The Question Most CIOs Cannot Answer
Dataiku's pitch for Agent Management opens with a comparison its CEO, Florian Douetteau, clearly enjoys making: ask a bank how many servers it runs, and you get an answer to the decimal. Ask the same bank how many AI agents it has running across its business units, and the answer is a shrug. That gap is not a minor operational embarrassment. It is a governance hole wide enough that finance teams cannot price the risk they are carrying and security teams cannot say with confidence which systems touch customer data.
Dataiku backs the point with a citation to IBM research showing fewer than one in five organizations maintain a complete and current inventory of their AI systems. Teams built agents faster than anyone could count them, in Douetteau's phrasing, and the count problem compounds because agents get built inside whichever platform a given team already had a seat for. One group ships an agent inside Snowflake Cortex, another inside Databricks, a third inside Copilot Studio, and nobody owns the aggregate view.
A Product That Sits Above the Platforms It Doesn't Compete With
What makes Agent Management worth a CIO's attention rather than a shrug of its own is that Dataiku built it to be platform-agnostic by design. It connects to AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, and Snowflake Cortex, plus Dataiku's own environment, and it supports OpenTelemetry for anything custom-built outside those seven. That is a deliberate bet that no enterprise of any size has standardized on a single agent-building platform, and that the governance layer has to sit above the stack rather than being bolted onto one vendor's corner of it.
The product automatically maps each agent it finds, including its tools and underlying models, and lets a compliance or finance team ask portfolio-level questions in plain language: which agents are unmonitored, where is risk concentrated, and whether a given agent is actually worth what it costs to run. That last question, cost against value, is the one most agent governance tools skip because it requires connecting technical telemetry to a business outcome, which is a harder integration problem than simply logging API calls.
What Happens to the Agents That Actually Matter
Not every agent needs the same scrutiny, and Dataiku's design reflects that. For agents that touch customers, sensitive data, or live transactions, Agent Management keeps a recorded certification status, a list of named risks, and a schedule of retests, building an audit trail a compliance officer can hand to an examiner without having to reconstruct it after the fact. That is a meaningfully different posture from dashboards that show uptime and latency but go silent the moment an auditor asks whether a specific agent was tested against a specific risk in the last quarter.
The emphasis on named risk and scheduled retesting, rather than a one-time certification that ages quietly on a shelf, matches where regulators are heading faster than most enterprise AI tooling has caught up to. A static sign-off at launch is close to worthless eighteen months later when the underlying model, the data it touches, or the business process it automates has all quietly changed. Scheduling the recheck as a default behavior, instead of relying on someone remembering to ask for one, is the difference between a governance program that survives contact with an audit and one that looks good only on the day it launched.
The Pricing Signal Worth Reading Closely
Dataiku priced Agent Management per instance annually, with monitoring metered per agent. That structure is worth more attention than a typical SaaS pricing note, because it means the bill scales directly with how much agent sprawl an enterprise has allowed to accumulate. A company with a disciplined agent inventory pays a predictable base fee. A company that let every product team spin up its own Copilot Studio or Bedrock agent without a central registry will discover the cost of that sprawl the first time it turns the lights on.
That dynamic creates an unusual incentive structure for a governance tool: the vendor benefits financially from exactly the sprawl it is selling enterprises a cure for. The honest conclusion is that procurement teams should treat the quote as a starting estimate, not a final number, and ask for a cost projection before rollout rather than after. The metered-per-agent line item could move substantially once the full inventory surfaces for the first time and the real count of orphaned, duplicated, and forgotten agents comes into view.
Why This Lands on the Data Platform, Not Just the AI Team
Agent Management is a data governance story as much as it is an AI governance story, because every one of the seven platforms it connects to is where an enterprise's governed data already lives. An agent running inside Snowflake Cortex or Databricks inherits whatever access that platform granted it, and until now there has been no consistent way to see whether that access was appropriate across every agent touching the warehouse at once. Dataiku's product effectively extends data governance discipline, the kind CIOs already apply to who can query which table, into the much messier world of what an autonomous agent is allowed to do with that same access.
For a CIO who already owns the data governance program, that framing makes Agent Management a natural extension of existing responsibility rather than a brand-new mandate to staff from scratch. The decision in front of these buyers is not whether agent sprawl is a real problem, that much is settled by the IBM figure Dataiku is citing. The decision is whether to buy a cross-platform inventory now, while the agent count is still in the hundreds, or wait until it is in the thousands and the retrofit costs far more than the annual-instance fee ever would have.
