Dataiku Agent Management: Cross-Platform Governance for Copilot Studio and Azure Foundry

What Dataiku announced​

Seoul Economic Daily covered the product for Korean readers. Its report is dated September 29 Korea time and says the product ships next month. The English-language press release came out several days earlier. Dataiku calls Agent Management a standalone product that finds every AI agent an enterprise is running, regardless of which platform built it, measures the business and technical performance, and flags agents that pose the greatest risk. It was unveiled at Dataiku Succeed, the company's annual flagship conference, and the company says it will be generally available in October 2026.

At the time of writing, that release is still a promise. No exact October date has been given, and there is no independent testing or customer deployment data yet. Everything below describes what the vendor says the product does. None of it is a benchmark.

The short version: this is a cross-platform inventory and oversight tool for AI agents, due in October, from a company best known for its data science and AI platform.

The platforms, and where Microsoft fits​

The Seoul Economic Daily report names only AWS Bedrock, "Microsoft Copilot" and Salesforce Agentforce. Dataiku's own list is longer and more specific. According to the company, Agent Management connects to the platforms that enterprise teams already use to run agents, including AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, Snowflake Cortex, Dataiku itself, with OpenTelemetry support for custom environments, and scans them into a single inventory. Dataiku's FAQ page also lists n8n.

The difference matters for Windows and Microsoft 365 administrators. "Microsoft Copilot" could mean anything from the Copilot app on your desktop to Microsoft 365 Copilot. What Dataiku actually names are Copilot Studio, the low-code tool where business users build agents, and Azure Foundry, where developers build and deploy them. If you're responsible for a Power Platform environment where makers have been creating agents, this product is aimed at your problem.

It is also a third-party product. Microsoft hasn't launched or endorsed anything here. Dataiku says its product is deliberately agnostic, and contrasts that with monitoring tools that sit inside a single vendor's stack. That's a marketing jab at platform vendors, Microsoft included. Microsoft has been building its own agent governance features into its admin and identity tools (that point comes from general industry knowledge, not from Dataiku's materials), so buyers will be comparing native controls against a separate layer that sits on top of several platforms.

What the product says it does​

Dataiku describes three main jobs.

1. Discovery and inventory. The company says the product automatically identifies each agent's structure, including the tools and models it relies on, giving supervisors full transparency into how an agent actually works, not just that it exists. Its product page adds that each inventoried agent gets an owner and a stated purpose.
2. Performance measurement. Forkast's coverage says it measures agents against business key performance indicators — usage, cost, quality, and value — rather than just uptime. Dataiku's point is that infrastructure monitoring can tell you a server is healthy but not whether an agent has drifted from what it's supposed to do.
3. Risk records for high-stakes agents. For agents handling customers, sensitive data, or live transactions, it keeps a standing record of certification status, named risks, and tests that rerun on a schedule, so the evidence trail already exists when a manager, auditor, or regulator asks for it.

Just as important is what it doesn't do. Dataiku's FAQ says Agent Management is not an AI gateway. It sits above agents and keeps track of them. It doesn't sit in the traffic path between agents and their tools, and the company's materials don't say it can block, change or shut down an agent. Think of it as a register and a record-keeper. If you need runtime enforcement, that comes from somewhere else.

The short version: you get an inventory, KPI tracking and an audit trail. You don't get a control layer that intercepts agent traffic.

How integration works, and the caveat​

Dataiku says the tool connects to third-party frameworks through native APIs and log streams. You pick a framework, authenticate, and the system scans and lists the agents it finds. The company also says integration depth depends on what each platform exposes, and that deeper integrations are still being built.

That caveat matters more than anything else in the announcement. A cross-platform inventory can only see what each platform's APIs and logs report. If Copilot Studio or Foundry doesn't expose a piece of metadata, Dataiku can't show it. Anyone evaluating the product should test that claim in their own environment before relying on it.

A reasonable evaluation checklist (this is our analysis, not Dataiku guidance):

• Test coverage in your own tenant. Connect a non-production environment and compare what Agent Management finds with what you know exists in Copilot Studio and Foundry.
• Review the permissions. Find out which accounts or app registrations the connector needs, what scopes it asks for, and whether that fits your least-privilege policy in Entra.
• Check metadata depth for each platform. Ownership, tool connections and last-run data may be fuller on some platforms than on others.
• Plan for custom agents. Agents built in-house need OpenTelemetry instrumentation to show up. Byteiota put it bluntly: if your custom agents are not already emitting OpenTelemetry traces, start now.
• Define risk tiers first. The certification and testing records are only useful if your organization has already decided what makes an agent "high risk."

Pricing: the model, but not the numbers​

Dataiku says product access is priced per instance annually, with monitoring metered per agent. It hasn't published list prices, minimums or a definition of what counts as a billable "agent." That last gap matters. If makers keep spinning up agents in Copilot Studio, per-agent metering could get expensive quickly. Ask for that definition in writing before signing anything.

The numbers behind the pitch​

The Seoul Economic Daily report cites IBM's "AI in Motion" research as finding that fewer than 20% of organizations have a complete, current inventory of their AI systems. Forkast gives a more specific figure: only 18% of organizations maintain a complete, current agent inventory, and less than half have visibility into what agents exist across their environments, according to IBM Institute for Business Value research. We couldn't check the IBM report directly, so treat the figure as attributed to IBM rather than confirmed. The same Forkast piece mentions an OutSystems 2026 survey of 1,900 IT leaders that found widespread agent use but little centralized management.

Dataiku also promoted its own survey at the conference. Run with The Harris Poll, it covered 685 CIOs: 84% agreed employees were building AI agents and apps faster than IT could govern them, and 75% said they couldn't consistently measure whether their agents delivered business results. That's vendor-sponsored research published alongside a product launch, so it tells you about the problem Dataiku sees. It doesn't tell you whether the product works.

The CEO's line about banks and servers appears in slightly different wording in different outlets because the Korean report was translated. In Dataiku's English press release, Florian Douetteau says: "Ask a bank how many servers it runs, and you get an answer to the decimal." The next line in the release says asking about AI agents gets you a shrug or a guess.

Analysis: does IT need another console?​

To be fair, the problem is real. Agent sprawl looks a lot like shadow IT, just faster. Low-code tools like Copilot Studio have made it easy for someone in finance to build an agent that reads SharePoint, calls a connector and answers customer questions, all in one afternoon. Each platform's native admin tools see their own agents well and everyone else's agents not at all. For a company running Microsoft, Salesforce and AWS side by side, one register covering all of them is a sensible idea.

There are counterarguments.

• Yet another pane of glass. Forkast summed up the market question: Whether enterprises want another governance layer is the open question. Every new console is one more thing to license, secure and keep in sync.
• Native tools are catching up. Platform vendors, Microsoft included, have every reason to build agent inventory and governance into their own admin centers. A single-vendor shop may not need a separate layer.
• Visibility is not control. An agent that shows up in the inventory still needs to be governed through identity, data loss prevention and access policy. Dataiku says openly that this product doesn't sit in the traffic path.
• Vendor claims, no track record yet. Until customers run it at scale after October, "finds every AI agent" is a goal, not a proven result.

One more note: Forkast reported that Dataiku first previewed Agent Management in March as part of a wider platform launch. The late-September announcement is the general availability commitment. So the idea has been public for months, and October is when buyers can actually get it.

The bottom line​

Dataiku's Agent Management, due in October 2026, promises one inventory of AI agents across Copilot Studio, Azure Foundry, AWS Bedrock, Salesforce Agentforce and other platforms. It adds KPI tracking and audit records for high-risk agents, with per-instance pricing and per-agent metering. If you manage Microsoft environments where agents are multiplying, it's worth a look, but test coverage and permissions in your own tenant before you believe the word "every." Knowing how many agents you have is only the first step. You still have to govern them.