Agent Sprawl as an Organizational Risk

Every team can deploy an agent in an afternoon. Almost no organization can tell you how many are running, what systems they touch, or who owns the judgment when they act.

In May, the Wall Street Journal gave enterprise AI a new problem statement. Not model quality. Not ROI. Too many agents. As platforms like Anthropic’s Claude Cowork made it easy for employees to spin up task-specific bots, companies including Lyft, DaVita, and GitLab found themselves navigating something the paper called “AI agent sprawl”: overlapping automations, duplicated workflows, and IT departments trying to govern a population that was multiplying faster than anyone could catalog.

The numbers stopped being abstract quickly. The Journal reported that FICO’s 3,500 employees were creating dozens of new AI agents every day across the hierarchy. DaVita’s chief information officer said the kidney-care company had passed 10,000 employee-built agents. GitLab’s leadership expected short-term sprawl and was betting the opportunity outweighed the mess. Lyft, a publicly traded company with heavy regulatory obligations, was building centralized controls so agents could share approved skills instead of reinventing them in isolation.

None of these firms were anti-AI. They were pro-AI and suddenly responsible for a footprint they had not planned to inventory.

When creation outruns the catalog

This is shadow IT returning with upgrades. For twenty years, IT learned to ask what was running, not only what was approved. Spreadsheets became systems of record. Macros became production logic. Personal databases became customer registries. The pattern was always the same: capability arrived through motivated individuals long before governance caught up.

Agents compress that timeline. Shadow scripts forgot what they did when the author changed teams. Agents remember. Shadow tools waited for a human click. Agents chain actions across systems. Shadow IT was annoying. Agent Sprawl is cumulative.

The Journal framed the immediate risk plainly: too many agents, especially multiple agents performing the same tasks, creates cybersecurity and management problems for corporate IT. That is the turn. The enterprise AI program still lives on a slide deck with twelve approved pilots. The actual agent population lives in Cowork folders, custom GPTs, API chains, and half-documented experiments someone built because the official path took eleven weeks.

When Understanding-as-a-Service argued that employees augmented by agents would generate more context and potential actions than leadership could process, this is the organizational version arriving in headlines. Decision-making becomes the constraint. Not model access. Not budget. The ability to metabolize autonomous capability into accountable practice.

Agent Sprawl: shadow IT with memory

Here is the frame those boardrooms needed before the meeting moved on.

Agent Sprawl is the unmanaged proliferation of AI agents across teams without a shared inventory, without shared metabolism, and without named judgment ownership. It is not the same as having many tools. It is many semi-autonomous actors whose permissions, memories, and failure modes nobody has mapped to the org chart.

Three differences from classic shadow IT change the risk profile:

1. Memory. An agent accumulates context across sessions. It learns phrasing, shortcuts, exceptions. Sunset the project and the pattern may survive in someone else’s clone.
2. Autonomy. Low-risk assistants graduate quietly. A summarizer gains send permissions. A research agent gains write access. Each step feels incremental. The compound is not.
3. Judgment debt. Every agent output that becomes a decision orphans accountability unless someone is explicitly on the line. Tool owners are not judgment owners.

Your adoption dashboard measures willingness to try: pilots launched, licenses purchased, training hours completed. It does not measure what is actually running. Gartner research cited alongside the Journal reporting suggests the gap will widen: Fortune 500 companies may manage more than 150,000 agents within two years, while only 13% of organizations say they have adequate agent governance in place today.

That gap has a name we already use elsewhere. The Digestion Gap describes what enters an organization versus what it can act on. Agent Sprawl is capture without absorption at machine speed.

What Verizon teaches about metabolism

The reflex is predictable. Discover sprawl, declare a freeze, route everything through a central gate, and wait for the backlog to teach patience. That response confuses visibility with prohibition.

Verizon offers a different case. Reporting from Constellation Research on the company’s Google Cloud Next 2026 briefing, consumer AI vice president Anil Kumar described agents as a workforce-wide strategy, not a niche tool: “If everybody is using agents then we can get much more done versus very few people using them.” Verizon encourages employees to build low-fidelity agents while walking a line against a Wild West scenario. Mission-critical agents register in advance. A LangChain control plane gives the company one layer across data stores and clouds.

The operational insight is usage, not intent. Kumar said actual adoption is the filter: agents with repeated use graduate to broader enterprise scope with stronger governance. Agents without use get retired. “If there’s no usage, we don’t want to run the [agent], because it’s a waste of computing, waste of resources,” he told analysts. Constellation summarized the cycle: encourage creation, observe usage, promote high-usage agents, retire low-usage ones, apply stronger governance to sensitive workloads. Managing sprawl, in Kumar’s telling, means creating agents and then paring back repeatedly.

That is organizational metabolism in plain language. Not a moratorium. A rhythm leadership can govern.

Scale with safety, or scale with sprawl

DaVita’s CIO told the Journal the company had to “scale with safety” given patient-care obligations. That phrase is the whole argument in three words. FICO can absorb dozens of new agents per day only if someone owns what those agents touch, what they remember, and who answers when output becomes decision. GitLab can tolerate short-term sprawl only if guardrails hold while consolidation catches up.

The problem is not too many agents. It is too little metabolism.

If you cannot name who owns judgment when the agent acts, you do not have a pilot. You have liability dressed in innovation language. The forward path is not slower deployment. It is faster integration: see what is running, assign ownership, absorb what works, retire what does not. Stepstones and protocols, not a twenty-page operating model before a list, as we argued when leading through the AI shakeup.

When every job becomes, in part, an IT job, agent-inclusive design is already here. The question is whether your organization can tell you how many are running, what they touch, and who owns the judgment when they act. The Wall Street Journal named the problem in May. Verizon is experimenting with the response. Most everyone else is still counting pilots.

The Thread: This continues the thread from Understanding-as-a-Service, where we argued that absorption capacity, not model access, is the binding constraint; Agent Sprawl is what that constraint looks like when the headlines start reporting agent counts instead of model releases.

Forward this to: the transformation lead whose enterprise AI dashboard still counts pilots while FICO counts dozens of new agents per day.

A Question for You: Does your organization know its agent count, or only its pilot count?

Madam I’m Adam