Enterprise AI Control Gap Leaves 90% of Firms Flying Blind

Enterprises expected AI governance to mature as deployments grew. The enterprise AI control gap shows the opposite: companies have standardized AI ambition before they standardized AI accountability.

A new VentureBeat Pulse Research survey of 145 qualified enterprise respondents found that AI programs are expanding faster than organizations can see, monitor, fund, or govern them, according to VentureBeat. The survey covers organizations with 100 or more employees and should be read as a directional signal, not a probability sample.

The sharpest finding is not that models are failing. It is that no one clearly owns the full AI stack. 58% of respondents are net-adding AI initiatives, while 85% run two or more platforms claiming to be the “primary” AI layer. Only 10% have active monitoring and alerting for production model failure.

The machinery to expand AI is running ahead of the machinery to control it.

That is the enterprise AI control gap in one sentence.

Enterprise AI expansion has outrun the control layer

The survey shows a clear split between growth and discipline. 33% of enterprises say they are “expanding significantly,” and another 25% report “net positive growth.” Combined, 58% are adding AI initiatives.

Restraint exists, but it is not the dominant posture:

• Rationalizing: 23% are scaling what works and cutting the rest.
• Holding flat: 12% are not expanding portfolios.
• Pausing for governance: Only 3% stopped to get control structures in order first.
• Unable to describe the portfolio: 4% could not classify their own AI footprint.

That last number matters more than it looks. If an organization cannot describe its AI portfolio, it cannot reliably assign ownership, measure cost, or detect failure.

The platform picture is even messier. 85% of respondents have at least two enterprise platforms claiming to be the primary AI layer. 36% report a four-way-or-more contest. Only 8% have consolidated to a single layer.

XOOMAR analysis: this is not just vendor overlap. Each added AI platform brings its own controls, permissions, telemetry, cost model, and failure modes. That turns governance into a coordination problem before it becomes a tooling problem.

For readers tracking AI control from the model-access side, XOOMAR has also covered related pressure points in Trump Drops Anthropic Export Controls After AI Lockout and Anthropic Fable 5 Roars Back After U.S. AI Freeze Ends. VentureBeat’s survey points to a different layer of risk: what happens inside enterprises after access expands.

The enterprise AI control gap starts with the missing owner

The most important control surface may now be the org chart.

Only 38% of respondents say a central team governs AI today. That is the leading answer, but it falls well short of a majority. The rest of the responses show fragmentation:

• Unclear or contested ownership: 21%
• Each platform team governs independently: 20%
• No one has addressed it: 19%

Role-level accountability is just as scattered. CIO/CTO/CISO leads at 27%. A Chief AI Officer or equivalent accounts for 22%. A striking 17% say no role holds formal accountability.

The barrier data sharpens the point. The single most-cited obstacle to governing AI across multiple platforms is the absence of a single accountable owner, cited by 32%. Vendor opacity follows at 25%, lack of tooling or infrastructure at 16%, leadership deprioritization at 17%, and lack of talent at 5%.

XOOMAR analysis: no monitoring platform can fully compensate for broken decision rights. If teams disagree on who can approve a model, shut down an agent, cap spend, or overrule a platform owner, observability becomes a dashboard without authority.

That is why the enterprise AI control gap is an ownership problem first. Technology can expose drift, spend, and policy violations. It cannot decide who has the mandate to act.

Manual monitoring breaks when AI moves into production

The survey’s confidence numbers look better until they are unpacked.

40% of enterprises say they are very confident they would detect a model drifting, behaving unsafely, or failing in production. But only 10% support that confidence with active monitoring and alerting. The other 30% rely on manual human review.

That is a weak fit for systems that can run across workflows and longer execution windows.

The reactive tail is also large. 19% say they would hear about failures from end users first. 8% report no systematic visibility. Another 32% expect to “catch most issues eventually.”

Before and after the AI shift, the control problem looks different:

• Before: Teams reviewed software behavior through tickets, logs, audits, and periodic checks.
• After: AI systems can drift, behave unsafely, fail tasks, and trigger actions faster than manual review can catch.
• Before: Cost overruns often appeared as procurement or infrastructure issues.
• After: Autonomous agents can generate runaway bills or operational failures during execution.
• Before: Shadow IT usually meant unsanctioned tools.
• After: Shadow AI can mean unauthorized agentic pipelines outside central oversight.

The source does not say every AI system can modify production systems or spend money. But it does document that autonomous agents are already producing financial and operational control failures. That changes the risk profile.

Fine-tuning disappointment pushed enterprises toward hedging

The survey also captures the aftershock from custom model investment.

Asked about proprietary foundation models fine-tuned over the past 18 months, 73% of respondents either failed to get custom models into productive use or deliberately avoided the effort. The largest group, 45%, described a “sandbox graveyard,” with projects too expensive or complex to maintain. Another 24% never started because they priced in the downstream maintenance burden. 27% said fine-tuned models are a reliable advantage.

That helps explain the vendor posture. 51% use a hybrid mix of open-weight and closed models. 32% remain deliberately committed to closed models, while 16% are making a hard pivot to self-hosted open models.

Vendor trimming is also showing up. Microsoft is the most-named target for phase-out or cutback at 29%, followed by OpenAI at 21%, Anthropic at 15%, and Google at 6%. 27% are downsizing no vendor at all.

XOOMAR analysis: the survey does not prove a mass vendor exodus. It shows a more practical pattern. Enterprises are keeping options open, cutting where ROI or control disappoints, and avoiding custom work unless the production payoff is clear.

Shadow AI has turned governance into a finance problem

The clearest evidence that the enterprise AI control gap is already costing money comes from autonomous agents.

49% of respondents cite shadow AI, described in the source as unauthorized agentic pipelines run on corporate cards outside central oversight, as their most severe financial or operational control failure. Another 25% report a runaway “infinite loop” agent bill. 6% cite an agent that degraded production databases.

Only 21% report guarded stability, meaning they imposed hard token throttling and budget caps at the infrastructure layer and avoided surprises.

That means roughly 79% have already experienced a real financial or operational control failure from autonomous AI.

This is where finance leaders enter the story. AI governance is no longer only about model accuracy or policy compliance. It is also about uncontrolled spend, unapproved pipelines, and the lack of budget enforcement at execution time.

XOOMAR analysis: the organizations in the strongest position are not necessarily the ones with the most AI activity. They are the ones that can answer basic operating questions: which agents are running, who approved them, what they can access, how much they can spend, and who can stop them.

The next AI operating model starts with authority, not another pilot

The survey points to a practical sequence for enterprise leaders.

First, assign a real owner for AI behavior across platforms. Second, map every platform claiming AI primacy. Third, require active monitoring and alerting for production systems. Fourth, enforce budget caps, token throttling, permissions, and shutdown paths before autonomous agents get longer execution windows.

The likely demand signal is clear: enterprises will want cross-platform control planes that track cost, model choice, drift, permissions, agent behavior, and policy enforcement from one layer. The source’s free-text responses also converged on a single accountable owner and a control plane that abstracts cost, drift, and model choice away from the end user.

The evidence that would confirm this thesis: rising adoption of centralized monitoring, fewer organizations relying on manual review, and a lower share reporting shadow AI or runaway agent bills in future surveys.

The evidence that would weaken it: enterprises consolidating naturally around one primary AI layer, or platform teams proving they can govern independently without producing new blind spots.

Until then, the warning is simple. Companies that do not assign ownership for AI behavior will still learn when systems fail. They will learn from users, invoices, degraded systems, or audit demands. That is not governance. That is discovery after damage.

Impact Analysis

• AI adoption is accelerating faster than enterprise governance structures can keep up.
• The survey suggests ownership and accountability, not model performance, are the core enterprise AI weaknesses.
• Low production monitoring raises operational risk as more companies deploy AI across multiple platforms.