Deloitte’s 2026 State of AI in the Enterprise report, based on a survey of more than 3,000 IT and business leaders across 24 countries, put a number on something most technology leaders already sense: only 21 percent of organizations have mature governance in place for agentic AI systems. The other 79 percent are deploying agents — systems that take autonomous action — with governance frameworks built for a different era of software.
These are not pilots. These are systems in production.
quadrantChart title Agentic AI: Deployment Scale vs. Governance Maturity x-axis Low Governance Maturity --> High Governance Maturity y-axis Low Deployment Scale --> High Deployment Scale quadrant-1 Controlled scale quadrant-2 The risk zone quadrant-3 Early exploration quadrant-4 Governance-ahead Finance automation agents: [0.21, 0.76] Code review agents: [0.38, 0.64] Customer service bots: [0.45, 0.48] Internal RAG systems: [0.58, 0.42] Data validation agents: [0.64, 0.30]
The rundown: what the Deloitte data says
The survey covers what enterprises are actually doing with AI in 2026, not what they plan to do. Agentic AI adoption has accelerated — agents handling procurement, customer communications, code review, and data pipeline management are now common in organizations that were running pilot programs 18 months ago.
Governance has not kept pace. The 21 percent with mature governance have structured oversight of agent actions: defined authorization boundaries, logging of agent decisions, escalation protocols, and periodic review of whether deployed agents still operate within intended scope. The 79 percent majority have some of these, or none, or have governance in place for traditional AI models but not for systems that act autonomously.
The distinction the Deloitte data makes — between AI systems that make recommendations and AI systems that take actions — is the right frame. A model that surfaces a recommendation keeps a human in the decision loop. An agent that makes a purchase, sends a communication, modifies a record, or calls an external API does not. The governance failure mode for a poorly governed recommendation system is a wrong suggestion. The failure mode for a poorly governed agent is a wrong action — and that action may already be irreversible.
For the working software engineer
Agentic AI governance is an engineering problem before it is a policy problem. The policy layer cannot function without the technical foundation underneath it, and that foundation is largely absent in the organizations the Deloitte data describes.
What the technical foundation requires:
Action logging at sufficient granularity. If an agent sends a customer communication or modifies a record, you need to reconstruct that decision after the fact. What data did the agent see? What action did it take? What was the authorization context? Most agentic deployments today log outcomes, not decisions. That is not enough for incident reconstruction or compliance review.
Authorization boundaries enforced technically, not by convention. The scope of what an agent is permitted to do is often set informally at deployment and drifts over time. Configuration changes, new integrations, and prompt updates alter effective agent capabilities without updating the authorization model. Authorization boundaries should be enforced at the integration layer — the agent calls an internal API that enforces rules, rather than relying on the agent to remember its own scope.
Escalation paths for edge cases. Some decisions should not be made autonomously. High-value actions, novel situations, and actions with compliance implications should route to a human checkpoint. Building that checkpoint into the architecture at deployment is significantly easier than retrofitting it after an incident.
For business owners and operators
The 79 percent who lack mature governance are not there because they are careless. They are there because agentic AI moved from pilot to production faster than governance frameworks moved from “traditional software” to “autonomous systems.” The mental model most organizations bring to AI governance was built for recommendation systems, not action-taking systems.
The practical exposure breaks into three categories:
Reversibility. When a human acts on an AI recommendation, the decision can be reviewed before execution. When an agent takes an action, the review happens after — often after the action is irreversible. Refunds, external communications, system modifications, and API calls can be difficult or impossible to undo. Governance needs to account for which agent actions are reversible and which are not, and build approval gates accordingly.
Compliance. Agents that interact with customer data, financial systems, or regulated workflows operate in the same regulatory environment as humans doing the same work. If your compliance framework covers human-executed transactions but not agent-executed transactions, you have a gap that auditors will find.
Scope creep. Agent capabilities expand over time. New integrations, updated prompts, and configuration changes alter what an agent can effectively do without necessarily triggering a governance review. The action inventory set at deployment becomes stale. Mature governance includes periodic review of deployed agent capabilities against what was originally authorized.
My take
When I was working with LERETA, leading architecture across a team of 30+ developers on a multi-year modernization program, one of the most useful things I did early in the engagement was produce a set of wall-sized enterprise architecture diagrams that made the full legacy-modernization critical path visible to the board. The board could not make informed decisions about a $20 million investment in a system they could not see. The phrase I used internally: you cannot govern what you cannot see.
That framing applies directly to agentic AI. Organizations have deployed agents that make decisions and take actions in systems that leadership cannot fully observe. The governance gap Deloitte measured is not, at its root, a policy gap — it is a visibility gap. Leaders cannot audit, authorize, or correct agent behavior they have no instrumented view of.
The fix is not a governance framework document. It is building the observability layer that makes governance possible: action logging, enforced authorization boundaries, and decision trails that survive the agent’s context window. That is engineering work, and it should happen before scale, not after the first incident.
The 21 percent who have mature governance built that layer before the question became urgent. The 79 percent will build it in response to something they did not anticipate. The gap between those two paths is usually one incident.
