AI agents are spreading fast. Their rules are still catching up.

AI agents are moving into enterprises faster than the playbook around them. In an IDC survey released Tuesday, 86% of respondents said they already use agents embedded in applications, but only 12% said the risks tied to sovereign AI are widely understood inside their organizations.

That matters because these systems are no longer just chatbots with a nicer name. They can reach into company data, call tools, and act inside other applications. Cohere, which commissioned the survey, says the real headache is observability and guardrails: who can see what, who can touch sensitive data, and who can tell what an agent actually did.

IDC’s sample was built from 508 IT and business decision-makers at organizations with more than $1 billion in annual revenue across the U.S., Canada, the U.K. and Germany. They came from healthcare, financial services, the public sector, energy, manufacturing, and telecommunications. Nearly everyone in the group used generative AI. But only 13% said sovereign AI concepts were widely or very widely understood across their organizations, and one in three said they had trouble explaining sovereign AI in their own words.

Cohere’s answer is to push control down into the stack. The company says its full platform, including its models, retrieval-augmented generation and North’s agents and tool use, can run on customer infrastructure with no external connection, including air-gapped and classified environments. Licensing is offline, updates arrive as signed packages, and the customer decides what gets in and when.

That control comes with more work. In the survey, infrastructure was the biggest barrier to sovereign AI readiness at 17%, with cost and budget at 10% and skills and talent at 9%. And the responsibility is already landing somewhere: 420 of the 508 organizations said someone is accountable for sovereign AI, but only 8% said that role was well-defined, and just 0.4% said it was fully formalized and governed.