The gap the data keeps confirming
The theme running through this month's enterprise AI reporting is uncomfortable and consistent: the technology is producing insight while failing to produce savings. A July 27 analysis from MarketScale pulled the threads together, and the underlying surveys are blunt. A study from Domino Data Lab with BARC Research, based on 639 senior AI leaders across North America and Europe, found that 57 percent of enterprises report AI ROI failing to outpace their spend, a figure unchanged since 2025 even as production capability improved. AI is surfacing business insight and better customer interactions, and it is not delivering the cost and time savings most business cases assumed.
That stall is the story CIOs need to bring to their next budget review. Improved capability that does not convert to return is exactly the pattern that gets AI programs cut when finance tightens. The Domino data adds a hopeful wrinkle, with 93 percent reporting improved production capability in 2026 versus 88 percent a year earlier, so the plumbing is getting better. The problem sits between a working model and a business user acting on it, which is where value either lands or evaporates. The regional splits sharpen the point, with North American respondents reporting a ROI shortfall around 51 percent against roughly 67 percent in the UK and continental Europe, a spread wide enough to suggest governance maturity explains the difference more than geography does.
The budget overruns are now measurable
The spend side is where the numbers turn sharp. A WitnessAI survey of 300 business executives, reported by CFO Dive on July 22, found 68 percent of U.S. companies ran AI initiatives over budget in the past year, and 33 percent said overruns happened mostly or always. Only 9 percent said more than three-quarters of their AI initiatives delivered a measurable financial return. Just as pointed, 30 percent said unmanaged or poorly governed AI usage drove those cost overruns, and 27 percent said it led to delayed or canceled initiatives.
Those figures name a specific culprit, which is ungoverned consumption. Agentic systems scale usage and cost faster than the copilots that came before them, and without spend controls the meter runs quietly until an invoice forces the conversation. A separate CloudZero reading cited in the same coverage found 87 percent of finance leaders feel pressure to connect AI spend to business outcomes while only 22 percent have achieved it. That 65-point gap between accountability and capability is the operational reality behind every board slide claiming AI is under control.
Governance is turning into the ROI variable
The most useful finding cuts against the reflex to treat governance as overhead. In the Domino study, organizations with fully integrated AI governance were 3.9 times more likely to have governed agentic AI running in production and reported roughly 3 times faster delivery velocity than firms where governance lagged behind their AI activity. Governance maturity correlated with more agents in production, not fewer, which inverts the usual assumption that controls slow you down. "Govern early, and build the applications that turn AI into something business users can actually use," said Shawn Rogers, CEO of BARC US.
Read alongside the spend data, the mechanism is clear. Teams that instrument identity, policy, observability, and cost controls up front can scale agents without the invoice shocks and cleanup that stall ungoverned programs. Thomas Robinson, COO of Domino Data Lab, framed the shift in milestones: getting a model into production used to be the goal, and the real milestone now is the moment a business user can act on it. Governance is what makes that moment repeatable at scale instead of a one-off demo.
Why agents change the cost math
The reason this is landing now is that agentic AI has a different cost curve than the tools it replaced. A copilot answers a prompt and stops. An agent takes multi-step actions, calls tools, and can spawn work that compounds token and compute consumption without a human in the loop to notice. The Domino data showed 43 percent of enterprises have agentic AI in governed production while 41 percent are piloting or scaling it without governance, which is a large cohort running the exact configuration the overrun data punishes.
For a CIO, that split is the risk map. The 41 percent scaling agents without governance are the population most likely to generate the surprise invoices and the delayed initiatives the WitnessAI survey measured. Shadow AI compounds it, with the CFO Dive reporting noting a large share of ungoverned activity concentrated in IT and infrastructure, precisely the teams with the access to run up cost fastest. The fix is unglamorous, which is metering, budget guardrails, and policy enforcement wired in before you scale, not after the first big bill.
What CIOs and CFOs should do with this
The practical move is to stop running AI governance and AI FinOps as separate initiatives. The surveys show they are the same problem viewed from two seats, the CFO worried about return and the CIO worried about control, and the teams that integrate them are the ones shipping governed agents and hitting their delivery targets. Put cost visibility and policy enforcement in the same platform layer, tie every agent to an owner and a budget, and make governed production the default path rather than the exception a project earns after the fact.
The blunt reason to act is that the ROI figure has not moved in a year, and finance is done waiting. Enterprises are already postponing a meaningful slice of planned AI spend into 2027 as scrutiny rises, and programs that cannot show governed, measurable value are the first to get deferred. The organizations that treat governance as the lever that unlocks return, rather than the tax that delays it, are the ones whose AI budgets will survive the next planning cycle intact. Start with the 41 percent of your agent estate running without governance, since that is where both the runaway cost and the fastest available ROI improvement sit, and make governed production the condition every new agent has to meet before it touches a customer or a ledger.
