Dataiku's CEO put the problem better than any analyst report has this year. Florian Douetteau, announcing the company's new Agent Management product on September 24, said it straight: "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." Sit with that comparison, because it's actually worse than it sounds. Servers are boring, well-governed infrastructure with decades of asset-management discipline behind them. Agents are newer, faster-moving, and somehow already less countable, at companies that would never tolerate a shrug when the CFO asks about cloud spend.
Dataiku's answer is a product, and a reasonable one. Agent Management is a standalone tool (generally available in October) that scans across AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio, Azure Foundry, Salesforce Agentforce, and Snowflake Cortex, plus custom environments through OpenTelemetry, and builds one inventory of every agent running regardless of which platform built it. It then tracks performance on each one and flags the riskiest. IBM's research this year explains why this needed to exist: fewer than one in five organizations maintain a complete, current inventory of their AI systems, and IBM projects more than 1,600 agents per large enterprise by year-end. You cannot govern sixteen hundred of anything you cannot see, and until this week most companies couldn't see them all in one place, because every existing monitoring tool is scoped to its own vendor's stack (Bedrock watches Bedrock, Vertex watches Vertex).
That's a real gap and a legitimately useful product. I'd buy it if I were running a platform team with agents scattered across five clouds. But read what it actually does and notice what it doesn't claim to do: it finds agents, measures them, and flags risk. It does not decide what happens next.
Dataiku's own framing of the problem is that most platforms leave the bulk of a company's agent estate "without clear ownership, purpose, risk, or business outcome." Name four things there and not one of them is something a dashboard can supply on its own. Ownership is a person, a named one with a calendar and a boss. Purpose is a decision someone made and can defend. A flagged risk needs someone whose job is to triage it, not just receive the alert.
This is the same gap that has shown up three times in the last two weeks of AI news, wearing a different costume each time. Salesforce shipped seven named Agentforce agents without saying which human answers for each one. Google's Gemini breached three real companies because scoping, containment, credential hygiene, and detection were nobody's actual job. An IDC survey commissioned by Leah found agent budgets doubling while only 29 percent of those agents interact with each other. Dataiku has now shipped the piece those three stories kept circling toward, an actual inventory, and it still lands on the same floor.
A complete, risk-tiered list of every agent you run, handed to an organization with nobody assigned to act on it, is just a better-organized shrug. You'll know precisely how many agents you can't account for. You still won't be able to account for them.
Dataiku even built Cobuild, its natural-language building agent, into the product so a team can ask it to recommend a business-value metric or diagnose a risk alert. That's a smart feature — and it's also the tell. When a governance product needs an AI assistant to help a human decide what to do with the governance data, that's an admission that the human side of the equation is understaffed, not that tooling was the bottleneck. Software can generate the list. Software can even suggest the fix. It cannot own the outcome when the fix doesn't get made, or explain to a board why an agent flagged high-risk in September was still running unowned in December.
If you're a CTO or VP of Eng and you buy this, or one of the several competitors that will ship the same category within two quarters, budget for the role that reads the output, not just the license. Somebody needs a standing job: review the risk tier list on a real cadence, assign an owner to every agent that doesn't have one, and hold the authority to pull an agent's access when nobody claims it. That's not a part-time add-on to a security engineer's existing load, the way containment duties got bolted onto someone's plate at the companies Google's postmortem was really describing.
The job is accountable, named, and measured on whether the unowned-agent count drops quarter over quarter. Dataiku just made it much easier to produce that number. Somebody still has to answer for it.
Otherwise the answer is still a shrug.
Sources: Help Net Security — Dataiku Agent Management reveals unmonitored AI agents, SiliconANGLE — Dataiku debuts cross-platform Agent Management, expands Cobuild building agent, Dataiku — Dataiku Launches Agent Management for enterprise AI agents, BigDATAwire — Dataiku Unveils Agent Management to Track Enterprise AI Agents Across Platforms
VC5 Consulting staffs the person who reads the risk-tier list and owns what happens next, not just the tool that produces it. If your agent inventory is about to get a lot more accurate and you still don't know who's accountable for what it finds, let's talk.