s move places a major European bank on the same timeline, with one difference: the boundary of what will not go to the cloud has been clearly stated.

In Brazil, where BNP Paribas operates in wholesale banking, the agreement does not signal any immediate change. The significant takeaway for local banks and consultancies is method. At this year's Febraban Tech, the CEOs of Itaú, Bradesco, Nubank, Santander, BTG, and Caixa presented scaling goals for AI and advocated that critical decisions remain under human oversight. What BNP released this week represents the other side of the equation: the list of what stays out.

For integrators and consultancies selling AI projects to banks, the message is clear. The five-year contract ensures Google has a seat at the table but not necessarily the volume. The volume will depend on how many use cases pass through the sensitivity screening that Camus described, which the bank will conduct on a case-by-case basis.