The Inventory Nobody Kept: Dataiku's Agent Management Sets Out to Count Every AI Agent a Company Runs

Every large company can tell you, to the decimal, how many servers it runs. Almost none can say the same about their AI agents. That asymmetry — between how fast agents are being built and how poorly they are being observed — is the gap Dataiku aimed at this week when it unveiled Agent Management at its Succeed conference in New York.

Announced on September 24 and set for general availability in October 2026, Agent Management is a standalone product, not a feature bolted onto Dataiku’s platform. Its job is blunt: find every AI agent an enterprise is running, regardless of which platform built it; measure each one’s business and technical performance; and flag the agents that pose the greatest risk.

The problem: agents multiplied faster than anyone could count them

The backstory is familiar to anyone running enterprise IT in 2026. Marketing stood up an agent to draft campaign briefs. Finance deployed another to reconcile invoices. Operations has several more in production. And somewhere, someone connected an agent to company data last week without filing a ticket. The ability to build AI has spread across the enterprise faster than the ability to manage it.

The numbers back up the anecdote. Fewer than one in five organizations maintain a complete, current inventory of their AI systems, according to IBM’s “AI in Motion” research cited by Dataiku. And the company’s own survey work is bleaker: the “Global AI Confessions Report: CIO Edition, 2026,” published at Succeed and built on a Harris Poll survey of 685 CIOs at companies above $500 million in revenue across eight countries, found that 81% of respondents admit they have lost oversight of agents built outside approved systems or formal channels. Nine in ten said they were confident they had complete tracking — right up until they were asked specifically about the agents nobody approved.

“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,” said Florian Douetteau, Dataiku’s co-founder and CEO. “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.”

How it works: connectors above the stack

The technical approach is what makes the launch notable. Most agent observability tools live inside a single vendor’s stack, which means they can only see — and tend to favor — that vendor’s own agents. Agent Management is deliberately agnostic. It sits above the stack and connects to the platforms enterprise teams already use to run agents: AWS Bedrock, Databricks Agents, Google Vertex and Gemini Enterprise, Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, Snowflake Cortex, and Dataiku itself. Custom and coded environments can report in through the OpenTelemetry standard.

Once connected, the product scans those platforms into a single inventory. Each entry automatically identifies the agent’s structure — the tools and models it depends on — so supervisors get transparency into how an agent actually works, not just that it exists. Every record carries the basics: who owns it, what it connects to, when it last ran, and what it was built to do.

From there, Agent Management monitors each agent’s health, usage, cost, quality, and behavior over time. An agent with three users and an agent with 3,000 users should not look the same on a dashboard, and a system that raises its hand when an agent quietly drifts from its original purpose is the difference between a list and a portfolio. Notably, the product does not operate agents or step into their decisions — teams keep building and deploying wherever the work happens, and each new agent joins the record the moment it goes live.

For the highest-risk agents — the ones handling customers, sensitive data, or live transactions — Agent Management keeps a standing record of certification status, named risks, and tests that rerun on a schedule. The intent is that when a manager, auditor, or regulator asks for evidence, the trail already exists. Teams can also ask portfolio-wide questions in plain language: which agents are unmonitored, where is risk concentrated, which ones earn their cost.

Cobuild grows up alongside it

Dataiku used the same New York stage to expand Cobuild, its natural-language agent for building AI projects that went generally available in June. A new feature, Cobuild Insights, lets any employee ask a question of governed company data and get what Dataiku describes as analyst-quality answers, then carry that line of questioning into deeper work inside Dataiku.

A second addition, Dataiku Headless, brings Cobuild to developers inside the coding agents they already use — Anthropic’s Claude Code, OpenAI’s Codex, and Cursor — so a developer can build or change pipelines, models, and agents from within those tools. Migration is another use case: an old Excel workbook or Alteryx workflow can be converted into a governed project, and the tool can be pointed at a live environment to investigate problems. Cobuild is also embedded in Agent Management itself, where it can recommend business-value metrics for an agent or diagnose and fix a risk alert.

Clément Stenac, Dataiku’s co-founder and CTO, framed the stakes: enterprises tend to run technical and business teams on separate platforms without shared data or shared guardrails. A few pilots can live with that. At scale, he argues, it becomes “a liability companies have to prepare for.”

The business model and the bet

Dataiku disclosed timing and pricing structure, not dollar figures: Agent Management will be generally available in October 2026, with product access priced per instance annually and monitoring metered per agent. That metering model is itself a signal — the company is pricing on the assumption that agent counts will keep climbing, and that the cost of watching each one should scale linearly with the fleet.

The strategic bet is that cross-platform beats native. Every hyperscaler would prefer that enterprises build and observe agents exclusively inside its own stack; Dataiku is wagering that the reality of 2026 — multi-vendor, multi-platform, part-approved and part-shadow — makes a neutral layer above the stacks more valuable than any single vendor’s dashboard. With 84% of surveyed CIOs saying employees are building agents faster than IT can govern them, and regulators on both sides of the Atlantic sharpening expectations around AI accountability, the demand for an evidence-ready inventory is unlikely to soften.

The open question is whether agent governance becomes a product category of its own or gets absorbed into the platforms it currently watches. For now, Dataiku has moved first with the most complete cross-platform answer — and for enterprises that literally cannot count their agents, an October release cannot come soon enough.