Ask a bank how many servers it runs, and you get an answer to the decimal. Ask how many AI agents it’s running, and you get a shrug or a guess.
That observation from Dataiku CEO Florian Douetteau captures the current state of enterprise AI. After a frantic wave of deployment, where organizations rushed to stand up autonomous agents, the industry is now hitting a wall of visibility. The infrastructure that allows companies to build agents is robust, but the infrastructure to control them is only just beginning to take shape.
Dataiku recently announced the general availability of its Agent Management product, set for October 2026. While the company first introduced these capabilities in March 2026 as part of its broader Platform for AI Success launch, the new GA announcement marks an incremental shift: packaging those existing tools into a standalone, platform-agnostic product. It is not a new invention, but rather a formalization of the need for a control layer that sits above the fragmented vendor stacks.
The core of the problem, according to Douetteau, is the difference between passive observation and active control.
Monitoring tells you an agent is running. Managing tells you whether it has earned the right to keep running, and right now, almost nobody can fire an agent.
This lack of control is backed by data, though the figures deserve a caveat. The most striking numbers come from a Harris Poll survey commissioned by Dataiku, fielded among 685 CIOs across eight countries in July 2026. In that study, 81% of respondents reported a lack of complete oversight for agents created outside approved systems. Seventy-two percent admitted they cannot consistently measure the business outcomes of their agents. And 47% said they have already decommissioned more than 20 agents this year — churn happening largely in the dark. These are self-reported numbers from a vendor-sponsored study, not independent benchmarks, but the directional signal aligns with broader research. IBM’s AI in Motion report from April 2026 found that fewer than one in five organizations maintain a complete, current inventory of their AI systems.
The Dataiku announcement is part of a larger pattern. Five different governance products have hit the market since September 12 alone. The industry is moving past the build-at-all-costs phase and into a period of consolidation and oversight. A specialized stack for agent management is emerging, distinct from the platforms used to build agents in the first place.
The market is placing a premium on orchestration and control. Yesterday, NiCE acquired Cognigy for $955 million, underscoring the massive value assigned to the routing layer of agent orchestration — a shift we covered in our analysis of NiCE’s routing-layer bet. Collibra recently introduced Guardian Agents, which provides runtime governance by reading machine-readable Agent Contracts to supervise behavior and block unauthorized actions in real time. The routing layer got a $955 million price tag. The governance layer is forming. And now the monitoring layer is shipping.
Dataiku’s positioning depends on being platform-agnostic. Unlike tools tied to a single vendor stack — the kind built into AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio, Salesforce Agentforce, or Snowflake Cortex — Dataiku’s product is designed to sit above all of them. It connects via native APIs and log streams, meaning organizations do not need to rewrite or redeploy their existing agents to gain visibility. It also supports custom environments via OpenTelemetry.
That agnostic claim is the real differentiator. Vendor-specific monitoring can only see agents built on that vendor’s platform. In a world where 67% of CIOs estimate they have 51 or more agents actively running in production — and those agents span multiple platforms — a single-vendor view leaves most of the estate invisible. The question is whether platform-agnostic monitoring can actually deliver on the promise, or whether the depth of integration will vary so much across platforms that the agnostic label becomes more marketing than substance.
There is also a gap between seeing and measuring. Monitoring an agent’s technical performance — is it running, is it consuming resources, is it hitting errors — is a necessary first step. But it is not the same as measuring whether the agent is actually delivering business value. IDC and Lenovo research found that 88% of custom agent builds fail. The control plane is catching up to the deployment wave, but the bottleneck has shifted from visibility to accountability. We can now see what our agents are doing. The harder question is whether they are doing it well enough to keep running.