The most important benchmark for the next phase of enterprise artificial intelligence isn’t arising around the model’s intelligence.
Instead, the most important benchmark is around a model’s data retention policies. That is, if the news last week that companies as varied as Nvidia, Palantir, Booz Allen Hamilton and Novo Nordisk are drawing hard boundaries around where third-party AI models can operate is any indication.
After all, the more useful AI becomes inside an enterprise, the more sensitive its context becomes. That changes the economics of model selection. A slightly better reasoning score may be valuable, but not if accessing that capability requires a company to alter longstanding rules governing confidential information.
For CFOs, CIOs and CISOs, the relevant question therefore shifts from “Which model performs best?” to “Which model can perform best inside the controls we already have?”
Read more: The 7 AI Terms Every CFO Needs to Understand
Privacy Is Becoming a Competitive Feature for AI
Companies spent the first years of generative AI comparing models on reasoning, accuracy, latency and price. But as AI moves out of experimentation and into finance, cybersecurity, legal, engineering and other information-rich functions, another variable is moving much closer to the top of the procurement checklist: What happens to the data after the model answers?
OpenAI, for example, says qualifying API customers can use zero data retention (ZDR), and in August announced a “Private Safety Processing” approach intended to identify dangerous patterns while keeping ZDR customer content on infrastructure controlled by the customer. Google Cloud similarly documents configurations customers can use to pursue ZDR deployments on its Gemini Enterprise Agent Platform.
As for Anthropic, the provider signaled in August that it would let enterprise customers keep required safety data on their own infrastructure, as PYMNTS reported. It formally unveiled that approach Sept. 1 as Enterprise Frontier Safeguards, or EFS. Under EFS, customers store retained data in their own cloud infrastructure under their own encryption keys, while Anthropic’s automated safety monitoring continues. Anthropic says the customer’s own security team, not Anthropic staff, reviews whatever the system flags. PYMNTS reported that the safeguards will roll out in phases, with broader availability targeted for this fall.
Read also: Good CFOs Automate but Great CFOs Anticipate
“There are existing bodies of law that, while not passed or promulgated for the reason of AI, are still applicable to AI solutions,” Alaap Shah, member of the firm at Epstein Becker Green, told Competition Policy International, a PYMNTS company, in an interview this month, adding that the question for firms is not simply whether an AI-specific law applies but whether the system creates risks covered by older legal obligations.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
For an AI laboratory, retaining data can function as a safety mechanism. For a multinational corporation, the same retention can create another system containing sensitive information that must be governed. In other words, AI companies want enough visibility into model activity to detect sophisticated misuse. Their largest customers often want precisely the opposite: the smallest possible external data footprint.
The result could be a new competitive front among AI providers: not simply who builds the smartest model, but who can deliver frontier intelligence without requiring enterprises to surrender control of the underlying data.
Read more: Is the Enterprise Insurance Market Agentic AI’s Hidden Ceiling?
The Office of the CFO Has to Answer a New AI Question
The cost-benefit calculation is no longer simply whether an AI system saves enough labor to justify its license and infrastructure costs. Finance leaders also have to consider what deploying that system does to the company’s risk surface.
A PYMNTS Intelligence report, “How Agentic AI Went From Zero to CFO Test Runs in 90 Days,” found in September 2025 that nearly 7% of enterprise CFOs in the United States had deployed agentic AI in live finance workflows, while an additional 5% were running pilots. And in a recent PYMNTS eBook, executives from Visa, FIS, Synchrony, WEX, Billtrust, i2c, Thales, Velera, Bottomline and other industry leaders describe what they are learning as AI agents move from demonstrations into real operating environments.
But every additional copy of sensitive information creates something that potentially must be secured, governed, audited and eventually deleted. An AI tool that produces better answers but requires additional data retention can therefore carry costs that do not appear in its per-token pricing.
PYMNTS Intelligence’s April “The Enterprise AI Benchmark Report” showed that 71% of executives at companies with at least $1 billion in annual revenue said organizational readiness is the chief limitation on AI performance. Only 11% said they think AI technology itself is the primary barrier.
For all PYMNTS digital transformation and B2B coverage, subscribe to the daily Digital Transformation and B2B Newsletters.
Add as Preferred Source
