Zero Data Retention in Generative AI: what companies should demand from their providers

Reuters, which attributes the information to The Information and sources familiar with the discussions, details several measures related to Anthropic: Palantir reportedly requested an irrevocable commitment to zero retention; Nvidia would limit its models to less sensitive work and use Nemotron for certain internal tasks; and Booz Allen reportedly banned its business model from work related to proprietary cybersecurity software. The report also notes that OpenAI faces similar scrutiny regarding the handling of its clients’ data.

The debate had a recent precedent. In June, Anthropic had introduced a 30-day retention period for certain advanced models, including Fable 5, intended to strengthen security oversight. The measure concerned clients who needed to use these models with zero-retention guarantees. Before the information gathered by Reuters was published, Anthropic had already announced a specific solution to offer this option to certain enterprise clients.

Beyond this specific case, zero retention will most likely become more relevant as the enterprise adoption of AI matures and its integration via APIs into corporate applications, processes, and data sources increases. But zero retention does not mean that data never leaves the company or that all metadata disappears: its scope depends on the service, the configuration, and the contractual terms.

Before choosing a tool, organizations should ask themselves what data the provider will see, where it will be processed, and what the provider will retain. It is not enough to know whether a service offers zero retention. It is also necessary to understand which product, model, features, and data that guarantee actually covers. The most powerful model will not always be the right one. Maturity means using AI where it adds value, with controls commensurate with the sensitivity of the information.

La retención cero responde a principios de privacidad y seguridad: recopilar solo los datos necesarios, limitar los registros de actividad y eliminar la información cuando deja de ser necesaria. El Reglamento General de Protección de Datos recoge la minimización de datos y la limitación del plazo de conservación, aunque no utiliza expresamente la denominación Zero Data Retention. La retención cero puede contribuir a aplicar estos principios, pero no equivale por sí sola al cumplimiento del RGPD.

Zero data retention is based on principles of privacy and security: collecting only the necessary data, limiting activity logs, and deleting information when it is no longer needed. The General Data Protection Regulation (GDPR) addresses data minimization and the limitation of retention periods, although it does not explicitly use the term Zero Data Retention. It can help implement these principles, but it does not, by itself, ensure compliance with the GDPR.

In generative AI, zero retention has gained prominence as organizations have integrated these models into their applications and processes via application programming interfaces (APIs) and have required greater assurances regarding the handling of their queries and documents. This is particularly relevant when AI works with source code, contracts, financial reports, or customer data. The persistence of this data in external systems increases exposure and raises concerns about access, jurisdiction, and incidents. However, zero retention does not mean that the data never leaves the corporate environment. The information leaves the application used by the employee, travels across the network, and may pass through API connectors or gateways before reaching the provider. There, it is processed, and the response is sent back to the company.

To assess this journey, the company must distinguish between three issues that are often confused: how long the content is retained, whether it is used to train models, and what technical logs are maintained.

Retention indicates whether queries, attachments, and responses remain stored after the request is processed. For example, by default, OpenAI retains certain logs for up to 30 days to detect abusive use, unless required by law or other specified exceptions apply. Customers who meet the provider’s requirements may access zero-retention options, the scope of which depends on the service and features used.

Anthropic typically deletes API inputs and outputs within that timeframe, but conditions may vary depending on the model, feature, and its security policies. Certain models, products, or features may be excluded from a zero-retention agreement.

The company should verify the timeframe, exceptions, handling of backups, and the exact scope of the service.

This usage determines whether the provider can use queries or responses to improve its models. Both OpenAI and Anthropic state that, by default, they do not use data from their enterprise products or APIs to train their models.

However, just because the data is not used for training does not mean it is not temporarily retained. The company must verify the default settings and ensure that the contract explicitly states that its information will not be used for that purpose.

Logs may contain information about the user, the date of the request, the IP address, the service used, the volume of data consumed, or security alerts. In some cases, they may also include the content of queries or responses.

The company must know which logs are generated, where they are stored, how long they are retained, who can access them, and whether monitoring systems or human review are involved.

Therefore, it is not enough to simply ask if a provider offers zero retention. You must identify the exact service, verify that the features used are compatible, and ensure that this guarantee is reflected in the contract and the technical configuration.

A company can use general-purpose services to work with public information. When internal data, intellectual property, or personal information is involved, it is advisable to require stronger safeguards. For critical processes, you may need a model deployed in a private cloud, on your own infrastructure, or under specific controls.

Therefore, before incorporating an artificial intelligence tool into a business process, it is advisable to seek specific answers to these questions:

• Which products, APIs, models, and features are actually covered by zero retention.
• Whether queries, files, or responses are retained, for how long, and what happens to backups.
• Whether the content can be used to train or improve models, and whether that option is disabled by default.
• What metadata is stored—user, IP address, date, data usage, or security alerts—and for how long.
• What path the information takes from the application to the model, in which region it is processed, and which providers or subprocessors are involved.
• What happens when using connectors, agents, or external tools—for example, a CRM, a search engine, a code repository, or a document management platform.
• How access is controlled and what evidence the provider provides—audit logs, certifications, or security reports.
• What obligations the provider assumes in the event of an incident, a change in terms, or the termination of the service.

These guarantees must be included in the contract governing the service and reviewed periodically. They may change with the launch of a new model, the connection to another data source, or the activation of an additional feature.

Provider commitments are one aspect of data protection. The company also needs to implement controls over its own environment and over how users, applications, and locations access AI services.

From the network, you can block unauthorized AI services, limit their use by users, locations, or applications, maintain visibility into connections, and reduce the risk associated with Shadow AI.

SAIWALL Secure SD-WAN helps enforce these policies across the infrastructure without replacing the contractual guarantees required of the AI provider.