Guides › Cloud AI platform data policies
Azure OpenAI, Amazon Bedrock and Vertex AI: Data Policies Compared for Enterprise Knowledge Bases
All three cloud model platforms state that prompts and outputs are not used to train foundation models without permission, and Microsoft and AWS also state that the model provider has no access to customer prompts and completions. They differ on what is retained for abuse monitoring, how zero retention is reached and where a knowledge base's files and vectors are stored. Azure OpenAI, now part of Microsoft Foundry, may keep a sample of flagged traffic for human review unless Microsoft approves modified abuse monitoring. Amazon Bedrock lets you set a data retention mode per Region, with some models requiring 30-day retention. Google Cloud's Gemini Enterprise Agent Platform, where the Vertex AI generative AI documentation now lives, has no single switch and lists the features to turn off. This page compares the three from each provider's documentation as of November 2026.
- Training. None of the three trains foundation models on customer prompts without permission.
- Abuse monitoring. Azure may store sampled flagged content for human review. Bedrock uses automated detection, with up to 30 days of retention for specific models. Google logs prompts for abuse monitoring for customers on its standard terms.
- Zero retention. Azure by approval, visible as ContentLogging false. Bedrock by setting the retention mode to none. Google by disabling a list of features per project.
- Knowledge bases. Azure On Your Data leaves data in your source. Bedrock keeps vectors in a store you manage or a managed store. In every case the index is storage you create and have to govern.
The Comparison
Each cell paraphrases the provider's documentation, linked at the end of the page. The provider policy comparison covers the assistants (ChatGPT, Claude, Microsoft 365 Copilot and Gemini). This page covers the platforms that teams build their own retrieval systems on.
| Question | Azure OpenAI (Microsoft Foundry) | Amazon Bedrock | Google Cloud (Gemini Enterprise Agent Platform) |
|---|---|---|---|
| Training and provider access | Prompts, completions, embeddings and training data are not used to train foundation models without permission and are not available to OpenAI or other model providers. | Models run in AWS-operated deployment accounts that model providers cannot access, so providers do not see prompts, completions or logs. | No training or fine-tuning on customer data without prior permission or instruction, including on pre-GA models. |
| Abuse monitoring | Automated review by default. Flagged samples may be stored for human review by authorized Microsoft employees, in a store separated by customer resource and kept in the resource's geography. | Automated detection under a zero operator access model. Classifier-flagged traffic on most OpenAI models is retained up to 30 days. All traffic on Claude Fable 5, 5.1 and 5.2 is retained up to 30 days with possible human review by AWS. | Prompt logging for abuse monitoring applies to customers on the standard Google Cloud terms. Models designated Advanced AI carry additional prompt and response logging. |
| Zero retention | Apply for modified abuse monitoring through Limited Access. Approval shows as ContentLogging false on the resource. Automated review may still run. | Set the data retention mode to none, per Region, at account or project level. Models that require retention are then blocked. ZDR for those models is by approval through the AWS account team. | No single switch. Disable caching, leave request-response logging off, set store to false on the Interactions API and avoid grounding with Google Search or Maps. Partner model ZDR is approved per project. |
| Features that store data | Files, vector stores, Responses API state, Assistants threads and stored completions, held in your Foundry resource in the same geography. | Stored responses on the Responses API in default mode. Model invocation logging if you enable it. It is off by default. | Implicit prompt caching is on by default and cleared within 24 hours. The Interactions API stores by default. Grounding with Google Search stores derived queries for up to 3 days, and with Google Maps for 30 days. |
| Knowledge base storage | On Your Data leaves the data in the source and location you designate and does not create a duplicate store. | Knowledge Bases write embeddings to a vector store you manage, such as OpenSearch Serverless or Aurora, or to a fully managed knowledge base that you can encrypt with your own KMS key. | A retrieval corpus is a resource you create in your project, and RAG Engine supports customer-managed encryption keys. It sits outside the zero retention list, which addresses the model call. |
| Residency | Processed in the resource's geography, except Global and DataZone deployments, which can process in any geography or anywhere in the zone. Batch is a Global deployment. | Retention settings are per Region. With cross-Region inference, retained data is stored in the Region that processes the request. | Regional and multi-regional endpoints keep durable cache data in the selected location. The global endpoint can serve from any region. |
What Each Platform Means for a Knowledge Base
Azure OpenAI
The model call is stateless and the retrieval source stays where you put it, which makes Azure the simplest of the three to reason about for an index you already run in your own subscription. Two things need attention. Files, vector stores, Assistants threads and stored responses are kept in your Foundry resource until you delete them, so a build that uses those features holds a second copy of the corpus. And the abuse monitoring store is the one place Microsoft may hold your prompts, with retrieved chunks inside them, until modified abuse monitoring is approved.
Amazon Bedrock
Bedrock gives the most direct control: a retention mode you can read and enforce through policy, and an explicit error when a model's requirement conflicts with it. An organization can require mode none across every account with a service control policy. The cost is model choice. The Claude Fable 5 family is unavailable unless the account opts in to 30-day retention with possible human review by AWS, or qualifies for the Enterprise Frontier Safeguards program, which AWS says allows zero retention for internal use through at least June 30, 2027. For the knowledge base itself, choose between a vector store in your account and the managed option, and supply your own KMS key in either case.
Google Cloud
Google's approach puts the work on the customer. Zero retention is the result of several settings: caching disabled or limited to memory, request-response logging left off, stored interactions turned off in each request, and grounding services avoided or replaced with the enterprise web grounding option. Prompt caching needs a specific decision for retrieval workloads, because the cached content is the retrieved context. Durable caching writes derived cache data to storage at rest for up to 24 hours and does not support customer-managed keys. It can be turned off per project while keeping in-memory caching.
The Same Model Under Different Terms
A model's data terms depend on who operates it. Anthropic states that on Amazon Bedrock and on Google-managed Claude deployments the cloud provider is the data processor, and directs customers to that platform's documentation. On Claude Platform on AWS and Anthropic-managed Claude on Google Cloud, Anthropic's own retention policy applies. Google states that zero retention approvals for the direct OpenAI API do not apply to OpenAI models on its platform. Microsoft states that prompts sent to Azure OpenAI are not available to OpenAI. The guide to getting ZDR covers what to request from whom.
What None of the Three Changes
These policies govern how the platform handles content after it arrives. They do not decide what your retrieval layer returns or whether a returned chunk contained an identifier. A knowledge base on any of the three still needs retrieval scoped to the asking user's entitlements, substitution of real identifiers before the prompt is assembled, and a retrieval record kept on your side. With those in place, the differences above determine how long a de-identified prompt is held, which is a smaller question than the one most reviews start with. The secure RAG architecture guide and the page on ZDR for knowledge bases cover the design.
Where Hardshell Sits
Hardshell runs in front of whichever platform hosts the model. It connects to the systems where enterprise data already lives without migrating it, enforces each user's existing permissions at query time, substitutes stand-ins for real identifiers before the prompt reaches Azure OpenAI, Amazon Bedrock, Google Cloud or a direct model API, and records every retrieval under opaque IDs. Because it is model-agnostic, a change of model or platform does not change the control. Hardshell deploys as a cloud service, in a private cloud inside your own account or as an offline bundle.
Frequently Asked Questions
Does Amazon Bedrock store my prompts?
It depends on the retention mode and the model. With the data retention mode set to none, AWS states that no request or response data is written to durable storage. In default mode the model's own policy applies, and specific models retain flagged traffic or all traffic for up to 30 days for abuse detection.
Does Azure OpenAI send my data to OpenAI?
No. Microsoft states that prompts, completions, embeddings and training data are not available to OpenAI or other model providers and are not used to train foundation models without permission.
Does Vertex AI retain prompts?
Google Cloud's generative AI platform, now documented as Gemini Enterprise Agent Platform, caches prompt data for up to 24 hours by default and may log prompts for abuse monitoring for customers on its standard terms. Google publishes the settings to change to reach zero data retention.
Which cloud platform is best for zero data retention?
Each reaches it by a different route. Bedrock offers a setting you control and can enforce by policy, limited by model requirements. Azure requires Microsoft's approval and shows the result on the resource. Google requires disabling several features per project. The right choice depends on the models you need and the features your application uses.
Where is my vector store on these platforms?
Where you put it. Azure's On Your Data feature leaves data in the source you designate. A Bedrock knowledge base uses a vector store in your account or a managed store. In every case the index persists until you delete it and is outside any zero retention setting on the model call.
Does using Claude on Bedrock change Anthropic's retention terms?
Yes. Anthropic states that on Amazon Bedrock the cloud provider is the data processor, so AWS's retention modes and abuse detection terms apply. A zero data retention arrangement with Anthropic for its direct API does not transfer.
Related guides
- What AI providers do with your data · OpenAI, Anthropic, Microsoft and Google compared.
- How to get zero data retention · Eligibility, request steps and verification on each provider.
- ZDR for RAG and AI knowledge bases · Which retrieval components stay stored under a ZDR agreement.
Sources (accessed November 2026)
Microsoft Learn, Data, privacy, and security for Azure Direct Models in Microsoft Foundry; Abuse monitoring. AWS, Amazon Bedrock data retention; Amazon Bedrock abuse detection; Data protection; Model invocation logging; Knowledge Bases; Encryption of knowledge base resources. Google Cloud, Gemini Enterprise Agent Platform and zero data retention; RAG Engine overview. Anthropic, API and data retention.