Microsoft's New Copilot Is a Bet That 'One App for Work' Beats Every Standalone AI Product. The Pricing Model Tells You Everything.

On September 24, 2026, Dataiku published the results of a Harris Poll survey of 685 CIOs across the United States, United Kingdom, France, Germany, UAE, Japan, South Korea, and Singapore. The headline finding: 81% of global CIOs say they lack full oversight of AI agents built outside formal IT channels. Not 81% who "could improve" their visibility — 81% who have already lost the thread.

The survey, titled the "Global AI Confessions Report: CIO Edition, 2026," arrived alongside Dataiku's announcement of Agent Management, a standalone product designed to inventory AI agents across enterprise platforms regardless of which vendor built them. The two announcements together tell the same story: AI agents are being deployed faster than governance infrastructure can track them, the business and compliance risks of that gap are materializing, and the enterprise software market is producing a new product category in response — AI agent inventory and governance.

For enterprise technology teams, the moment represents a familiar pattern playing out at an unfamiliar speed: a new class of software gets deployed bottom-up across the organization before centralized governance infrastructure exists, the risks become visible through incidents, and the market produces tooling only after the sprawl is already significant. Shadow IT took fifteen years to produce a governance product category. AI agent sprawl took eighteen months.

The Numbers Behind the Governance Crisis

The Dataiku survey's topline statistics are stark, but the operational details are more specific:

• 84% of CIOs agree that employees are creating AI agents and applications faster than IT can govern them
• 83% lack standardized agent lifecycle management across the organization
• 72% cannot consistently confirm business outcomes from their deployed agents
• 60% lack a central AI governance layer entirely
• 47% have already decommissioned more than 20 agents this year — meaning the lifecycle management problem is active, not theoretical

The agent decommissioning statistic is particularly significant. An organization that has decommissioned 20+ agents in a single year has already discovered, after deployment, that those agents should not have been running — whether because they violated policy, failed to deliver value, created unexpected costs, or introduced risk that became visible only in production. The retrospective clean-up cycle is already underway in most large enterprises; governance infrastructure is lagging behind the clean-up effort, not ahead of it.

The most consequential number: 53% of CIOs in the UK sample said an AI agent had violated policy and directly affected customers. This is not a theoretical risk. It is an event that more than half of UK CIOs have already experienced — an AI agent interacting with a customer in a way that violated company policy, potentially creating legal, regulatory, or reputational exposure.

Only 21% of CIOs report full, near-real-time visibility into AI costs with attribution by business unit, team, or use case — which means 79% are running AI agent programs without the cost visibility required to assess ROI, identify waste, or defend budget allocations to finance.

Why AI Agent Sprawl Is Different From Shadow IT

Enterprise IT organizations have managed shadow IT — technology deployed outside formal IT channels — for decades. The conventional response involves discovery tools, usage policies, vendor rationalization, and bring-your-own-device programs that bring unauthorized technology under governance.

AI agent sprawl presents a structurally different problem in three ways:

Agents act, not just process. Traditional shadow IT — an unauthorized SaaS subscription, a personal device on the corporate network — processes data or enables communication. AI agents take actions: they make decisions, send communications, update records, execute transactions, and interact with external systems. The governance failure mode is not just data exposure; it is unauthorized action taken in the organization's name.

Agents are created faster and more frequently. The democratization of agent creation — Salesforce Agentforce, Microsoft Copilot Studio, AWS Bedrock, and dozens of no-code agent builders — means that a business user with no programming background can deploy a functional AI agent in hours. The creation rate is orders of magnitude higher than the rate at which business users were deploying unauthorized SaaS subscriptions in 2015. The governance tooling gap that took years to develop in the shadow IT era is being recreated in months for agents.

Agents have no natural audit trail. Traditional software systems generate logs. Those logs exist, are retrievable, and can be analyzed. AI agents built outside formal IT channels may not emit structured logs at all, or they may emit logs that are not connected to the organization's SIEM, monitoring, or audit infrastructure. An agent that has been running for three months may have no audit trail that the CIO can review when a compliance question arises.

The Career Stakes for CIOs

The Dataiku survey captures a second, less technical dimension of the agent sprawl problem: the personal stakes for CIOs navigating it.

• 88% of CIOs say their professional reputation or career trajectory will be shaped by their success with AI
• 76% believe their role as CIO will be at risk if their company fails to deliver measurable AI business gains by the end of 2027
• 72% say their AI budget is likely to be cut or frozen if performance targets are not met by the end of 2026

These numbers reflect a specific pressure dynamic: CIOs are being held accountable for AI business outcomes in organizations where AI deployments are happening outside their direct control. A CIO who cannot account for the AI agents running in their organization cannot credibly report on AI ROI, cannot confidently respond to a compliance question, and cannot manage AI risk on behalf of the board.

The SAFA self-regulation framework that Google, OpenAI, and Anthropic are building addresses governance at the model level — what frontier models can do, how they are tested, what safety standards apply before deployment. The Dataiku survey reveals the complementary problem at the deployment level: even if every frontier model is perfectly governed by its developer, the enterprise organizations deploying agents built on top of those models may have no inventory of what they have deployed, no measurement of whether it is working, and no governance framework for what it can do.

Dataiku Agent Management: What the Product Does

Against this backdrop, Dataiku announced Agent Management, a standalone product designed to inventory, monitor, and govern AI agents across the enterprise regardless of which platform built them.

The core capability is cross-platform discovery: Agent Management connects to the platforms that enterprise teams already use to deploy agents — AWS Bedrock Agents, Databricks Agents, Google Vertex AI Agent Builder, Microsoft Copilot Studio and Azure Foundry, Salesforce Agentforce, Snowflake Cortex, and Dataiku's own agent infrastructure — and scans them into a unified inventory. OpenTelemetry support allows custom agent environments to be included without a dedicated integration.

The product addresses three governance functions:

Inventory and classification. Agent Management enumerates all agents running on connected platforms, classifies them by risk tier — based on access to sensitive data, autonomy over customer interactions, and ability to execute consequential transactions — and surfaces the classification to the IT governance team. The classification output answers the question that most CIOs currently cannot answer: what agents do we have, what can they do, and what is the risk profile of each?

Performance monitoring. The product measures both technical performance (latency, error rates, availability) and business performance (task completion rates, outcome quality, cost-per-outcome). The business performance monitoring layer directly addresses the 72% of CIOs who cannot confirm business outcomes: if an agent is deployed to handle customer inquiries, Agent Management tracks whether those inquiries are being resolved, how often the agent escalates to human handling, and what the cost per resolved inquiry is.

Certification and evidence. For high-risk agents — those handling customer interactions, accessing sensitive data, or executing financial transactions — Agent Management maintains a standing record of certification status, named risks, and scheduled test results. When an auditor or regulator asks for evidence of AI governance practices, the evidence trail exists and is retrievable without a manual reconstruction effort.

Pricing is per instance annually, with monitoring metered per agent. General availability is October 2026.

Why Cross-Platform Visibility Is the Structural Challenge

The Dataiku Agent Management product's design addresses a specific structural problem: enterprise AI agent deployments are not platform-homogeneous. The same organization may simultaneously run agents in Salesforce Agentforce (customer-facing service agents), AWS Bedrock (backend data processing agents), Copilot Studio (workflow automation for Microsoft 365 users), and Snowflake Cortex (data analytics agents for the finance team).

Each of those platforms has its own monitoring interface. A CIO who wants a consolidated view of agent activity must aggregate four or more separate dashboards, each with different data models, different classification schemes, and different reporting timescales. The aggregation cost is high enough that most enterprise IT teams simply do not do it — which is a primary mechanism by which the 81% oversight gap exists.

Obsidian Security's $85M raise in September 2026 for AI agent runtime security addresses a related but distinct layer of the same problem: Obsidian focuses on behavioral anomaly detection and threat response for AI agents in production, while Dataiku focuses on inventory, lifecycle management, and performance measurement. The two products are complementary — runtime security without an inventory is incomplete governance; an inventory without runtime security is incomplete risk management.

The emergence of multiple products addressing different layers of agent governance in a compressed timeframe reflects accumulated demand that has been building since agent deployment accelerated in early 2026. Market observers have noted a wave of governance product launches in mid-to-late September — five products addressing different layers of enterprise AI control in less than two weeks, from agent inventory (Dataiku) to runtime security (Obsidian) to model governance (the SAFA initiative). The enterprise software market is building governance infrastructure for AI agents after the fact, responding to demand that has been latent since agent deployment outpaced governance capability.

How to Audit Your AI Agent Portfolio

For enterprise technology teams who want to address the 81% oversight problem in their own organization, a structured agent audit provides the governance foundation:

1. Identify all agent creation surfaces. Map every platform in your technology stack that allows non-IT employees to create agents: Microsoft Copilot Studio, Salesforce Agentforce, Power Automate, Zapier, Vertex AI Agent Builder, and any internally built agent frameworks. The inventory of platforms is the prerequisite for the inventory of agents.

2. Run a discovery scan on each platform. Most enterprise agent platforms provide administrative views of deployed agents. Export the agent inventory from each platform and consolidate into a single registry. For platforms without native administrative views, use API scanning or OpenTelemetry instrumentation to identify running agents.

3. Classify agents by risk tier. For each agent in the inventory, assess: Does it interact with external customers? Does it access personally identifiable information or sensitive business data? Can it execute irreversible transactions or commitments? Does it generate content attributed to the organization? Each "yes" increases the risk tier and governance requirements applicable to that agent.

4. Establish lifecycle management policies. Every agent should have a named owner, a defined business justification, and an expiration date or periodic review requirement. Agents without owners should be treated as ungovernanced — in most organizations, this is most of them — and should be subject to a review decision: confirm and assign ownership, modify to meet policy, or decommission.

5. Implement performance and compliance monitoring for high-risk agents. High-risk agents — those in risk tiers 2 and above — require ongoing monitoring of business outcomes and policy compliance. Establish baseline metrics at deployment (what should this agent achieve?), implement automated monitoring against those metrics, and define escalation procedures for when an agent deviates from expected behavior.

The Measurement Problem: 72% Cannot Confirm Business Outcomes

The most operationally actionable finding in the Dataiku survey is the measurement gap: 72% of CIOs cannot consistently confirm business outcomes from deployed AI agents. This is not primarily a governance problem — it is a product management problem. An AI agent deployed without defined success metrics is a project without a brief.

The AI customer success frameworks Signal has documented demonstrate the same pattern: enterprises that define specific outcome metrics before deployment consistently achieve higher ROI and better governance than enterprises that deploy AI and attempt to measure impact retrospectively.

The three-question framework for outcome measurement at deployment: - What business outcome does this agent improve, and what is the current baseline? - How will we measure the outcome, at what frequency, and through what mechanism? - What threshold of underperformance triggers a review or shutdown?

For a customer service agent, this translates to: current first-contact resolution rate is 67%; agent target is 75% FCR within 90 days; if FCR falls below 60% on agent-handled tickets in any two-week period, escalate to review. A governance framework built on specific, measurable, time-bounded outcomes is far more tractable than one built on general aspirations, and it directly addresses the 72% who currently cannot confirm whether their agents are delivering what they were deployed to deliver.

What Good Agent Governance Looks Like

The enterprises well-positioned as AI agent deployment continues to accelerate are those that establish governance infrastructure before the agent count grows further. Governance should operate at three levels:

Policy layer. Define who can create agents, under what conditions, for what purposes, and subject to what review. Most enterprises lack this policy document. A practical policy is brief: a one-page matrix of who can create what type of agent under which approval pathway is sufficient governance policy for most organizations at current scale.

Operational layer. Inventory, lifecycle management, and performance monitoring — the functions Dataiku Agent Management addresses. This layer answers operational questions: what do we have, is it working, and what needs attention?

Risk layer. Behavioral monitoring and anomaly detection — the functions that products like Obsidian Security address. This layer answers security questions: is any agent behaving in ways that violate its intended design, are there signs of compromise or manipulation, and is agent activity consistent with its authorization scope?

The governance maturity model for AI agents follows the same arc as cybersecurity maturity: discovery → inventory → classification → monitoring → response. Most enterprises are at discovery or early inventory stage today. Regulatory pressure will accelerate the arc: as AI incidents become more frequent and regulators begin asking governance questions directly, organizations that have not completed the inventory and classification stages will face compliance exposure on a timeline that is measured in quarters, not years.

Takeaway: The 81% CIO oversight gap is not a finding about negligent IT organizations — it is a finding about the speed at which AI agent deployment is outrunning the governance infrastructure that enterprise technology teams know how to build. Dataiku Agent Management represents the enterprise software market's first cross-platform response to the inventory layer of this problem: connecting to the major enterprise agent platforms and producing a consolidated view of what agents exist, what they do, and what risk they carry. The product addresses one layer of a multi-layer governance challenge; the risk layer (runtime security) and the policy layer (governance frameworks) require complementary investments. For enterprise CIOs navigating the 88% who say their career trajectory depends on AI success, the foundational step is understanding what AI is already running in the organization. You cannot govern what you cannot see — and right now, most enterprises cannot see most of what they have deployed.