The story enterprises tell about AI is a story of capability: can the agents do the work? By mid-2026, that question is mostly answered. The agents can do the work. Ninety-six percent of organizations are already running them. The success has been real and fast.

The story enterprises are not yet telling is what that success quietly created. Enterprises deployed 28.8 million AI agents in 2025. By the end of 2026, analysts expect them to be managing roughly eighty times that number. That growth didn't come from a plan. It came from thousands of individual teams each deploying agents that solved a real problem — and no one tracking the total. A $400 million collective cloud-spend leak surfaced across the Fortune 500 in April 2026, driven by agent sessions running without per-session cost ceilings. Gartner now forecasts that more than 40% of agentic AI projects will be cancelled by 2027, largely because of escalating costs and unclear value.

This is agent sprawl, and it is the same shape as a problem enterprises have seen before. It is shadow IT, returning in a faster and more expensive form.

Why Agent Sprawl Repeats the Shadow IT Pattern

Shadow IT — the unsanctioned SaaS tools that proliferate across an organization — happened because adopting a tool became easier than governing it. Agent sprawl happens for exactly the same reason, with the dynamics intensified.

Deployment outran inventory. A team finds an agent useful and deploys it. The deployment is easy, local, and solves a real problem, so it happens without central registration. Multiply by every team, and the organization ends up running agents it has no list of. You cannot govern, cost, or retire what you cannot see — and most enterprises currently cannot see their agents.

Agents consume resources continuously, not on a license. A shadow SaaS subscription had a fixed, visible monthly cost. An agent consumes compute every time it runs, and a poorly bounded agent can run far more than anyone intended. The cost is variable, invisible until the bill arrives, and — without per-session ceilings — effectively unbounded. That is what produced the $400 million leak.

Agents generate other agents. This is the dynamic shadow IT never had. A multi-agent system spawns sub-agents; an orchestrating agent delegates to specialists. Sprawl is no longer just humans deploying agents. It is agents deploying agents. The growth is compounding, and the eighty-fold projection is what compounding looks like.

Why FinOps Is the Discipline That Fits

The reflex when costs spiral is to call for tighter control — approval gates, deployment freezes. That reflex tends to fail, because it also kills the value the agents create. FinOps offers a better-fitting model, which is why FinOps for AI is now the top forward-looking priority for FinOps teams, with 98% of practices managing AI spend.

FinOps assumes decentralized decisions. FinOps emerged for cloud cost management precisely because cloud spending decisions are made by many teams, not one. It doesn't try to centralize the decisions. It makes them visible, attributes cost to the team that incurred it, and holds teams accountable. That is exactly the structure agent sprawl needs — many teams deploying, with visibility and accountability rather than a central gate.

FinOps balances cost against value, not cost alone. A pure cost-control approach asks "how do we spend less." FinOps asks "are we getting value for what we spend." That framing matters for agents, because the goal is not the cheapest agent fleet — it is retiring the agents that don't earn their cost while protecting the ones that do. Cost visibility without value visibility just tells you what to cut blindly.

FinOps is built around continuous review. Agent value is not static. An agent that was worth its cost last quarter may be redundant now, or running against a workflow that changed. FinOps's habit of regular, recurring review fits an agent fleet far better than a one-time approval at deployment, which is exactly the moment you know least about an agent's long-run worth.

Where This Shows Up Across the Organization

Agent sprawl is felt differently depending on where you sit, and the differences explain why it's so hard to address from any single seat.

Finance. Finance sees the aggregate AI bill rising and often cannot explain it, because the spend isn't attributed to anything. Agents run under shared service accounts; cost lands in an undifferentiated pool. Finance's first task is not to cut — it is to make spend attributable, so that questions about value can even be asked.

Engineering and platform teams. These teams deployed the agents and rarely see the cost of what they built. An agent that works is considered done; whether it's efficient is invisible to its builders. Closing that loop — putting cost data in front of the people who deploy — is one of the highest-leverage moves available.

Security and risk. Every untracked agent is an untracked access path. Agent sprawl is not only a cost problem; it is an attack-surface problem. An agent no one remembers deploying, still holding credentials no one reviews, is exactly the kind of exposure security teams exist to prevent.

Leadership. Executives hear that AI is working and that AI costs are alarming, and struggle to reconcile the two. The reconciliation is sprawl: the value is real and the waste is real, and they are not the same agents. Leadership's job is to fund the discipline that tells them apart.

What to Actually Do About It

Addressing agent sprawl is a governance build, and it works best in a specific order.

Build the inventory first. Everything else depends on knowing what you are running. Establish an agent registry — every agent, its owner, its purpose, its access, its cost. An organization that cannot list its agents cannot manage them, and most cannot. This is unglamorous and non-negotiable.

Put per-session cost ceilings on everything. The $400 million leak was caused by agents running without spend limits. Every agent should have a per-session and per-period cost ceiling. This single control prevents the most expensive failure mode and should not wait for a broader program.

Attribute cost to owners. Move agents off shared service accounts so that spend lands with the team that deployed the agent. Cost that no one owns is cost no one optimizes. Attribution is what turns a spend report into accountability.

Review on a schedule and retire deliberately. Set a recurring review — quarterly is reasonable — that asks of each agent whether it still earns its cost. Retiring agents has to become as routine as deploying them. An organization that only ever adds agents is guaranteed to drift back into sprawl.

The uncomfortable truth is that agent sprawl is a symptom of success. It happens because agents work, because deploying them is easy, and because every team has real problems they solve. The answer is not to make agents harder to deploy — that would forfeit the value. It is to make them visible, attributable, and reviewable, so the fleet can grow without the waste growing with it. The enterprises that build that discipline now will keep the agents that earn their cost. The ones that wait will be among the 40% explaining to their boards why the agentic program is being cancelled.