CIOs Own the Blame for Agents They Cannot Kill

A Harris Poll survey of 685 CIOs found that 81% lack a full view of AI agents staff build outside approved systems. Dataiku released the Global AI Confessions Report: CIO Edition on September 24, 2026, at Succeed, its New York conference. Florian Douetteau, the company’s co-founder and CEO, said almost nobody can fire an agent.

Those same CIOs say their CEOs have already tied job security to AI results, and most cannot prove the results exist.

Tracking Stops Short of a Kill Switch

The Harris Poll ran the study online from July 9 to 29, 2026, among CIOs in the United States, the United Kingdom, France, Germany, the UAE, Japan, South Korea, and Singapore. Dataiku, a Paris-founded firm now based in New York, paid for the work and published it as it showed a new product to more than 1,000 AI, IT, data, and business leaders.

On the agents they already count, CIOs sound confident. 90% say they have complete tracking of that fleet. 67% estimate 51 or more agents already run in production. The hole is everything that never entered the count, plus the test of whether the counted ones earn their keep.

WHAT THE 685 CIOS SAID ABOUT THE FLEET

• Known agents: 90% say they have complete tracking of every agent they already count.
• Shadow build: 81% lack a full view of agents created outside approved systems or formal channels.
• Cost view: Only 21% report full, near-real-time visibility into AI costs by business unit, team, or use case.
• Shared process: 83% lack a shared process for starting, running, and stopping agents across the company.

Sixty percent said they have no central control layer across apps, tools, and IT environments. Eighty-four percent agreed staff are creating agents and applications faster than IT can govern them. Tracking answers whether something is on. It does not answer whether it should stay on.

Monitoring tells you an agent is running. Managing tells you whether it’s earned the right to keep running, and right now, almost nobody can fire an agent.

Florian Douetteau, co-founder and CEO of Dataiku, Succeed conference, New York

72% cannot consistently confirm that the agents they track deliver the business outcomes they were built for. Douetteau put the industry’s pitch in blunter terms, saying it sold CIOs on autonomy and forgot to mention the bill: autonomous systems need to be managed, not just watched.

Boards Tied CIO Careers to Unmeasured Returns

The personal exposure is already in writing. Eighty-eight percent of CIOs said their reputation or career path will be shaped by success with AI. Eighty-six percent believe their own pay will be linked to proven AI outcomes. 87% said their CEO has told them that job security depends on those outcomes.

Seventy-six percent believe the CIO job itself will be at risk if the company shows no proven AI gains by the end of 2027. Ninety-seven percent reported at least some rise in board pressure to show a return on AI, and 77% called that rise significant or moderate. The nearer clock is money, not title: 72% said their AI budget is likely to be cut or frozen if performance targets are missed by the end of 2026.

HOW U.S. CIOS COMPARE WITH THE GLOBAL SAMPLE

In the United States the same questions run hotter, and 92% of U.S. CIOs said they would currently stake their job on delivering proven AI results. The people who will be judged on the return are the people who cannot see the full bill.

Who Owns an Agent When It Breaks?

Even when an agent is flagged as a problem, CIOs do not agree who carries it. Dataiku asked who is primarily responsible when agents go awry, and the room split.

WHO CIOS SAY SHOULD OWN A FAILED AGENT

• Shared teams: 23% put the failure on groups that share the work.
• Central IT: 21% pointed at the core IT function.
• Data or AI team: 20% named the specialists who built the stack.
• Security, risk, and compliance: 18% said those control functions should carry it.

Those four answers cover 82% of the sample. Dataiku posed the ownership riddle a day before the conference, asking whether the team that built an agent is the same team accountable once it is live, and answering that they usually are not. A CIO who cannot name an owner also cannot name who is allowed to shut the thing down.

The Cull Started Before Governance Caught Up

Forty-seven percent of the CIOs had already shut down more than 20 agents in 2026 by the time they took the survey. That churn, Dataiku said, is happening largely in the dark. The shutdowns are not a future forecast. They are this year’s operating reality for almost half the sample, while 83% still lack a shared lifecycle process that would make a shutdown a normal step rather than an ad hoc raid.

THE PATH FROM HYPE TO SHUTDOWNS

1. June 25, 2025: Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, or weak risk controls.
2. July 9 to 29, 2026: The Harris Poll fields Dataiku’s CIO survey in eight countries.
3. September 23, 2026: Swami Sivasubramanian, AWS vice president of agentic AI, tells HumanX in Amsterdam that almost 90% of Amazon’s early agent prototypes never reached production.
4. September 24, 2026: Dataiku releases the CIO report and launches Agent Management at Succeed in New York.
5. October 2026: Agent Management is due to go on general sale, priced per instance each year, with monitoring charged per agent.

Anushree Verma, a senior director analyst at Gartner, said most agentic projects right now are early experiments or proofs of concept driven by hype and often misapplied. In a January 2025 Gartner poll of 3,412 webinar attendees, 19% said their organization had made significant investments in agentic AI, 42% had made conservative investments, 8% had made none, and 31% were waiting or unsure. Verma also said many so-called agentic pitches lack a real return, because current models do not have the maturity to hit complex business goals on their own.

Sivasubramanian described Amazon’s own miss rate from two years earlier, when the company was pouring money into agents and most prototypes never shipped. Citing analysts, he said 17% of organizations have successfully deployed AI agents and only 7% can measure the return. The CIO survey sits on the same fault line: the agents that made it into production are numerous, the proof is thin, and the ones that fail are already being retired without a common rulebook.

Regret Hits 82% of U.S. CIOs on Vendor Picks

U.S. CIOs also reported the sharpest hangover from the last round of buying. Eighty-two percent regret at least one major AI vendor or platform choice from the past 18 months. That regret is showing up as a hedge, not a pause. Seventy-five percent of CIOs worldwide plan to use more or different models to keep AI running if one source wobbles, and 74% are considering open-source or open-weight models as a hedge if a model becomes unavailable.

The people who will be blamed for a bad platform bet are trying not to be trapped by the next one. They are not recentralizing the build. They are spreading it across models and vendors while the measurement layer stays immature. An inventory that cannot fire an agent still leaves the last 18 months of contracts in place, and the next budget review will ask what any of it returned.

Dataiku Ships an Inventory Layer in October

On the same day as the survey, Dataiku launched Agent Management, a standalone product that finds agents wherever they were built, scores business and technical performance, and flags the ones that carry the most risk. It is due for general availability in October 2026. Douetteau said a bank can tell you how many servers it runs to the decimal, and that the same bank shrugs when asked how many AI agents it runs.

The company pointed to IBM’s AI in Motion research, which found that fewer than one in five organizations keep a complete, current inventory of AI systems. Most agent platforms, Dataiku said, can see only the agents built on them. Agent Management is meant to sit above that stack and scan AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio, Azure Foundry, Salesforce Agentforce, Snowflake Cortex, and Dataiku itself into one inventory of agents across platforms.

WHAT THE SCAN IS BUILT TO SEE

• Owner and purpose: Each entry is supposed to carry who owns the agent, what it connects to, when it last ran, and what it was built to do.
• Live behavior: The product tracks usage, cost, quality, and drift, and raises a flag when an agent leaves its normal pattern.
• High-risk record: For agents that handle customers, sensitive data, or live transactions, it keeps certification status, named risks, and tests that rerun on a schedule.
• No hands on the wheel: Dataiku says the product does not operate the agents or step into their decisions, so teams keep running what they already built.

That last point is the limit of the new layer. It can list an agent, score it, and keep an audit file. The shutoff still belongs to a person, and 83% of CIOs said their companies do not have a shared process for that act. An inventory without a recorded decision to keep, pause, or kill an agent is still a dashboard, and dashboards do not survive a budget freeze on their own.

For forty years, enterprise software didn't need a manager.

Your CRM doesn't decide anything. People use it, and people get managed. The software just has to stay up.

Agents broke that deal. They're the first software that does the job instead of helping someone do it. So they… pic.twitter.com/xlyK7G3LVt

— Florian Douetteau (@fdouetteau) September 24, 2026

Douetteau’s own post on September 24 argued that ordinary enterprise software never needed a manager, because people used it and people got managed, while agents are the first software that does the job instead of helping someone do it. That is the bet Dataiku is selling, and it is also the bind the survey describes. Boards want a return by the end of 2026. CIOs can see the agents they already registered. The ones they cannot fire, and the ones they never counted, are still running when the budget meeting starts.

Frequently Asked Questions

How Does Dataiku’s CIO Study Compare With Its CEO Survey?

The CIO study covered 685 people in eight countries in July 2026. Dataiku’s earlier Global AI Confessions Report: CEO Edition, released May 4, 2026, surveyed 900 CEOs in the same markets and found that 80% said their own jobs would be at risk by the end of 2026 if AI failed to deliver, a shorter clock than the end-2027 deadline CIOs now attach to their own roles.

Do Most CIOs Want to Centralize Agent Building?

No. Ninety-one percent agreed that the most effective AI strategy is not to centralize every application and agent, but to let business teams build inside a governed environment, which is why a control product that leaves existing agents running is the path Dataiku is selling rather than a lock-down of staff tools.

What Else Did Gartner Forecast Besides Project Cancellations?

Gartner also estimated that only about 130 of the thousands of vendors marketing agentic tools are real, a practice it called agent washing, and it predicted that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, while 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.

Which Extra Connectors and Dates Sit Around Agent Management?

Beyond the major clouds named at launch, Dataiku’s product page lists n8n as a source and OpenTelemetry support for custom environments, and the company scheduled a live demo of Agent Management for October 1, 2026, ahead of general sale later that month, with access priced per instance each year and monitoring metered per agent.