Companies Are Losing Control of AI Spending, WitnessAI Report Finds

Companies around the world are ramping up AI spending faster than they can manage it. A new report from WitnessAI, a firm specializing in enterprise AI security and governance, puts numbers to the scale of the problem: 68 percent of large organizations admit that at least some of their AI projects have gone over budget in the past year.

The WitnessAI survey, run by a company founded by former Zscaler and Palo Alto Networks executive Rick Caccia, polled 300 decision-makers at large organizations, ranging from vice presidents to board members. Respondents came from companies with more than a thousand employees, making it one of the more representative snapshots of how corporations have handled the rapid rollout of AI over the past year.

Budgets Spiraling Out of Control

The key figure in the report is 68 percent, the share of companies admitting that at least some of their AI initiatives have exceeded their planned budget. One in three of them say overruns happen regularly rather than occasionally. At the same time, 30 percent of respondents directly link these overruns to a lack of proper oversight of how AI is used within the organization, and 27 percent admit that poor governance has led to projects being delayed or scrapped entirely.

The problem isn't limited to deployment costs alone. Pressure to demonstrate return on investment is rising in step with spending, analysts have been warning for months that the market is starting to expect concrete financial results rather than just declarations about digital transformation.

Where the Returns Disappear

The most troubling finding in the survey is the gap between the scale of investment and its results. Only 9 percent of respondents said that more than three-quarters of their AI projects delivered measurable, quantifiable financial returns. By comparison, 91 percent of companies worry that AI agents increase their organization's financial risk, yet 64 percent still believe the benefits of agentic AI outweigh the risk enough to keep deploying it.

This paradox, being aware of the risk while accelerating deployment at the same time, is, according to the report's authors, the main source of the problems. Companies aren't abandoning AI despite rising costs, they're instead trying to catch up on control procedures after the fact, once systems are already running in production.

Agents Running Without Oversight

The report pays particular attention to autonomous AI agents, systems capable of carrying out multi-step tasks on their own without constant human supervision. As many as 70 percent of surveyed companies are already using such agents or piloting their deployment, but only 18 percent have a complete, formal inventory of these systems approved by their security team.

That means AI agents at most large organizations are operating beyond the full control of the teams responsible for security and compliance. The financial consequences of these gaps can be severe: 43 percent of companies reported at least $2 million in combined annual costs from AI-related incidents, and 21 percent estimated the cost of their single most serious incident at $1 million or more. As many as 86 percent of companies investigated at least one AI-related incident over the past year.

The shift to agentic AI introduces a completely new level of financial and operational liability, and individual incidents now run into the millions of dollars - Rick Caccia, CEO and co-founder of WitnessAI

What This Means for Companies in Poland

The WitnessAI survey covers US enterprises, but the dynamic it describes is universal and familiar to Polish companies rolling out AI as well: pressure for fast deployment outpaces the buildup of oversight procedures, tool inventories, and clear lines of accountability. Similar signals have already shown up in Polish studies on so-called shadow AI, employees using AI tools without their managers' knowledge or approval.

For businesses, that translates into a concrete recommendation: before another department starts piloting an AI agent for customer service, sales, or data analysis, it's worth having a ready list of questions about who is accountable for its actions, what data it processes, and how quickly it can be shut down if something goes wrong. The WitnessAI report shows that companies that skip this step end up paying for it later in cash, not just in reputation.

The report's authors stress that the problem won't simply disappear as the technology matures. If anything, the more companies adopt agentic AI capable of acting on its own, the wider the gap grows between the pace of deployment and the pace of building real oversight over it. According to the survey, it's this gap, not the technology itself, that generates the biggest costs.