Three in Four Enterprises Have No Oversight of Their AI Agents

Seventy-five per cent of EMEA enterprises have no clear oversight of the AI agents handling their sensitive data. Seventy per cent admit that automated AI workflows are touching sensitive corporate information without full visibility into what those processes are doing or what they have changed. And 67 per cent of IT departments report that employees are standing up autonomous AI agent workflows that the IT function cannot track.

These numbers come from a new survey commissioned by Veeam, the data resilience and AI trust company, and conducted by Censuswide across EMEA enterprises. Published in mid-September 2026, the research puts hard figures on a problem that compliance teams, CISOs and general counsel have been raising quietly for months: the "shadow agent" phenomenon has already arrived, and it is moving faster than governance can follow.

The term matters. Shadow IT was about employees buying software outside procurement. Shadow agents are about employees deploying autonomous digital workers that can read, write and alter enterprise data without any central record of what they have done. The difference is the difference between an unsanctioned spreadsheet and an unsanctioned employee with database access and no manager.

What the data shows

Veeam's survey of EMEA IT and business leaders surfaces five findings that together paint a picture of an enterprise governance gap widening in real time.

First, the oversight gap is structural. Three-quarters of enterprises lack clear visibility into what their AI agents are accessing, processing or modifying. Among those, 70 per cent acknowledge that automated AI processes interact with sensitive corporate data without full supervision. These are not edge cases. They are the norm.

Second, the IT department is losing track. Sixty-seven per cent of IT teams say employees are creating autonomous AI workflows that IT cannot monitor. The speed at which individuals can stand up agentic workflows, often through consumer-grade tools connected to enterprise accounts, has outpaced every conventional IT control. A marketing manager can wire an AI agent to a CRM, pull customer data into a third-party model, analyse it and generate a campaign in the time it takes IT to schedule a review meeting.

Third, regulatory pressure is intensifying. Fifty-eight per cent of surveyed enterprises are now subject to corporate accountability laws that impose personal legal responsibility on senior executives for cyber resilience and data compliance. One in eight say the allocation of individual responsibilities under those laws is unclear inside their own organisations. The EU AI Act, GDPR, NIS2 and DORA each impose overlapping obligations around data governance, algorithmic transparency and operational resilience. An AI agent operating outside visibility is by definition operating outside compliance.

Fourth, the personal stakes are rising. Forty per cent of leaders say they personally fear legal consequences from AI non-compliance. Nearly a third of organisations report that regulatory scrutiny is already generating boardroom tension or conflict among executives. This is not a theoretical risk. Under the EU AI Act, which entered its enforcement phase in August 2026, penalties for non-compliance can reach the higher of EUR 35 million or 7 per cent of global annual turnover.

Fifth, the readiness gap is severe. Separate data from Veeam's broader Data and AI Trust Gap report, also released in September, found that 88 per cent of enterprises are already using or piloting AI agents, yet only 7 per cent consider themselves adequately prepared for the governance, security and compliance demands those deployments create. The gap between deployment velocity and governance maturity is not closing. It is accelerating.

Why this is different from shadow IT

The enterprise has seen this pattern before. Shadow IT arrived when cloud services became cheap, easy and invisible to procurement. The response took a decade: discovery tools, cloud access security brokers, SaaS management platforms and, eventually, a grudging acceptance that the genie was not going back in the bottle.

Shadow agents share the same root cause: technology that is useful, frictionless and available to anyone with a corporate email address. But the risk surface is an order of magnitude larger. A shadow SaaS tool stores data. A shadow agent processes it, transforms it, shares it with an external model, makes decisions based on it and, in many cases, writes results back into production systems. Where shadow IT leaked documents, shadow agents can rewrite them. Where shadow IT created a data silo, shadow agents create a decision silo: a workflow whose inputs, reasoning and outputs are invisible to the organisation that bears legal responsibility for the outcome.

Tim Pfaelzer, General Manager and Senior Vice President for EMEA at Veeam, framed the issue in terms of alignment. "While the increased stress and tension are completely understandable," he said, "if the result is better alignment at the board level, the organisation will be far better positioned to trust its data and AI governance in the agentic era. With greater alignment ultimately comes better data governance, visibility and protection across the organisation."

How enterprises are responding

The survey reveals two reactive strategies. Forty-one per cent of enterprises are building local or sovereign AI models specifically to keep sensitive data under tighter control. Forty-nine per cent are deploying hybrid architectures that isolate sensitive workloads from general-purpose AI infrastructure.

Both approaches address the symptom rather than the cause. Sovereign models reduce the risk of data exfiltration to external providers but do nothing to track what agents are doing internally. Hybrid isolation limits blast radius but does not provide visibility. The missing piece in both strategies is runtime governance: the ability to see what every agent in the enterprise is doing, to verify its identity and intent at the point of action, and to intervene when it exceeds its authorised scope.

The market is beginning to respond to that need. In the past month alone, Broadcom launched AgentMinder, a runtime governance platform that treats agents as first-class identity principals. Microsoft open-sourced its Agent Governance Toolkit, covering the full OWASP Agentic Top 10. At AGNTCon Europe in Amsterdam last week, governance startups dominated the exhibition floor. The common thread is that control is no longer an afterthought bolted onto agent deployment. It is the starting point for the next phase of enterprise adoption.

What it means for enterprise buyers

The Veeam data carries three practical implications for organisations that are deploying or evaluating agentic AI.

Discovery comes first. Before an enterprise can govern its agents, it must know they exist. The 67 per cent of IT departments reporting untracked autonomous workflows need agent discovery tools that work across SaaS platforms, cloud services, local development environments and consumer-grade AI products with enterprise access. Without a complete inventory, every other governance investment rests on incomplete data.

Governance must be real-time, not retrospective. Audit logs tell you what happened after the fact. Runtime governance tells you what is happening now and blocks what should not happen at all. For enterprises in regulated sectors, the distinction is the difference between compliance and a regulatory finding. Buyers should treat real-time policy enforcement as a procurement requirement, not a nice to have.

Executive accountability is personal. With 40 per cent of leaders fearing personal legal consequences and 32 per cent reporting boardroom tension, AI governance is no longer an IT concern. It is a board-level risk management issue. The EU AI Act, GDPR fines up to 4 per cent of turnover and NIS2 obligations around supply chain security mean that ignorance of what an agent did is not a defence. It is an aggravating factor.

What it means for suppliers

For companies building, selling or integrating agentic AI platforms, the Veeam data is both a market signal and a product requirement.

Governance is a sales differentiator. In a market where 75 per cent of buyers have no agent oversight, the vendor that ships with built-in visibility, identity, policy enforcement and audit trails is selling safety, not just capability. Procurement teams in regulated industries will increasingly filter vendor shortlists by governance maturity before they evaluate functionality.

Integration with existing identity and compliance stacks matters. Organisations have invested heavily in identity providers, security information and event management platforms, and cloud access security brokers. An agent platform that requires a parallel governance stack adds cost and complexity. One that integrates with what is already in place removes friction from the purchase decision.

The EMEA market is a governance bellwether. The EU AI Act, combined with GDPR, NIS2 and DORA, creates a regulatory environment that is stricter and more enforceable than anywhere else in the world. Products that pass the EMEA governance bar will find the rest of the world a lighter lift. For suppliers targeting global enterprise accounts, getting EMEA right first is commercially rational.

The Agentic Expo angle

Agentic Expo 2027, taking place at Olympia London on 23-24 March 2027, sits at the centre of this conversation. The event exists because the distance between agent capability and enterprise trust is where meaningful commercial value is created, and where the most important purchasing decisions in the agentic AI market will be made between now and March 2027.

For buyers, the Veeam data confirms that governance cannot be deferred until after deployment. The question to ask of every exhibitor and speaker is the same: how does your product help me see what my agents are doing, prove it to a regulator, and stop it before it causes harm? Agentic Expo is the forum where those answers can be evaluated side by side, not through a sales pitch but through live demonstration and peer conversation.

For suppliers, the data is a product roadmap in survey form. Seventy-five per cent of your target buyers have a governance gap. Forty per cent of their leaders are personally worried about legal exposure. Sixty-seven per cent of their IT departments are already losing sight of agent workflows. The vendors who close those gaps are the ones who close the deals. Agentic Expo is where those conversations turn into commercial relationships.