The survey that puts a number on agent sprawl
Dataiku, working with The Harris Poll, surveyed 685 CIOs across the US, UK, France, Germany, UAE, Japan, South Korea, and Singapore between July 9 and 29, 2026, for its Global AI Confessions Report: CIO Edition, published September 24, 2026. The headline numbers describe an organization that has lost the thread on its own AI footprint: 84 percent say employees build agents and applications faster than IT can govern them, 83 percent lack standardized lifecycle management for those agents, and 81 percent admit they have no complete oversight of agents built outside approved systems.
The gap between confidence and reality is the detail worth sitting with. 90 percent of CIOs say they are confident they fully track their agents, yet 72 percent cannot consistently confirm those agents actually deliver business outcomes. This is a measurement problem at its core, and measurement is squarely data infrastructure work: instrumentation, logging, and a pipeline that connects agent activity to an outcome someone can audit, the same discipline data teams have spent a decade building for dashboards and reports, now overdue for extension to agents.
The data team is holding the bag without holding the authority
When an agent fails, accountability splits four ways in Dataiku's data: shared teams take 23 percent of the blame, central IT 21 percent, the data or AI team 20 percent, and security, risk, and compliance 18 percent. No group holds a clear majority, which in practice means the data team absorbs real responsibility for a fifth of agent failures without necessarily having the authority to set the governance rules that would have prevented them in the first place.
This is the predictable outcome of letting agent creation outrun governance. Business teams build what they need, data and platform teams get pulled in after the fact to explain why an agent made a bad decision or ran up an unexpected bill, and the org chart never catches up to where the actual accountability has settled. If your data team is already fielding postmortem questions about agents it did not build and did not approve, that pattern is already live inside your organization, whether it has been named yet or not.
Cost visibility is the gap that will force the issue
Only 21 percent of CIOs report full, near real time visibility into AI costs attributed by business unit, team, or use case. Combine that with 67 percent estimating 51 or more agents in production and 47 percent having already decommissioned more than 20 agents this year, and the picture is an environment generating real infrastructure spend that finance cannot trace to a line of business, let alone a return on investment.
That combination tends to resolve itself the same way every time an unmetered internal platform scales past a threshold: finance notices the bill before anyone notices the value, and a blunt cost freeze lands on the data and platform teams to execute, regardless of which agents were actually worth running. Building cost attribution into the data pipeline now, while the agent count is still in the dozens per company rather than hundreds, is materially cheaper than retrofitting it under a mandate.
Shift the frame from monitoring to management
Dataiku CEO Florian Douetteau frames the needed shift precisely: 'Monitoring tells you an agent is running. Managing tells you whether it's earned the right to keep running.' That distinction maps directly onto data platform maturity. Monitoring is a dashboard that confirms an agent executed. Management is a lifecycle: provisioning tied to a business case, instrumentation tied to an outcome metric, a review cadence with a real decommissioning trigger, and a named owner who can actually pull the plug when the metric says the agent is not earning its keep, rather than letting it run indefinitely because nobody owns the decision to turn it off.
Most data platforms already have the technical building blocks for this: audit logs, usage metering, and access controls, inherited from years spent building governed pipelines for traditional analytics. The gap is organizational rather than technical. Those capabilities were built for pipelines and dashboards that business teams requested through a known intake process, with a ticket, an approver, and a budget line, while agents increasingly get spun up directly against a model API with no ticket filed anywhere and no budget line attached until finance asks where the spend went. Closing that gap means extending the existing intake discipline to agent workloads, not inventing a new governance process from scratch.
The architectural choice CIOs are already making
91 percent of CIOs in the survey agree the most effective strategy is letting business teams build within a governed environment rather than centralizing every application through IT. That is the right call, and it is also the harder one to execute, because a governed environment for agent building is a data platform decision before it is a policy decision. It means the catalog, the lineage tracking, the access controls, and the cost metering all have to extend to agent workloads with the same rigor they apply to a BI dashboard today, which for most organizations means a genuine re-architecture rather than a policy memo circulated to engineering leads.
75 percent of CIOs plan to use more or different models for continuity, and 74 percent are weighing open source or open weight models as a hedge against any single vendor's availability. Both of those decisions increase the surface area a data platform needs to govern consistently. The organizations that treat agent governance as an extension of their existing data platform, rather than a parallel system bolted on afterward, are the ones that will answer the accountability question before the next CIO survey asks it again.
