Major enterprise technology companies are restricting how they use advanced AI models as concerns over proprietary information and data retention begin to influence which frontier systems are trusted with sensitive workloads.
Reuters reported, citing The Information, that Palantir and Nvidia have limited their use of Anthropic models, while Booz Allen Hamilton has restricted employees from using Anthropic’s commercial AI for some proprietary cybersecurity work. The companies are seeking stronger guarantees over the handling of sensitive corporate data.
The development exposes an increasingly important enterprise-AI constraint: model quality alone does not determine whether software can be deployed inside sensitive corporate systems.
Palantir wants stronger zero-retention commitments
Reuters reported that Palantir has sought irrevocable zero-data-retention commitments from Anthropic, while Nvidia has limited Anthropic models to less-sensitive tasks. Booz Allen has reportedly prohibited some use involving proprietary cybersecurity work.
Anthropic’s own commercial-data policy says inputs and outputs from its API are automatically deleted within 30 days of receipt or generation under its standard retention arrangement. Exceptions include services where customers control longer retention, separate contractual agreements, enforcement of Anthropic’s Usage Policy and legal requirements.
The company also offers zero-data-retention arrangements to some approved commercial customers. Anthropic says certain API and Claude Code Enterprise customers may have agreements under which it does not store inputs or outputs except where required by law or to combat misuse or harm.
Those arrangements are not universal. Anthropic says zero-retention requests are reviewed on a per-organization basis and that some covered models require limited retention and review as part of its safety work.
Privacy and AI safety create competing requirements
The dispute highlights a difficult technical trade-off.
Enterprise customers often want confidential prompts, source code, security findings or business records deleted immediately. Frontier-model providers, meanwhile, may need visibility across interactions to detect abuse or sophisticated safety risks.
OpenAI has been developing another approach to that problem. In August, it previewed Private Safety Processing, designed to identify patterns across related interactions without giving OpenAI personnel access to the underlying content.
OpenAI says its Zero Data Retention option gives eligible API customers a commitment that prompts and model responses are not retained after a request is processed. For those deployments, customer content remains on infrastructure controlled by the customer while automated systems can return limited safety signals.
OpenAI also states that enterprise and API customer data is not used for model training unless customers explicitly opt in.
Data governance is becoming part of model competition
The implications go beyond Anthropic, Palantir or Nvidia.
As companies move AI from generic productivity tasks into cybersecurity, software development, financial analysis and internal data systems, prompts can contain some of an organization’s most valuable information.
That changes how enterprises compare model providers.
Benchmarks, latency and API prices still matter, but so do retention periods, contractual guarantees, access controls, customer-managed infrastructure and whether sensitive interactions can be inspected by the provider.
For frontier-model companies, privacy architecture is therefore becoming a product feature rather than a compliance footnote.
A model may outperform a rival on reasoning benchmarks and still lose an enterprise deployment if its data-handling guarantees do not meet the customer’s security requirements.
As AI moves deeper into proprietary workflows, the next competitive advantage may be not only what a model can do, but how little of the customer’s data the provider needs to keep while doing it.
