An enterprise AI platform runs models for you under business terms
An enterprise AI platform is a cloud service that runs AI models for a company, under business terms rather than consumer ones. An AI agent inside your ERP or CRM sends records to one of these platforms each time its model reads them, unless the model runs on your own servers. That puts your sales orders and customer records under the platform’s terms.
We put four questions to each platform’s own pages:
- Does it train on our data?
- How long does it keep prompts and outputs?
- Can the work stay in our country or region?
- Which limits does an agent hit?
The answers below are the vendors’ words, read on 18 November 2026. We did not test any platform’s claims ourselves.
All five say they do not train on business data, but they keep and place it differently
OpenAI API
OpenAI says that, as of March 1, 2023, data sent to its API is not used to train its models unless you opt in.
It keeps abuse monitoring logs for up to 30 days by default, and its Responses API keeps app state for 30 days too. Zero Data Retention exists, but it needs OpenAI’s prior approval. For residency, you pick a region per project, and it cannot change after the project is made. OpenAI’s region table lists storage and processing apart. We counted 10 regions with storage, and 6 of them also run inference there. Europe (EEA and Switzerland) has both. The UK region stores data but runs inference elsewhere. System data, meaning account data, metadata and usage data, falls outside residency.
Rate limits sit at the organization and project level, not per user. A request rate that climbs too fast returns a 429 error, and so does a hard spend limit.
Anthropic Claude API
Anthropic says it does not train on inputs or outputs from its commercial products by default. Its commercial terms go further, saying Anthropic may not train models on customer content. The customer keeps rights to its inputs and owns its outputs. Feedback is different. If you choose to report feedback, such as a thumbs up or down, Anthropic may use it for training.
API inputs and outputs are deleted within 30 days. Chats flagged for breaking the usage policy are kept longer, up to 2 years. Their safety scores may be kept up to 7 years.
Anthropic has two location settings. The inference_geo setting picks where inference runs. It defaults to global, so inference may run in any geography, and it works on Claude 4.6 and later models. Workspace geo picks where data is stored at rest. It is set when a workspace is made, it cannot change after, and “us” is the only option today. Limits are counted per minute for each model class. On the Start tier, Claude Sonnet 5.5 gets 4,000 requests, 2,000,000 input tokens and 400,000 output tokens a minute, and US-only and global requests share that pool.
Google: Gemini API and Gemini Enterprise Agent Platform
Google sells model access two ways, with different terms. On the Gemini API, paid tier content is not used to improve Google’s products, but free tier content is. On Google Cloud’s Agent Platform, Google won’t train or fine-tune models on your data without your permission.
Google lists several things it keeps, and zero retention needs action on each. It may log prompts for abuse monitoring. Its published Gemini models cache data by default, in memory and in durable storage, and the memory cache is deleted within 24 hours. Grounding with Google Search keeps query logs for up to 3 days, with no way to turn that off while you use it. The Interactions API stores prompts and responses unless you set store to false. Request-response logging stays off unless you switch it on.
A global endpoint gives no residency promise. The EU multi-region endpoint covers EU member states only, so it leaves out the United Kingdom and Switzerland. Google also lists single regions, including the United Kingdom (europe-west2), for some models, so check your model’s row. Gemini API rate limits apply per project, not per API key, and daily quotas reset at midnight Pacific time.
Microsoft Foundry, with Azure OpenAI
Microsoft says prompts and completions are not used to train foundation models without your permission. OpenAI and other model providers cannot see them. Models sold by Azure do not talk to services OpenAI runs, such as ChatGPT or the OpenAI API. Stored data is encrypted at rest with AES-256 by default, and you can delete it at any time.
Prompts are processed in the geography you choose, unless you pick a Global or DataZone deployment. Stored data stays in your geography either way. Data zone deployments keep processing inside the US, EU or Asia Pacific zone. Regional deployments come in up to 27 regions.
Modified Abuse Monitoring is an approval that lets a customer skip Microsoft’s abuse data storage and human review. Microsoft says it will end for models made available on or after November 1, 2026, from a date not yet announced, with at least 30 days’ notice.
Limits are set per model, deployment type and quota tier. On quota Tier 1, a gpt-4.1 DataZoneStandard deployment gets 300 requests and 300,000 tokens a minute. Account limits sit above that: 30 Azure OpenAI resources per region per subscription.
Amazon Bedrock
Amazon says neither AWS nor the model providers use Bedrock inputs or outputs to train Amazon Nova, Amazon Titan or any third-party model. Inputs and outputs are not shared with model providers. Customer content is encrypted and stored at rest in the AWS Region where you use Bedrock.
Bedrock now lets you choose a data retention mode, per Region. You set it for the whole account or for one project, and it does not carry over to other Regions. The mode called none means zero data retention, with nothing written to durable storage. Some models require retention, though. Claude Fable 5, 5.1 and 5.2 require human review by AWS, so with none set, Bedrock blocks those requests with an error. Amazon says most models do not need review. On 18 November 2026, Amazon’s page said there was no console screen for this setting yet.
An inference profile tied to a geography, such as the US or EU, lets Bedrock pick a Region inside that geography for each request. Profiles can also be global. Quota increases are not granted automatically. Bedrock’s two inference endpoints are called bedrock-runtime and bedrock-mantle, and each keeps its own per-model quotas.
Side by side, the five differ most on where the work can stay
Scroll the table sideways to see every column.
| Platform | Trains on business data? | How long prompts are kept | Where the work can stay | Where limits are set |
|---|---|---|---|---|
| OpenAI API | Not unless you opt in | Abuse logs up to 30 days. Responses app state 30 days | Storage in 10 regions, inference in 6, chosen per project. UK: storage only | Organization and project |
| Anthropic Claude API | Not by default. Commercial terms forbid it | Deleted within 30 days. Flagged chats up to 2 years | Inference global by default, or US. Storage: “us” is the only workspace geo | Per model class, per minute |
| Google Agent Platform and Gemini API | Not without permission. Free Gemini API tier content improves products | Search grounding logs up to 3 days, if you use it. Cache on by default | US or EU multi-region, or a single region such as the UK, by model | Per project, not per API key |
| Microsoft Foundry | Not without permission | Stored data can be deleted by the customer at any time | Processing in your geography or a US, EU or Asia Pacific data zone. Storage stays in your geography | Per model, deployment type and quota tier |
| Amazon Bedrock | No | Your choice of mode, per Region | The Region you use, or a US or EU profile | Per model, per endpoint |
The AI inside your business systems adds its own data terms
When you use the AI built into an ERP or CRM, two sets of terms apply: the business system vendor’s, and the model platform’s behind it. Here is what each vendor says, from its own pages.
- Salesforce. The Einstein Trust Layer includes zero data retention and toxicity detection.
- HubSpot. Third-party AI providers are barred by contract from training their models on your data.
- Oracle NetSuite. The N/llm module uses Cohere’s Command A model unless a script names another. Oracle says the data is not used by third parties for model training.
- Xero. JAX does not use your data to train the language models it uses, and its providers include AWS, Microsoft Azure, GCP and OpenAI. Xero also bars anyone from training an AI model on data taken through its APIs.
- QuickBooks. Intuit says QuickBooks data is not used to train Claude through its connector.
- Odoo. Odoo 20 keeps conversations with AI agents for up to 30 days after the last message.
- SAP Business One. The Web client’s AI features reach a model through the SAP AI Core service on SAP BTP.
- Dynamics 365 Finance. The customer page summary runs on Azure OpenAI Service.
- ERPNext. Frappe Flow works with any server that speaks the OpenAI API, including Ollama and LM Studio. That means the model can run on your own servers.
How to choose a platform for an ERP or CRM agent
Start from where your data must stay, then check retention, then limits. Rules 1 to 4 come from the vendor facts above, and rule 5 from our own work.
- If data must stay in the EU, set it up before the first project. OpenAI’s region is fixed when the project is made. Use Google’s EU multi-region endpoint, Microsoft’s EU data zone or a Bedrock EU profile. Anthropic’s only workspace geo today is the US.
- If UK data is in scope, check storage and inference apart. OpenAI’s UK region stores data but runs inference elsewhere. Google’s EU endpoint leaves out the UK, but Google lists a UK single region for some models.
- If you need zero retention, plan for the trade-offs. OpenAI’s Zero Data Retention needs approval. Bedrock’s none mode blocks models that require retention, such as Claude Fable 5. On Google, set store to false, skip Grounding with Google Search, and follow Google’s steps for caching and abuse-monitoring logs.
- If several agents share one project, they share its limits. OpenAI and Google set limits per project, and Amazon will not raise a Bedrock quota on its own.
- Count your own servers too. The model platform is only one place your data sits. The crash news agent our team built for a US personal injury law firm runs on a dedicated server in the US. An AI model reads the articles, and plain rules make the decisions (case study).
We sign your NDA before you share any data, if you ask. For help with these four questions on your own systems, see AI consulting. AI integration covers linking the chosen platform to your ERP or CRM. Our list of AI agent builders covers the tools that sit on top of these platforms.
Sources
Every source behind a figure or a vendor fact in this article, with the day we read it.
- OpenAI: Data controls in the OpenAI platformRead 18 Nov 2026
- OpenAI: Data residencyRead 18 Nov 2026
- OpenAI: Rate limitsRead 18 Nov 2026
- Anthropic: Is my data used for model training?Read 18 Nov 2026
- Anthropic: Commercial Terms of ServiceRead 18 Nov 2026
- Anthropic: How long do you store my organization’s data?Read 18 Nov 2026
- Anthropic: Data residencyRead 18 Nov 2026
- Anthropic: Rate limitsRead 18 Nov 2026
- Google: Gemini Developer API pricingRead 18 Nov 2026
- Google: Rate limitsRead 18 Nov 2026
- Google Cloud: Gemini Enterprise Agent Platform and zero data retentionRead 18 Nov 2026
- Google Cloud: Data residencyRead 18 Nov 2026
- Microsoft: Data, privacy, and security for Foundry Models sold by Azure in Microsoft FoundryRead 18 Nov 2026
- Microsoft: Understanding deployment types in Microsoft Foundry ModelsRead 18 Nov 2026
- Microsoft: Azure OpenAI in Microsoft Foundry Models Quotas and LimitsRead 18 Nov 2026
- Microsoft: Azure OpenAI ServiceRead 18 Nov 2026
- Amazon: Amazon Bedrock FAQsRead 18 Nov 2026
- Amazon: Route model inference requests across AWS Regions with cross-Region inferenceRead 18 Nov 2026
- Amazon: Quotas for Amazon BedrockRead 18 Nov 2026
- Salesforce: How Does Agentforce Work?Read 18 Nov 2026
- HubSpot: AI Customer Agent for Customer SupportRead 18 Nov 2026
- Oracle NetSuite: SuiteScript 2.1 Generative AI APIsRead 18 Nov 2026
- Xero: Just Ask XeroRead 18 Nov 2026
- Xero: Pricing and Policy Updates FAQsRead 18 Nov 2026
- Intuit: Use Claude to enter transactions and get insights in QuickBooksRead 18 Nov 2026
- Intuit: QuickBooks pricing pageRead 18 Nov 2026
- Odoo: Odoo 20 Release NotesRead 18 Nov 2026
- SAP: Enabling AI for Web ClientRead 18 Nov 2026
- Microsoft: Agents, Copilot, and AI capabilities in Dynamics 365 appsRead 18 Nov 2026
- Microsoft: Sales Order Agent overviewRead 18 Nov 2026
- Microsoft: Payables Agent Overview in Business CentralRead 18 Nov 2026
- Frappe: Set Up FlowRead 18 Nov 2026
- Amazon: Data retentionRead 18 Nov 2026
- CadenceWorks: Case studies: AI agents and ERP buildsRead 18 Nov 2026
- HubSpot: Run, Build, and Manage Your AI AgentsRead 18 Nov 2026