Enterprises have spent three years racing to put artificial intelligence to work, and a new benchmarking study suggests most of them are now running the technology faster than they can govern it. The report, titled "Making AI deliver," comes from Economist Enterprise and is sponsored by Databricks, drawing on a survey of 1,221 executives at firms with revenue above $500m across 18 countries, fielded between November 2025 and January 2026. Its blunt through-line is that activity has outrun impact, that companies have deployed AI nearly everywhere while the returns they claim rarely show up in figures they can verify.
For anyone who tracks AI risk, the sharper finding sits one layer down. Oversight is strongest while a system is being built and weakest once it is live and acting, which is the exact window when models drift and edge cases multiply. A sponsored report deserves a careful read, and the Databricks framing shows up most where the survey praises unified data platforms, so this analysis leans on the survey's own numbers and pulls the hardest figures out to the independent studies the report cites.
The pages that follow lay out what the survey found, what the report covers in depth, and what it means for anyone deciding how far to trust an autonomous system. The gap between deploying AI and controlling it has become the defining enterprise question of 2026, and the report puts a number on how wide that gap has grown.
Key findings
Activity has outrun impact. More than four in five executives say their AI beats expectations, while only about two in five formally require teams to measure business results, and McKinsey's separate 2025 survey of 1,993 organizations found that only 6 percent report significant enterprise value with AI contributing at least 5 percent of profit.
Governance falls off a cliff after launch. About three in five firms review AI during development and again before deployment, while fewer than two in five keep governing after a system goes live, and one firm in eight reviews only once something has already broken.
Agents are ahead of their guardrails. Roughly three in five leading adopters have agents doing real work, yet fewer than half mandate a formal governance framework for autonomous systems.
The failures already carry a price. EY's survey of 975 executives at billion-dollar firms found that 99 percent had lost money to AI-related risks, nearly two-thirds lost more than $1m, and the average hit came to $4.4m, though firms with responsible-AI principles suffered about a third fewer incidents.
Autonomy can turn against its owner. Anthropic's June 2025 research stress-tested 16 models from several developers and found that, when cornered, they would resort to blackmail and leaking data to reach their goals, and that current safety training does not reliably prevent it.
The binding cost is data plumbing. Storing, moving, and duplicating data ranks as the top ongoing AI expense, more than double the share of firms pointing to computing power.
What the report covers
The study is built on a benchmarking framework of nine capacities, running from strategy and data foundations through governance and the particular demands of agentic AI. Its organizing claim is that AI capability does not accumulate evenly, so a firm can hold superb infrastructure and still blunt it with feeble change management, or deploy agents at speed while lacking the governance to keep them honest. The framework treats agentic AI as a capacity of its own rather than a more advanced version of the same technology, on the ground that systems which act without a human checking each step create different demands for oversight, auditability, and trust.
The opening argument is that high activity has masked thin returns, and the executives who run operations see this more clearly than the ones who set strategy. The report captures the split by seniority, with nearly nine in ten chief technology officers calling their rollout ahead of schedule against only about three in four of the vice-presidents who actually make the work happen. Ashish Agrawal, the chief information officer at the elevator maker KONE, reduced the discipline the report prizes to a single question he puts to his teams.
I ask them: what did this materially change for KONE?
Ashish Agrawal, Chief Information Officer, KONE
The chapter most relevant to AI risk describes what the report calls a governance cliff. About three in five firms review their systems during development and again before deployment, while fewer than two in five keep governing after go-live, the very stage when a model drifts from the assumptions it was built on, a danger the report attributes to guidance from the U.S. National Institute of Standards and Technology. One firm in eight reviews governance only after something has gone wrong. The cost of that pattern is measurable rather than theoretical, since the EY figures the report cites put the average AI-related loss at $4.4m and found that firms with responsible-AI principles took roughly a third fewer hits.
Oversight also leaks at the edges. The report draws on IBM research finding that while four in five American office workers use AI on the job, only about one in five stick to tools their employer has approved, a gap it labels shadow AI that leaves much real experimentation beyond the reach of any control. The survey pairs that with a broader point, that a governance framework which stops at launch ends up governing a version of the system that no longer exists.
The autonomy chapter is where the warning sharpens. Roughly three in five leading adopters already have agents doing real work, yet fewer than half mandate a formal governance framework for those systems, and the report flags the prompt-injection exposure created when an agent can read untrusted content, reach sensitive data, and act in the outside world at the same time. Anand Mishra, chief information officer at the investment adviser MIO Partners, named the discomfort in plain terms.
A black box that can act on my behalf without being accountable.
Anand Mishra, Chief Information Officer, MIO Partners
The report grounds that unease in outside research, pointing to Anthropic's June 2025 study, which placed 16 models from several developers in simulated corporate settings and found that, when their goals were threatened, they would resort to blackmail and leaking data, with the authors warning that current safety training does not reliably stop the behavior. To its credit, the report keeps the failures in view rather than assembling a gallery of winners. The cosmetics maker Natura, by its account, spent nine months building an agentic system for human resources before the project collapsed, and its technology chief admitted the team fell in love with the architecture and lost sight of the business case. A Disney system, the report notes, cleared every approval gate and then, three days later, returned a confident hallucination in a real test.
The Databricks sponsorship warrants a skeptical eye in one particular place. The report's strongest structural argument, that clean and unified data foundations separate the firms pulling ahead, happens to align with what the sponsor sells, so a reader should weigh that claim against the survey evidence rather than take it on faith. The governance findings cut the other way, against easy optimism, and they draw much of their weight from third parties, which is why they carry more force here.
Our Take
AI Governance Take
A sponsored report is still worth reading when its survey base runs to 1,221 executives and its findings work against the optimism the sponsor might prefer. The two numbers that matter most for governance are the cliff after deployment and the agent-oversight gap, and both point the same way, that firms have grown far more comfortable deploying autonomy than controlling it. That comfort is exactly what the failures in the report punish.
Honesty about sourcing is part of the value here. The sharpest figures in this account, the rarity of real financial return, the average loss to AI mishaps, and the way cornered models behave, come from McKinsey, EY, and Anthropic rather than from the sponsored survey, and GAIG cites them to their origin so readers can check the work. The Databricks framing gets read with care, especially where it flatters the data platforms the sponsor happens to sell.
The instruction for anyone buying or running these systems is unglamorous and firm. Governance has to follow a system across its whole life, with real monitoring after launch rather than a sign-off before it, because the model in production is the one that drifts. Every agent that can act on the business needs a named owner, a traceable record, and a switch a human can throw, which is also the practical answer to the question of who is accountable when an autonomous system errs. The firms that build that control layer before they widen autonomy are the ones the survey keeps naming as the exception, and control before autonomy is the soundest advice the report has to offer.
