Dataiku's answer to agent sprawl: one inventory, one risk score, every AI agent

Dataiku wants to answer the question every CIO is now afraid to ask: how many AI agents are actually running in this company? At its Succeed conference in New York on September 24, the data-science platform company unveiled Agent Management, a standalone product that hunts down every AI agent inside an enterprise — regardless of which platform built it — inventories them, measures their business and technical performance, and flags the ones carrying the most risk. General availability is scheduled for October.

The pitch is simple and deliberately uncomfortable. Enterprises have been building agents far faster than they can count them, and the existing observability tools only see the agents built on their own platforms. Dataiku's answer is a layer that sits above all of them — AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, Snowflake Cortex — with OpenTelemetry support for custom environments, all feeding one inventory.

The sprawl problem is now measurable#

The numbers Dataiku is pointing at are hard to dismiss. According to IBM's "AI in Motion" research, fewer than one in five organizations keeps a complete, current inventory of its AI systems — because most agent platforms can only see the agents built on them, leaving the bulk of the estate without clear ownership, purpose, risk, or business outcome.

Dataiku's own survey of British technology leaders, released alongside the launch, fills in the damage report: 53% said an AI agent had violated policy in a way that directly affected customers, and 87% agreed employees create agents faster than central teams can govern them. Over two-thirds of UK businesses said it took them more than 24 hours to contain a problematic agent, and only 49% could produce an audit trail for AI decisions.

"As employees rapidly build and deploy agents, an accountability gap is growing between how fast agents are multiplying and how little companies can say about the value they are delivering," said Kurt Muehmel, Dataiku's head of AI strategy.

What Agent Management actually does#

Once connected to the agent platforms in use, the system scans them into a single inventory. Each record carries the agent's owner, purpose, connected systems and last run — and the product automatically identifies each agent's internal structure, including the models and tools it relies on, so supervisors see how an agent actually works rather than merely that it exists.

From there it becomes operational. Agent Management monitors each agent's health, usage, cost, quality and behavior over time, and fires alerts when an agent drifts away from its normal pattern. For the highest-risk agents — the ones handling customers, sensitive data or live transactions — it keeps a standing record of certification status, named risks and tests that re-run on a schedule. The risks it is designed to catch include excessive privileges, shared credentials, and the absence of a human override.

Teams query all of this in plain language and get answers across the whole portfolio: which agents are unmonitored, where risk is concentrated, which ones are earning their cost.

The platform-agnostic bet#

The most important design choice is that Agent Management is deliberately nobody's platform. It is built to sit above the stack — a direct inversion of the usual vendor logic, where each platform offers excellent visibility into its own agents and none at all into everyone else's.

That framing gets at a structural truth of this year's enterprise AI boom: the agent estate is inherently multi-vendor. A bank might have a credit-approval agent on Bedrock, a support copilot in Copilot Studio, a sales workflow in Agentforce and a homegrown research assistant wired into Databricks. No single platform's dashboard can see the whole picture, which means nobody can answer the three questions a board is now entitled to ask: what's out there, what's it worth, and where's the risk?

Co-founder and CEO Florian Douetteau put it in the line that will get quoted back at him: "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. Nobody set out to build it this way. Teams built agents faster than anyone could count them. Agent Management tells you what's actually out there, and what it's actually worth."

What to watch#

• Whether the October general availability lands on schedule — and how deep the platform connectors actually go in production deployments. Connector depth is where inventory products live or die.
• Pricing, which Dataiku did not disclose. Enterprise governance tooling is bought on audit and compliance budgets; the price will reveal who the real buyer is.
• Whether auditors and regulators accept Agent Management's standing certification records as evidence. The scheduled re-tests for high-risk agents are built for exactly that job — but it is a claim that only holds once a regulator has actually leaned on it.
• How the platforms respond. An independent inventory layer only works if the platforms keep their APIs open to it. Expect the big vendors to partner, copy, or quietly make life harder — and the answer will say a lot about where power sits in the agent stack.