For the last two years, the enterprise AI conversation has been dominated by the “pilot.” We asked if LLMs could code, if they could summarize meetings, or if they could draft emails. In 2026, that era is officially over. We have moved from managing individual AI pilots to managing autonomous AI “fleets.”
The scale of this shift is difficult to overstate. According to research from Gartner, by 2028, the average Fortune 500 company will have over 150,000 AI agents operating within its environment. To put that in perspective, most of these organizations currently employ fewer than 50,000 humans. We are entering the age of the “Invisible Enterprise,” where the vast majority of operational workflows are executed by non-human actors that do not appear on an org chart.
This rapid proliferation has a name: Agent Sprawl. It is the uncontrolled growth of autonomous agents deployed by different teams, on different platforms, with varying levels of data access and zero centralized oversight. It is the new Shadow IT, but with a critical difference: unlike a rogue SaaS app that just sits there, a rogue agent acts.
The “Sprawl Tax”: The Three Hidden Risks of Unmanaged AI
When I look at the current AI competitive landscape, it’s clear that the “Stack War” isn’t just about who has the best model; it’s about who can orchestrate the chaos. Organizations currently ignoring agent sprawl are paying a heavy “Sprawl Tax” across three specific vectors.
1. The Identity Crisis
Agents need credentials to do their jobs. They need to access CRMs, databases, and internal Slack channels. Currently, most enterprises allow agents to “borrow” human credentials or use loosely governed service accounts.
Research into AI Identity management highlights a terrifying gap: when an agent makes an irreversible machine-speed mistake, like deleting a production database or leaking PII, there is often no clear audit trail. If Agent A uses User B’s credentials to perform Action C, the traditional security stack sees User B. This creates a massive liability for compliance and cyber-insurance. Without a unique machine identity for every agent, you aren’t just scaling productivity; you’re scaling “untraceable” risk.
2. Logic Loops and Emergent Failures
In a recent case study involving a global logistics firm, two uncoordinated agents, one designed to optimize shipping speed and another designed to minimize fuel costs, became locked in a “logic loop.” The speed agent would book a premium carrier; the cost agent would immediately cancel it and rebook a slower route. By the time a human intervened, the company had incurred $2 million in “re-booking” fees and missed several critical delivery windows.
This is the “Translation Problem” in action. Agents operating in silos often work at cross-purposes because they lack a shared objective function. As multi-agent systems become more common (a trend extensively documented in recent Arxiv research on agentic coordination), these unmonitored interactions will lead to emergent failures that no human could have predicted.
3. Context Fragmentation (The Schizophrenic Brand)
Every agent has its own “memory” (vector store). When a customer talks to a Support Agent, then a Sales Agent, and later a Billing Agent, they expect a coherent experience. But in a sprawled environment, the Support Agent has no idea what the Sales Agent promised. This leads to what we call “Context Fragmentation”, a schizophrenic brand experience where the left hand of the AI doesn’t just ignore the right hand; it contradicts it.
Building the “DMV for AI”: A Governance Framework
You don’t solve sprawl by banning innovation. You solve it by creating a “DMV for AI Agents.” If an agent is going to operate on your company’s behalf, it needs to be registered, tested, and issued a license.
Effective governance in 2026 requires a two-tier approach:
The Centralized Agent Registry
Every agent, whether it’s a bespoke internal build or an “embedded agent” inside a SaaS tool, must live in a central registry. This registry should track:
- Purpose: What is this agent’s specific goal?
- Ownership: Which human is accountable for its actions?
- Permissions: Exactly what data can it read, and what systems can it write to?
- Lifespan: When does this agent expire or require re-certification?
Two-Tier Governance (Policy vs. Operations)
Central leadership (the CIO or CAIO) should set the Policy Layer: global rules around security, ethics, and “kill-switch” protocols. However, the Operations Layer, the day-to-day management of what the agent actually does, must remain with the domain experts. A Sales agent should be governed by Sales leadership, provided they stay within the guardrails set by the central policy.
Kill-Switches and Lifecycle Management
We are seeing a collapse in AI SaaS valuations because many “agents” are being replaced by native platform features. This means your agent inventory is constantly in flux. You need an automated “offboarding” process for AI. If an agent hasn’t been utilized in 30 days, or if its “owner” leaves the company, the system should automatically revoke its credentials.
The Board’s Mandate: 4 Critical Questions
Board members can no longer afford to ask generic questions about “AI strategy.” They need to dig into the operational reality of agentic density. I advise boards to ask leadership teams these four specific questions:
- “Who is the single point of accountability for our non-human workforce?” If the answer is “every department is doing their own thing,” you have a major governance hole.
- “Do we have a real-time inventory of every agent touching our customer data?” If it takes more than 10 minutes to pull this list, you are at risk.
- “What is our ‘Machine-Speed’ risk protocol?” How do we stop an autonomous agent from executing a million-dollar error before a human even realizes it’s happening?
- “How are we measuring the ROI of an agent vs. a human head?” If we are just adding agents without seeing a corresponding lift in organizational health or margins, we are just adding “work about work.”
Strategic Synthesis: From App-Centric to Agent-Centric
The Great Agent Sprawl is the inevitable growing pain of the most significant architectural shift in computing history. We are moving away from an era where humans “use” software (App-Centric) to an era where humans “delegate” to software (Agent-Centric).
The winners of this era will be the ones with the best Orchestration Layer. This means treating AI agents not as “tools,” but as a new class of digital employees that require the same, if not more, rigor in onboarding, management, and performance review as their human counterparts.
At LBZ Advisory, we help leadership teams bridge this “Translation Problem.” The goal is to move from a chaotic sprawl to a structured agentic fabric that drives genuine competitive advantage without compromising the integrity of the enterprise.
FAQ: The Strategic Reality of Agent Sprawl
Is Agent Sprawl just the same thing as Shadow IT?
No. While Shadow IT involved unauthorized software, Agent Sprawl involves unauthorized agency. A rogue SaaS app might store data insecurely, but a rogue agent can proactively interact with customers, spend company money, or alter codebases. The “blast radius” of an agent is exponentially larger than that of a static application because agents possess the ability to execute multi-step workflows autonomously.
How does Agent Sprawl affect my AI ROI?
Sprawl is a silent killer of ROI. When multiple departments build redundant agents to solve the same problem (e.g., three different “RFP assistants”), you are paying for redundant token usage, redundant compute, and redundant maintenance. Furthermore, fragmented agents create a “coordination tax” where humans have to spend more time fixing agent errors than they would have spent doing the task themselves.
Do I need a “Chief AI Officer” to manage this?
Not necessarily, but you do need a single point of functional accountability. Whether that sits under the CIO, CTO, or a new CAIO, the key is having a cross-departmental mandate. This person must have the authority to shut down non-compliant agents and the budget to build the shared “Context Layer” that all agents must plug into.
Related reading
Executive Keynotes and Board Briefings
Liat speaks to executive teams and boards about what AI changes in how a company operates, and what it does not.
