As artificial intelligence becomes embedded in core business operations, data security and intellectual property protection are emerging as decisive factors in how companies select AI models.

According to a September 14 report from The Information, major enterprises including NVIDIA, Palantir Technologies Inc., and Booz Allen Hamilton have recently restricted the use of Anthropic's flagship model Fable, following Anthropic's June policy change that requires retaining customer usage logs on a rolling 30-day basis.

These concerns extend beyond Anthropic alone. OpenAI faces similar scrutiny from enterprise clients over its data usage practices. Reports indicate that Microsoft is actively leveraging this situation to court OpenAI's corporate customers, having pitched at least one large client on an AI solution that runs entirely on private servers. For businesses, the appeal of such an offering lies in reducing the risk of sensitive data entering external AI platforms.

As these worries proliferate, some enterprises are moving beyond contractual "zero data retention" (ZDR) commitments, seeking instead to directly control both data and model operating environments—including deploying on their own servers or even developing proprietary models to keep sensitive information isolated from external AI vendors.

Fable Policy Change Triggers Data Retention Concerns

Both Anthropic and OpenAI state that under default enterprise contract terms, they do not use customer inputs or model outputs for training purposes. However, both companies still collect a degree of metadata.

In June, Anthropic adjusted Fable's data retention policy to require keeping customer usage logs for 30 days, heightening concerns among some enterprise clients about data security. This month, Anthropic introduced a self-storage option allowing eligible companies to keep relevant data on their own servers, though companies must apply separately and Anthropic retains the right to revoke this option.

OpenAI faces similar disputes over data usage. Chief Research Officer Mark Chen has stated the company uses some metadata to improve models, including recent work on the Navier-Stokes mathematical problem. Enterprises worry that so-called "de-identified" data may still contain behavioral records of employee software usage, and the boundaries of how such data is used remain opaque.

From Palantir to NVIDIA: Enterprise AI Security Strategies Diverge

Among enterprise clients, Palantir Technologies Inc. maintains particularly stringent data protection requirements. Citing informed sources, the report notes that Palantir previously negotiated with Anthropic seeking irrevocable ZDR guarantees covering all models. After Fable's data retention policy change, Palantir suspended offering Fable to its customers through its platform and demanded ZDR commitments that could not be revoked retroactively.

Palantir CEO Alex Karp stated last week that enterprises have grown weary of AI labs "abusing" their data. The company has also distributed handbooks to clients guiding them in negotiating ZDR terms with AI vendors. Currently, Palantir has secured access to OpenAI GPT-6 Astra under zero data retention conditions.

NVIDIA has adopted a more flexible approach: using Fable only for lower-sensitivity tasks like open-source software, while shifting core operations such as supply chain monitoring to its proprietary Nemotron model. Justin Boitano, NVIDIA's vice president of enterprise AI, said the company believes ZDR "should be on by default."

Sensitive Industries Move Toward Private Deployment

These data protection demands are even more pronounced in defense and energy sectors, where some enterprises are no longer satisfied with negotiating retention terms with AI vendors, instead seeking to control the model operating environment itself.

Booz Allen Hamilton has prohibited employees from using Fable in work involving proprietary cybersecurity software. A major U.S. utility company abandoned testing Fable for core power infrastructure after Anthropic refused to provide irrevocable ZDR, though it continues using Anthropic products in lower-sensitivity areas like finance and human resources.

U.S. defense technology firm Northrop Grumman has long operated open-source AI models on air-gapped servers without network connectivity, and deploys some external vendor models on its own infrastructure. Novo Nordisk allows Claude to process public materials and general information but prohibits employees from entering proprietary data.

Courting Enterprise Clients, Microsoft Strengthens AI Deployment and Governance

Enterprise demands for data control are also creating opportunities for competitors like Microsoft to win over clients. Citing sources involved in these efforts, the report indicates that Microsoft has begun lobbying OpenAI's enterprise customers to switch to its AI products, pitching at least one major client on an AI solution running entirely on private servers, with costs potentially reaching millions of dollars over several years, still under evaluation.

Simultaneously, Microsoft is strengthening governance of its own AI offerings. The company recently released interim AI conduct guidelines emphasizing that AI should always serve human objectives, should not form its own goals, and should not undermine user judgment or autonomy. Microsoft also plans to further refine model restrictions for high-risk scenarios involving cybersecurity and hazardous materials, and intends to apply updated guidelines to model development beginning in 2027.

Microsoft is enhancing enterprise AI controllability on both product and governance fronts: offering deployment methods with greater data control while reducing enterprise concerns through strengthened model safety rules. In the near term, Anthropic and OpenAI's model capabilities in areas like enterprise automation and legal research remain difficult to replace. However, as enterprises raise the bar on data retention and usage requirements, enabling companies to maintain data control while leveraging top-tier models is becoming the new competitive battleground in the enterprise AI market.