AI adoption now measured by what you can audit

The main case was announced on September 12, when Cloudera and Mistral revealed a sovereign enterprise AI alliance. The trend, rather than a unique event, emphasizes open and customizable models within a hybrid data platform, with data, models, and compute under customer control. Two days earlier, NeuroWatt unveiled NeuroTeam, an enterprise agentic AI workforce that bundles agent reasoning, tool execution, identity, access, policies, human approvals, auditability, and monitoring into a unified architecture.

This architecture includes multi-model routing via an internal LLM gateway and an on-prem option for sovereignty and latency.

Both announcements share a common trend: the model is no longer the central selling point. Instead, the focus is on the wrapper that ensures auditable and manageable AI. A concerning statistic emerges: 92% of respondents deem governing agents critical for security, yet only 44% have implemented policies for this governance. Only 18% of MCP server deployments incorporate any scope to tool permissions, and only 52% can trace and audit the data their agents access.

An MCP server, for those unfamiliar, is the component that exposes tools and data to an AI agent. Without scope, an agent could call any tool published by the server, with no limitations. The Cloud Security Alliance's Agent Identity Governance Framework suggests treating agents as first-class identity subjects, complete with just-in-time access, expiration, and human sponsorship. However, rotating credentials is crucial to prevent abandoned accounts with permanent access.

Gartner advises against applying the same governance to all agents, proposing four levels of autonomy with varying controls. They caution that human approvals can lead to fatigue and a false sense of security. Despite these concerns, the human-in-the-loop approach remains essential.

Regulatory measures are also in place. Article 50 of the AI Act's transparency obligations, effective August 2, along with the AI Office's sanctioning powers, could result in fines of up to 15 million euros or 3% of global turnover. Palo Alto documents the "rug-pull" attack, where an MCP server, after passing initial review, silently changes its tool definition. Fines of around 47 million euros have reportedly been issued within weeks, though this figure is unconfirmed as an official source.

The pattern is clear: the AI that gains traction in 2026 will be the one equipped with identity, scope, logging, and built-in safeguards. Everything else is still a pilot or less. However, the lack of comprehensive agent inventory and governance presents significant challenges. Prioritizing inventory and scope before implementing policies is paramount, as it allows organizations to address potential vulnerabilities before they become issues.