A bank generally cannot use the consumer versions of ChatGPT or Claude with real customer data without risking a breach of banking secrecy obligations, because those consumer products send prompts to a third-party provider outside the bank's control and, depending on settings, may retain that data for service improvement. Enterprise agreements, such as Azure OpenAI Service or Claude accessed through AWS Bedrock with a zero data retention contract, change this picture by keeping data within a defined cloud boundary and adding contractual protections, which can satisfy some regulators for lower-sensitivity use cases if a proper data processing agreement and risk assessment are in place. Many banking secrecy regimes, including Swiss banking law and Turkish banking legislation, and many internal bank policies still treat any transmission of identifiable customer data to an external processor as requiring explicit legal review, and some prohibit it outright for the most sensitive data categories. The safest architecture for genuinely sensitive customer and transaction data is a self-hosted open-weight model running entirely on infrastructure the bank controls, which removes the third-party transmission question altogether. Any use of a hosted API should go through the bank's data protection and legal teams before rollout, not after. Nanobase AI helps banks choose between private cloud APIs and fully self-hosted models based on their specific secrecy and compliance obligations.
The question is which version, not whether at all
Framing this as "can a bank use ChatGPT or Claude" skips the detail that actually matters, which is that these products exist in materially different deployment forms with different data handling guarantees. The decision that determines banking secrecy exposure is which deployment tier a given use case is routed to, based on data sensitivity, not a blanket yes-or-no answer applied to the product name. A bank that treats "ChatGPT" as one monolithic risk category either blocks genuinely low-risk uses unnecessarily or, worse, fails to distinguish between a safe enterprise deployment and an unsafe consumer one.
A decision framework by data sensitivity
| Data sensitivity tier | Example use case | Acceptable deployment | Why |
|---|---|---|---|
| Public or synthetic data | General research, drafting non-customer content | Consumer ChatGPT/Claude.ai | No customer data transmitted |
| Low-sensitivity internal | Internal style guide questions, generic policy drafting | Enterprise API with contractual protections | Contained within a defined cloud boundary |
| Customer-identifiable, non-transactional | Customer service macros referencing account type only | Enterprise API with zero data retention contract, after legal review | Requires DPA and risk assessment |
| Real customer/transaction data | Case investigation, KYC review, credit file review | Self-hosted open-weight model only | Removes third-party transmission entirely |
The safest architecture for genuinely sensitive customer and transaction data is a self-hosted open-weight model, which is the only tier in this table that removes the third-party data transmission question rather than managing it contractually.
What actually changes between consumer and enterprise tiers
Consumer versions of ChatGPT or Claude send prompts to the provider's infrastructure outside the bank's control, and depending on account settings may retain that data for service improvement, which is the core banking secrecy problem. Enterprise agreements, such as Azure OpenAI Service or Claude accessed through AWS Bedrock with a zero data retention contract, change the picture meaningfully by keeping data within a defined cloud tenancy boundary and adding contractual terms that can satisfy some regulators for lower-sensitivity use cases, provided a proper data processing agreement and risk assessment are completed first. This is a real improvement over the consumer tier, but it is a contractual and architectural mitigation, not the same guarantee as never transmitting the data to a third party at all.
Jurisdiction-specific rules that override the general framework
Banking secrecy regimes vary, and some are stricter than the general enterprise-API mitigation can satisfy. Swiss banking law and Turkish banking legislation, along with many banks' own internal policies, often treat any transmission of identifiable customer data to an external processor as requiring explicit legal review, with some prohibiting it outright for the most sensitive data categories regardless of contractual protections. This means the framework above is a starting point, not a substitute for jurisdiction-specific legal review, since a use case that would be acceptable under a general enterprise API tier in one jurisdiction may require full self-hosting in another.
A rollout process that keeps legal comfortable
- Classify the intended use case by data sensitivity tier before selecting a deployment option, not after a team has already started building against a specific API.
- Confirm the applicable banking secrecy regime's specific requirements for third-party data transmission, since general industry practice does not override local law.
- Route anything above the low-sensitivity tier through legal and data protection review before any real data touches the system.
- Default to self-hosted open-weight models for any use case where jurisdiction or sensitivity creates ambiguity, since it removes the underlying question rather than managing around it.
- Document the deployment decision and its rationale, since this becomes the evidence an examiner or auditor will expect to see.
Routing the deployment decision through legal before development starts, rather than after a pilot is already built, avoids the common and expensive outcome of a working system that legal ultimately blocks from touching real data.
Frequently asked questions
Does a zero data retention contract fully solve banking secrecy concerns?
It substantially reduces risk for lower-sensitivity use cases by removing data retention by the provider, but it does not eliminate the fact that data is transmitted to a third party, which some banking secrecy regimes still restrict regardless of retention policy.
Can a bank use consumer ChatGPT or Claude for any work at all?
Yes, for tasks involving no customer or confidential data, such as general research or drafting generic, non-sensitive content, consumer tools present materially lower risk, though many banks still restrict this by policy to avoid employee error.
Is self-hosting always more expensive than an enterprise API?
Self-hosting requires upfront infrastructure investment that an API model avoids, but for sustained, high-volume use touching sensitive data, the total cost comparison depends heavily on usage volume and should be evaluated against the specific workload rather than assumed.
Who should make the final call on which deployment tier a use case gets?
This decision should involve the bank's legal and data protection teams alongside the technical team, since the classification depends on regulatory interpretation as much as technical architecture.
How Nanobase AI helps
Nanobase AI helps banks choose between private cloud APIs and fully self-hosted models based on their specific secrecy and compliance obligations, then builds whichever deployment the classification calls for. See our solutions, or continue with how banks deploy LLMs on-premise for data security for the self-hosted architecture referenced above.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.