Dataiku Agent Management: count your agents first

You cannot govern an agent you cannot name. On September 24, Dataiku launched Agent Management, a standalone product whose entire pitch is finding the AI agents a company is already running — across whatever platform built them — and putting them in one list with an owner next to each. It is not a model. It is a headcount. That tells you something about where enterprise AI actually is right now.

What actually happened

Agent Management connects to the platforms teams build agents on — AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, Snowflake Cortex, n8n and Dataiku itself — plus OpenTelemetry for anything custom. It scans them into a single inventory, maps the tools and models each agent depends on, and tracks usage, cost and quality per agent. It also handles certification: assessing an agent against named risks with documented mitigations, on a scheduled test cadence, with an audit trail.

The company announced it at its Dataiku Succeed conference. SiliconANGLE reports general availability in October 2026, priced as an annual per-instance fee metered by agent count.

The stats Dataiku cites to justify the product are the interesting part. It points to IBM research finding fewer than one in five organizations keep a complete, current inventory of their AI systems. A Harris Poll survey of 685 CIOs in July 2026, also cited in the launch coverage, found 81% lacked complete oversight of agents built outside approved systems and 84% said employees were building agents faster than IT could govern them. Treat those as vendor-selected numbers making a vendor's case — but the direction is not controversial.

Why an agent inventory matters for your business

You are not buying an enterprise governance suite at your size. You still have the problem it exists to solve, and yours is cheaper to fix.

Your inventory is a spreadsheet, and it should exist today. One row per agent: what it is called, what triggers it, which model and API keys it uses, what it can write to, who owns it, and the date someone last confirmed it still works. That is an afternoon. The enterprises paying per-instance fees are buying automated discovery because they lost track. You have not lost track yet.

Shadow agents come from your best people. The ops lead who wired a Zapier-plus-model workflow to chase invoices did the right thing. The problem is that it runs on their personal API key, nobody else knows the prompt, and it will break silently the week they take vacation. Inventory is not a crackdown; it is making that work survivable.

Cost per agent is the metric nobody tracks. Dataiku measuring usage, cost and quality per agent exists because most organizations can see one aggregated model bill and cannot attribute it. Tag your API calls per workflow now, while you have five of them, and you will never have to reverse-engineer the invoice later. Run the numbers on the two you think are earning their keep.

"Last confirmed working" beats any dashboard. An agent that quietly stopped firing in July is worse than no agent, because someone stopped checking the thing it was supposed to handle. A date column and a monthly ten-minute review catches that.

Sources: Dataiku, SiliconANGLE.