Two major studies dropped this week with the same uncomfortable message: enterprise AI production capability is at an all-time high, and ROI is still broken. Nearly seven in ten companies (68%) say AI initiatives ran over budget last year. Only 9% of organizations report that more than three-quarters of their AI projects delivered measurable financial returns. The technology is working. The business model around it isn't.
This isn't a technology problem. It's a governance problem. And the data makes that gap impossible to ignore.
WitnessAI surveyed 300 U.S. business executives for their "Hidden Cost of Enterprise AI" report, released this week. On the same day, Domino Data Lab published their Fifth Annual Enterprise AI Report drawing on 639 senior AI leaders globally. Together they paint a consistent picture: enterprises are deploying AI faster than they can govern it, and the financial consequences are materializing in real time.
If you're a CIO, CFO, or business leader trying to justify next year's AI budget, you need to understand exactly what's happening here—and what the organizations succeeding are doing differently.
The Numbers Are Getting Harder to Ignore
Start with the ROI stagnation. Domino Data Lab found that 57% of enterprises still report AI ROI that fails to outpace their investment. That's the exact same number as 2025. Two years in a row, no movement. Meanwhile, 93% of those same organizations say their AI production capability improved this year, up from 88% in 2025.
Think about what that means. More AI in production than ever before. Same broken ROI.
The WitnessAI data adds urgency on the cost side. One-third of respondents (33%) say AI projects were always or mostly over budget in the past 12 months. And it's not just budget variance—43% of enterprise executives reported $2 million or more in total costs from AI-related security incidents in the past year. Twenty-one percent say their single most significant AI incident cost $1 million or more. Seventeen percent say annual AI incident costs exceeded $10 million.
These aren't rounding errors. These are material financial events that boards need to know about.
The CloudZero research adds context: 87% of finance leaders feel pressure to connect AI spending to measurable business outcomes within the next year. Only 22% have actually achieved that goal.
That gap—87% under pressure, 22% succeeding—is where most AI programs are trapped right now.
Why Production Success Doesn't Translate to Business Value
Here's what's confusing the picture: AI teams are genuinely improving. More models in production, faster deployment, expanding use cases. But getting a model into production used to be the milestone that mattered. It isn't anymore.
The real milestone is the moment a business user can act on what the model found. And that moment still isn't happening at the pace or scale that business demands.
Domino Data Lab found that 34% of organizations use a mix of AI access methods that varies by business unit—no consistent delivery layer, no standard experience. Another 40% still rely entirely on mediated access: a scheduled report from a data science team, or requests submitted to an analyst who runs the analysis and returns results. Business users are waiting in line. The value is sitting in production, unactualized.
Part of the difficulty is that AI ROI rarely surfaces as a single number. It appears as fragments scattered across budgets, productivity reports, and individual business units. The WitnessAI report describes CFOs as driving further AI investment without a unified, accurate picture of whether existing initiatives are paying off. That's not a CFO failure—it's a structural visibility failure. Nobody built the financial instrumentation to track AI spend and return across the organization.
Shadow AI Is Where Your Money Is Actually Going
The governance gap has a specific shape, and it starts where you'd least expect it.
The WitnessAI report found that IT and infrastructure departments are the largest single source of shadow AI activity at 47%. Sales and business development came in at 34%, marketing at 33%.
Let that sink in. The department most responsible for governing AI use within the organization is also the department most likely to be operating outside its own policies.
Thirty percent of respondents directly attributed cost overruns to unmanaged or poorly governed AI usage. Another 27% said poorly governed AI caused delayed or canceled AI initiatives—money spent, initiatives killed, nothing to show for it.
Meanwhile, agentic AI is scaling faster than anyone expected. The percentage of organizations orchestrating multiple AI agents across workflows doubled from 9% to 18% in Q2 alone, according to a June KPMG report. WitnessAI found that 70% of enterprises are already using or piloting AI agents capable of autonomous actions. But fewer than one in five (18%) report that all agents are formally inventoried and approved by their security team.
Autonomous agents making decisions with incomplete inventories and no approval process. That's the combination that creates $2 million security incidents.
The C-Suite Has a Confidence Problem
There's a specific dynamic that keeps this problem from getting fixed: leaders at the top believe things are under control. The people doing the actual work don't.
WitnessAI found that 68% of C-suite executives report having full confidence in their visibility into AI tools and agents. Only 46% of VPs executing deployments said the same.
That 22-point gap is dangerous. When the people accountable for AI governance don't feel the urgency that frontline AI teams feel, interventions get delayed, budgets don't get reallocated, and the cost overruns continue.
The reporting structures that would close this gap—AI-specific financial modeling, unified spend visibility, real-time incident tracking—are often either incomplete or absent. WitnessAI found that only 46% of CFOs actively model AI-specific risk and ROI. Another 38% say finance reviews AI spending but doesn't separately model AI-related risk exposure.
In conversations with finance and technology leaders, this is exactly what I hear: the CFO dashboard shows total software spend, but AI is buried inside it. Nobody can tell you which AI initiatives are delivering and which are burning cash.
Governance Is the Actual Differentiator
Here's where the Domino Data Lab data gets actionable. They found that governance maturity is the clearest dividing line between organizations getting value from AI and those stuck in the ROI plateau.
Among organizations whose governance is fully keeping pace with their AI activity, 67.5% have agentic AI running in governed production. Among organizations where governance is only partially keeping pace, that figure drops to 17.2%. That's a 3.9x difference based entirely on governance maturity.
The velocity numbers are even more striking. Among organizations with fully integrated AI governance, 75% report significantly improved AI delivery velocity. That's more than three times the rate among organizations where governance is falling behind (23%).
This matters because it breaks a common misconception: that governance slows things down. The data shows the opposite. Governance enables faster deployment because teams aren't spending time managing incidents, cleaning up unauthorized tools, or explaining budget overruns to finance leadership.
The organizations proving this are in financial services. Among the heavily regulated verticals in Domino's survey, banks, insurers, and financial services firms lead every industry on both governance maturity and production velocity. They built governance infrastructure first, then scaled. The enterprises succeeding today are following the same playbook.
What the Top Performers Do Differently
The pattern across both studies is consistent. Organizations beating the ROI plateau share four characteristics.
They inventory everything before deploying at scale. The WitnessAI finding that only 18% of organizations have all AI agents formally inventoried and approved isn't just a security gap—it's a financial control gap. You can't optimize spend you can't see. Comprehensive agent inventories are a prerequisite for financial accountability.
They assign ownership before incidents happen. The survey found that accountability for agentic AI is genuinely unclear at most organizations. Thirty percent identify the CIO as primarily responsible for managing AI risk, while only 15% say the CISO. Six percent say the CISO is primarily liable when an AI agent causes financial or regulatory harm. Twenty-six percent say it's the CIO. Ambiguous liability structures produce slow incident response and diffused accountability. Top performers define this before deployment.
They connect AI finance to business finance. The top performers have CFOs who actively model AI-specific risk and ROI, separate from general software spend. That means AI investment decisions are made with the same financial discipline as capital expenditures—not as technology experiments that report back if they feel like it.
They treat the last mile as the primary problem. For organizations stuck in the ROI plateau, the issue isn't model performance. It's access. Business users can't act on AI-generated insights because access methods are inconsistent, mediated, or buried. The fix is building direct business-user access layers—dashboards, APIs, embedded AI in existing workflows—not more models in production.
They govern agentic AI before it governs them. Domino found that 41% of organizations are piloting or scaling agentic AI today without the governance to manage it. With agents scaling outnumbering pilots by more than two to one, the window to build governance ahead of deployment is closing fast. Organizations that govern early move faster later.
What This Means for Leaders Right Now
If you're a CIO or CTO: your production metrics may look good. But if business users still require mediated access to AI insights—scheduled reports, analyst requests—your ROI will stay stuck. The infrastructure problem to solve isn't model performance. It's last-mile delivery. Build the access layer.
If you're a CFO: 87% of finance leaders are feeling the same pressure you are. The 22% who are succeeding aren't smarter—they built separate AI cost modeling earlier. Treat AI spend as its own P&L center, not a line item in software. You can't govern what you can't see in the budget.
If you're a CISO or Head of AI Governance: the shadow AI problem starts in IT, not in sales or marketing. That's where unauthorized tools proliferate and where ungoverned agents are most likely to operate. Start your audit there.
If you're a CEO or board member: the 22-point confidence gap between C-suite and VP-level leaders on AI visibility is a board-level risk. Consider commissioning an independent AI inventory and governance assessment. The $2M+ incident costs in the WitnessAI data don't just affect operations—they create regulatory exposure that can reach 3-5% of annual revenue.
The Bottom Line
Enterprise AI has a production problem. The models work. The ROI measurement, governance, and last-mile delivery systems don't.
The organizations closing the gap share one thing: they built governance infrastructure before scaling. Financial services firms lead not because they're more innovative, but because regulation forced them to instrument their AI programs properly from the start.
The rest of the enterprise world is getting there the hard way—through cost overruns, ungoverned agents, and ROI that still isn't outpacing spend two years into the scaling era.
The good news is the playbook exists. The financial impact of getting this right—3.9x better agentic deployment rates, 75% faster AI delivery velocity—is measurable. The cost of not getting it right is increasingly measurable too.
Govern early. Build the access layer. Connect AI to the CFO's dashboard. The $2 million incident is the thing you can't afford to learn from.
Sources: WitnessAI "Hidden Cost of Enterprise AI" report (July 2026, 300 U.S. business executives); Domino Data Lab Fifth Annual Enterprise AI Report (2026, 639 senior AI leaders); KPMG Q2 AI Pulse (June 2026); CloudZero Finance Leaders Survey (June 2026).
