BNP Paribas and Google Cloud announced a new five-year partnership on September 24, 2026, expanding the French bank’s access to Google’s AI-optimized infrastructure, Gemini models, and Gemini Enterprise. The headline detail isn’t the AI rollout itself, though. It’s what the bank is explicitly keeping out of it: sensitive customer data, medical records tied to its insurance arm, and critical operations will stay on BNP Paribas’s own on-premises infrastructure rather than move to Google’s public cloud.
That split, agentic AI for internal work paired with a hard line on where certain data physically lives, is becoming the template for how large regulated banks are approaching generative AI in 2026. BNP Paribas, Europe’s largest bank by assets, just gave that template one of its highest-profile test cases yet.
What BNP Paribas and Google Cloud Actually Signed
The agreement, announced jointly by both companies, expands BNP Paribas’s existing relationship with Google Cloud into a broader five-year commitment. Under the deal, the bank gets wider access to Google Cloud’s AI-optimized infrastructure alongside Gemini models and Gemini Enterprise, Google’s platform for building and deploying enterprise AI agents.
Google Cloud CEO Thomas Kurian framed the announcement around agentic AI specifically. “Through this five-year partnership, BNP Paribas is combining its banking leadership with Google Cloud’s secure infrastructure and Gemini Enterprise to deploy intelligent agents within platforms across the bank,” Kurian said in the announcement, published via PR Newswire.
A Google Cloud spokesperson added more detail on the mechanics: “Through this partnership, BNP Paribas will have access to Gemini Enterprise to build, evaluate and deploy purpose-built agents tailored to specific business needs.” That’s a meaningfully different pitch from a generic chatbot rollout. Gemini Enterprise is built for banks and other regulated industries to assemble narrow, task-specific agents rather than hand employees a single general-purpose assistant.
No deal value was disclosed in the announcement, and neither company named an individual signatory from BNP Paribas’s side. What both sides did specify clearly is scope: this is an infrastructure and AI-platform agreement layered on top of BNP Paribas’s existing cloud footprint, not a wholesale migration.
The Five Jobs BNP Paribas Wants AI Agents to Do
The partnership is meant to support AI agents across five specific functions inside the bank: corporate credit memos, sales, trading, research, and structuring. That list matters because it tells you where BNP Paribas sees the fastest return on agentic AI, and it’s almost entirely on the corporate and institutional banking side rather than retail.
Credit memo drafting is a natural target. It’s document-heavy, follows fairly standard templates, and consumes analyst hours that banks would rather redirect toward judgment calls than formatting. Research and structuring work similarly, both involve synthesizing large volumes of market and client data into something a human still has to sign off on, which is exactly the kind of task where an AI agent can draft a first pass and a banker reviews it rather than starts from a blank page.
Sales and trading are the more interesting additions. Neither function tends to be first in line for automation at large banks, given the regulatory scrutiny around trade execution and client communications. Their inclusion here suggests BNP Paribas is treating this less as a pilot program and more as an attempt to push agentic AI into revenue-generating desks, not just back-office support functions.
Why “Public Cloud” Still Makes Banks Nervous
The part of this deal that’s drawing the most attention isn’t the AI agents, it’s what BNP Paribas refused to put on them. According to the announcement, Google Cloud use inside the bank will follow BNP Paribas’s existing security and data-governance requirements, which determine what data and workloads are allowed to run in a public-cloud environment at all. Certain categories of data will not be stored in a public-cloud environment, full stop.
Medical information tied to BNP Paribas’s insurance operations was singled out as an example of data that will stay off public cloud entirely. Insurance arms of large banking groups routinely hold health records tied to life and disability policies, and that category of data carries some of the strictest handling requirements under European law of anything a financial institution touches.
More broadly, BNP Paribas is keeping sensitive customer data and critical operations on its own on-premises infrastructure rather than migrating them to Google’s servers. That’s not a rejection of the cloud, it’s a segmentation strategy: agentic AI and general infrastructure workloads go to Google Cloud, while the highest-sensitivity data categories stay under the bank’s direct physical control.
What Stays On-Premises vs. What Moves to Google Cloud
The bank hasn’t published a full data-classification map, but the announcement is specific enough to sketch the broad split between what’s in scope for Google Cloud and what isn’t.
| Category | Handling Under the Deal | Notes |
|---|---|---|
| Corporate credit memos, sales, trading, research, structuring | AI agents via Gemini Enterprise on Google Cloud | Named explicitly as target workloads for agentic AI |
| General AI-optimized infrastructure | Google Cloud | Expanded access under the five-year term |
| Insurance medical information | Kept off public cloud | Cited by the bank as an example of highly confidential data |
| Other sensitive customer data | Kept off public cloud | Governed by BNP Paribas’s existing security requirements |
| Critical operations | Remain on-premises | Bank retains direct infrastructure control |
What that split shows is a bank drawing a line between AI-agent workloads it’s comfortable running on shared infrastructure and data categories it isn’t willing to move at all, regardless of how the AI layer is architected on top.
Data Ownership and the “We Won’t Train On Your Data” Promise
Google addressed the other half of the trust equation directly: data ownership. According to the announcement, BNP Paribas will retain ownership of its data, and that data will not be used to train, tune, or refine Google’s models. That commitment has become close to standard language in enterprise AI deals over the past two years, but it’s worth stating plainly given how much scrutiny bank-cloud partnerships get from regulators and shareholders alike.
BNP Paribas also said its AI agents will be authenticated, restricted to the resources required for their specific tasks, and monitored. That’s a fairly standard least-privilege access model applied to AI agents rather than human users or service accounts, and it reflects a pattern showing up across enterprise agent deployments generally: treat an autonomous agent’s permissions the same way you’d treat a contractor’s, narrow by default, logged, and revocable.
Access controls like these have become more prominent industry-wide as agentic AI moves from chat interfaces into systems that can take real actions. Google’s own cloud security documentation and its Gemini Enterprise product page both emphasize scoped agent permissions and audit logging as core selling points for regulated customers, which lines up with what BNP Paribas described.
Google Cloud’s Financial Services Push
BNP Paribas isn’t Google Cloud’s first attempt to court a heavily regulated financial institution with a hybrid data model, and it won’t be the last. Google has spent much of 2026 positioning Gemini Enterprise specifically for industries where “we won’t touch your sensitive data” is a prerequisite for any conversation at all, not a nice-to-have. Banking, insurance, and healthcare share that requirement more than almost any other sector.
The pitch to financial institutions has consistently leaned on three pillars: keep the customer’s data ownership intact, don’t train shared models on customer data, and let the customer set the boundary on what workloads are eligible for public cloud in the first place. BNP Paribas’s deal follows that pattern closely enough that it reads less like a custom-negotiated exception and more like Google’s standard enterprise financial-services offer, applied at scale to one of Europe’s largest banking groups.
That consistency matters for competitive positioning. A bank considering a similar deal with Google Cloud today has a public reference case to point to, both for the AI capability and for the data-governance carve-outs that made the deal palatable to BNP Paribas’s compliance and risk teams.
How Google Cloud, AWS, and Microsoft Azure Compare on Bank-Grade AI
The three major hyperscalers have each built a distinct enterprise-agent product aimed at exactly this kind of regulated-industry deployment. Google’s is Gemini Enterprise, layered on top of its broader Google Cloud infrastructure. Microsoft’s competing push runs through Azure AI Foundry, and Amazon’s runs through Bedrock and its AgentCore runtime.
| Provider | Enterprise Agent Platform | Underlying Models | Financial-Sector Positioning |
|---|---|---|---|
| Google Cloud | Gemini Enterprise | Gemini model family | Data ownership retained by customer, no training on customer data, hybrid public/on-prem split |
| Microsoft Azure | Azure AI Foundry | Multiple, including OpenAI models | Deep existing footprint in enterprise IT and Microsoft 365 workflows |
| Amazon Web Services | Bedrock / AgentCore | Multiple, including Anthropic and Amazon models | Broad infrastructure incumbency, model-choice flexibility |
None of the three platforms is a drop-in substitute for the others once a bank starts negotiating data-residency and model-training terms, which is exactly where BNP Paribas spent its negotiating effort with Google. The technical capability gap between the major agent platforms has narrowed considerably industry-wide; the governance terms attached to each deal are increasingly what actually differentiates one hyperscaler’s pitch from another’s for regulated customers.
Shattered.io has covered this three-way comparison in more technical depth, including pricing structure differences across the major agent platforms.
A Decade of Banks Warming Up to the Cloud
It’s worth remembering how recently “bank” and “public cloud” were treated as nearly incompatible terms. A decade ago, most large banks kept core banking systems entirely in owned data centers, citing regulatory requirements, latency, and a general institutional wariness about handing infrastructure control to a third party. That posture shifted gradually through the late 2010s and accelerated hard once cloud providers built out region-specific compliance tooling and financial-services-specific contract terms.
Generative AI has compressed that shift further. Banks that spent years running pilot programs for basic cloud storage and compute are now signing multi-year AI-agent partnerships in the same conversation, in part because the AI capability itself is the reason to move any given workload off-premises at all. BNP Paribas’s approach, aggressive on AI agents, conservative on data location, reads as the current default posture for large regulated banks rather than an outlier.
The European Regulatory Backdrop
BNP Paribas operates under some of the strictest data-governance expectations of any major bank globally, shaped by EU banking supervision and financial-data-handling rules. The European Banking Authority has spent recent years pushing guidance on outsourcing and cloud usage specifically for banks, and the Bank for International Settlements has published extensively on operational resilience expectations tied to third-party technology providers.
That regulatory backdrop is the practical reason BNP Paribas structured this deal the way it did. A European bank moving sensitive customer or medical data onto a shared public-cloud environment without a clearly documented governance framework would be inviting supervisory scrutiny it doesn’t need. Structuring the deal around explicit categories, AI agents on public cloud, sensitive data on-premises, is a defensible position a compliance team can actually document and audit.
How the Market and Industry Read the Announcement
Coverage of the deal has emphasized the data-governance angle over the AI-capability angle, with reporting from Reuters and Finextra both highlighting that BNP Paribas is keeping sensitive data off public cloud despite deepening its Google relationship. That framing is telling. In a year when nearly every large enterprise has announced some kind of AI-agent partnership, the detail that actually distinguishes this one is the boundary the bank drew, not the AI itself.
For Google Cloud, the deal is a visible proof point in a market where Microsoft and AWS both have longer-standing relationships with major global banks. Landing a five-year commitment from Europe’s largest bank by assets, with a public data-governance framework attached, gives Google’s financial-services sales teams a concrete reference to cite in future pitches to peer institutions.
Risks and Open Questions
The announcement leaves several practical questions unanswered. Neither company disclosed the financial value of the five-year agreement, so it’s not possible to gauge how large a commitment this represents relative to BNP Paribas’s total technology spend. No specific timeline was given for when the corporate-credit, sales, trading, research, and structuring agents will actually go into production use, versus remain in pilot or evaluation phases.
There’s also an open question about audit and oversight in practice, not just policy. Saying AI agents will be authenticated, scoped, and monitored is a governance commitment; whether that holds up under real production load, across five years and an unspecified number of deployed agents, is the kind of thing regulators and the bank’s own internal audit function will be testing on an ongoing basis rather than something settled by the announcement itself.
Finally, the hybrid split itself carries operational complexity. Running some workloads on Google Cloud and others strictly on-premises means BNP Paribas’s technical teams have to maintain a consistent security and access model across two very different environments, which is harder in practice than it sounds in a press release.
Five Predictions for Where Bank-Cloud AI Deals Go Next
- More named hybrid-cloud AI deals from European banks. BNP Paribas’s public data-governance framework gives other large EU banks a template to negotiate from, and expect at least a few peer institutions to announce structurally similar arrangements with one of the three major hyperscalers within the next year.
- Data-location terms become the headline, not the AI capability. As agentic AI capability converges across Gemini Enterprise, Azure AI Foundry, and Bedrock/AgentCore, the differentiating detail in future announcements will increasingly be what data stays off public cloud, not which model powers the agents.
- Insurance-linked medical data stays a hard carve-out industry-wide. Given how explicitly BNP Paribas singled out insurance medical information, expect other banking groups with insurance arms to draw the same line in their own cloud contracts.
- Trading-desk AI agents expand cautiously. BNP Paribas including trading among its target functions is notable; expect other banks to follow into trading-adjacent AI use cases, but slower and with more compliance review than back-office functions like credit memo drafting.
- Hyperscalers publish more financial-services-specific agent governance tooling. Expect Google, Microsoft, and AWS to each expand documentation and product features specifically addressing agent authentication, scoping, and monitoring, using deals like this one as the proof case for regulated-industry customers.
What This Means for Enterprise AI Buyers Generally
Outside of banking specifically, the BNP Paribas deal is a useful data point for any large enterprise negotiating an AI-agent contract with a hyperscaler right now. The structure that made this deal workable for BNP Paribas’s risk and compliance teams wasn’t a novel technical architecture, it was clear, documented boundaries: which workloads are eligible for public cloud, which categories of data are excluded outright, and explicit commitments on data ownership and model training.
That’s a reproducible negotiating framework for any regulated or data-sensitive organization evaluating Gemini Enterprise, Azure AI Foundry, or Bedrock, not just banks. The specifics will differ by industry, but the underlying pattern, agentic AI on shared infrastructure paired with an explicit, auditable carve-out for the data that can’t leave the building, looks increasingly like the default shape of enterprise AI deals for the rest of 2026 and into 2027.
Frequently Asked Questions
What did BNP Paribas and Google Cloud actually announce?
A five-year partnership, announced September 24, 2026, expanding BNP Paribas’s access to Google Cloud’s AI-optimized infrastructure, Gemini models, and Gemini Enterprise for building internal AI agents.
Is BNP Paribas moving all its data to Google Cloud?
No. The bank is explicitly keeping certain categories of sensitive customer data, medical information tied to its insurance operations, and critical operations on its own on-premises infrastructure rather than public cloud.
What is Gemini Enterprise?
Gemini Enterprise is Google Cloud’s platform for building, evaluating, and deploying purpose-built AI agents tailored to specific business functions, rather than a single general-purpose assistant.
Which parts of BNP Paribas’s business will use AI agents under this deal?
The announcement names corporate credit memos, sales, trading, research, and structuring as the target functions for agentic AI deployment.
Will Google use BNP Paribas’s data to train its AI models?
No. Google said BNP Paribas retains ownership of its data and that the data will not be used to train, tune, or refine Google’s models.
How much is the deal worth?
Neither company disclosed a monetary value for the five-year agreement in the official announcement.
How does this compare to Microsoft Azure or AWS deals with banks?
Microsoft’s equivalent enterprise-agent platform is Azure AI Foundry, and Amazon’s is Bedrock paired with its AgentCore runtime. All three hyperscalers are pursuing similar regulated-industry deals, with data ownership and no-training commitments becoming standard terms across the market rather than a Google-specific differentiator.
Why does BNP Paribas single out insurance medical data specifically?
BNP Paribas’s insurance operations hold health records tied to life and disability policies, a category of data that carries some of the strictest handling requirements of anything a European financial institution manages, which is likely why it was called out as an explicit example of data staying off public cloud.
