DM Explains · No. 017 · Governance
Agent sprawl: why your company will have 150,000 AI agents by 2028
Every team can now spin up an AI agent in minutes. Without a registry, a lifecycle and an owner, those agents multiply, duplicate, overreach and outlive their purpose — quietly.
· 13 min read
Read me first
Agent sprawl genuinely is coming for every large organisation, but it isn't the agents themselves that cause the damage — it's the missing registry that means nobody can say how many exist, who owns them, or what they can touch.
The expense-approval agent: 1 finance builds it with a low-code tool → 2 two other departments copy it, unknowingly duplicating → 3 it drifts outside review, gaining permissions quietly → 4 the policy it served changes but it keeps running, an orphan → 5 a registry audit flags it, moving governance from L1 toward L5, where the sprawl index falls 94.6% → 6 instead of banning agent-building, the two-tier committee catalogues and retires it on schedule
A manager in sales asks for an AI agent to draft follow-up emails. It takes an afternoon. Marketing already built one that does almost the same thing last quarter, and nobody told IT either existed. Multiply that across every team in the building, every quarter, for three years, and you get Gartner's forecast: the average global Fortune 500 enterprise goes from fewer than 15 agents in 2025 to over 150,000 by 2028 (Gartner, 2026). Only 13% of organisations believe their own governance is adequate for that scale. Most companies will hit six figures of agents blind.
Everyday picture. Amma lets everyone in the house bring home a helper robot whenever they want — Arjun gets one for homework, Meera for revision schedules, Dadi once asked for one to fetch milk while she was unwell. Months later three robots are watering the same garden, and Dadi's milk-robot is still showing up every morning though she's long recovered, quietly using up the household budget nobody is tracking.
Where the picture breaks. Robots in a house are physical and visible in the kitchen; enterprise agents are invisible software running inside dozens of applications, so nobody "notices the extra robot" the way Amma would.
Why it exists now
The problem isn't that AI agents are new — it's that building one no longer requires a project, a budget line, or a request to IT. Agent builders are now embedded directly inside the CRM, the ERP, the productivity suite everyone already uses. Any team can spin one up the way they'd spin up a shared document, with no central approval step. Creation outpaced tracking the moment that became true.
Gartner put a number on where that leads: fewer than 15 agents per average Fortune 500 enterprise in 2025, over 150,000 by 2028 (Gartner, 2026). Gartner's Max Goss has pointed to the consequences of letting that growth stay ungoverned: misinformation, oversharing, and data loss. Meanwhile a 2025 survey of 360 IT application leaders found only 13% believe they have the right governance in place, and just 15% were considering, piloting or deploying fully autonomous agents at all — meaning most organisations are at the very start of this curve, not the end of it (The Register, 2026).
What changed to make 2026 the moment this became urgent isn't a single model release — it's that agent-building tools crossed from "specialist capability" to "feature inside software everyone already has". The stakes rose because the thing that got easy to create is also the thing hardest to track: something that acts, not just something that sits on a shelf.
How it works
Anyone can build one
Low-code tools make agents as easy to spin up as a doc
What it is. Agent-building tools embedded in CRM, ERP and productivity suites let any team create an autonomous agent without going through a central request or approval process.
Why it exists. Vendors built these tools to remove friction — the same instinct that gave every team its own spreadsheet template now gives every team its own agent.
Where it breaks. It works fine with a handful of agents built by people who talk to each other. It breaks once dozens of teams do this in parallel with zero shared visibility between them.
How that's handled. A central inventory has to exist before growth, not after it — the point Gartner's six steps start from.
In technical terms
In technical terms: low-code agent builders inside common enterprise software let a non-specialist configure triggers, data connections and output actions without writing orchestration code or registering the result anywhere central.
Duplicates pile up
Different teams solve the same problem twice, unknowingly
What it is. Multiple agents across different teams end up doing identical or near-identical tasks because nobody can see what already exists.
Why it exists. Without a shared inventory, each team solves its own version of a problem from scratch, wasting compute and producing inconsistent outputs from differently configured copies.
Where it breaks. Obvious when the duplicated task is a single visible chore. Much harder to spot when two agents do overlapping work that looks superficially different on the surface.
How that's handled. An academic governance maturity model names this pattern explicitly — functional duplication — as one of five recurring sprawl patterns worth checking for by name, not by accident (arXiv, 2026).
In technical terms
In technical terms: functional duplication is one of five sprawl patterns (SP-01 through SP-05) catalogued by the maturity model, alongside shadow agents, orphaned agents, permission creep and unmonitored delegation.
Shadow & creep
Agents drift outside governance and gain access quietly
What it is. Agents get deployed outside IT oversight entirely, or stay inside it but quietly accumulate permissions beyond what they were originally granted.
Why it exists. Without a registration requirement or a periodic permission review, there is nothing forcing anyone to check what an agent can now reach versus what it could reach on day one.
Where it breaks. A household helper's job is visible to everyone in the kitchen; an agent's expanding data access is not, so the warning signs only arrive after something has already gone wrong.
How that's handled. Gartner recommends AI trust, risk and security management (AI TRiSM) tools specifically to discover and catalogue agents across both sanctioned and shadow deployments (Gartner, 2026).
In technical terms
In technical terms: AI TRiSM tooling scans for agent activity across sanctioned and unsanctioned deployments to build a discovery layer that a manual registry alone cannot maintain at scale.
Orphans never retire
Agents outlive their purpose and keep costing and risking
What it is. Agents built for a task that has since ended keep running indefinitely, because nothing forces their retirement.
Why it exists. There's no lifecycle step equivalent to "delete the project" — an agent doesn't know its job is done, and nobody is checking.
Where it breaks. Cancelling a forgotten household chore is trivial. Revoking a forgotten agent's access to a sensitive system is not, because nobody remembers the agent exists at all.
How that's handled. The maturity model names this pattern "orphaned agents" directly, flagging ongoing cost and security exposure for zero remaining value as its defining business impact (arXiv, 2026).
In technical terms
In technical terms: orphaned agents (SP-03 in the maturity model's taxonomy) retain their original credentials and permissions after their triggering task or policy has expired, with no automatic deactivation step.
Maturity cuts sprawl
A registry plus lifecycle rules shrink the problem sharply
What it is. Building a centralised agent inventory, defining identity and permissions and lifecycle rules, governing data access, and monitoring behaviour together make up a governance maturity ladder.
Why it exists. Each rung closes one of the gaps the earlier steps describe — the registry solves duplication, the lifecycle rules solve orphaning, the permission review solves creep.
Where it breaks. The numbers behind this come from a modelled analysis built to test a proposed framework, not from measured outcomes across real organisations that have actually climbed the ladder yet.
How that's handled. Even as a modelled result, the direction is stark: moving from the least to the most mature governance level cut a modelled sprawl index by 94.6% and risk incidents by 96.5% (arXiv, 2026).
In technical terms
In technical terms: the Agentic Enterprise Governance Maturity Model defines five levels (L1 to L5) across inventory, identity/lifecycle, data governance and monitoring dimensions, with the largest single reduction attributed to automated sprawl detection.
Govern, don't block
Banning agents just pushes people to unsanctioned ones
What it is. Organisations that sanction agent use and train staff on it, rather than banning agents outright, keep adoption inside governance instead of pushing it underground.
Why it exists. Blocking a capability people need doesn't remove the need — it just removes the visibility into how that need gets met.
Where it breaks. This comes from a single vendor survey correlation, not a controlled study, so it shows association rather than proof of cause.
How that's handled. Organisations that expanded access broadly under strong governance were 3.3 times more likely to report higher value from generative AI tools than those restricting access to trusted users only (The Register, 2026).
In technical terms
In technical terms: Gartner's recommended structure is two-tier — a centralised AI governance committee sets policy, operational governance teams embedded in each application domain enforce it.
The expense-approval agent: 1 finance builds it fast, no registry entry → 2 two other departments copy it, becoming shadow/orphaned → 3 it keeps permissions after the policy that created it changes → 4 AI TRiSM finally discovers it among the fleet → 5 the two-tier committee assigns an owner and policy → 6 scored L2, re-governed toward L5, where Gartner's model shows sprawl index down 94.6% and risk incidents down 96.5%
Compared with what came before
| Traditional IT/SaaS sprawl | Agent sprawl | |
|---|---|---|
| What accumulates | Unused software licences, dormant accounts | Autonomous agents that act, chain into other agents, and acquire permissions dynamically |
| Visibility problem | Who's still paying for what | Who can still do what, with what data |
| Worst outcome | Wasted spend | Wasted spend plus live security and compliance exposure |
| Fix that doesn't work | Freezing all new purchases | Blocking all agent creation, which pushes use into shadow AI |
Agent sprawl isn't a bigger version of software sprawl — it's a different kind of risk entirely, because the thing left unmanaged can act on its own, not just sit idle on a licence sheet.
Where it breaks
Governance maturity reduces sprawl, but only when it's an actual registry with lifecycle rules enforced, not a policy document nobody checks against.
Modelled, not yet measured
The 94.6% sprawl-index reduction and 96.5% drop in risk incidents come from a modelled analysis built to test a proposed governance maturity framework — not a measurement of real enterprises that have actually climbed from level one to level five.
Restriction creates the worse problem it was meant to prevent: organisations that block agent use outright don't get fewer agents, they get agents nobody can see, built with tools IT never approved. The correlation data behind "govern broadly, don't block" rests on one vendor's survey of 360 leaders — real signal, not proof.
Numbers to carry
| Figure | Value |
|---|---|
| Agents per average Fortune 500 firm, 2025 | fewer than 15 |
| Agents per average Fortune 500 firm, forecast 2028 | over 150,000 |
| Organisations confident in their agent governance | 13% |
| Leaders considering/piloting/deploying fully autonomous agents | 15% (n=360) |
| Value uplift from broad access + strong governance vs. restricted access | 3.3 times more likely to report higher value |
| Sprawl index reduction, least to most mature governance | 94.6% (modelled) |
| Risk incident reduction, least to most mature governance | 96.5% (modelled) |
| Composite net business value improvement, same comparison | 51.0% (modelled) |
Myths worth killing
| Commonly repeated | What is actually true |
|---|---|
| "Agent sprawl means too many AI tools, like SaaS sprawl" | Agent sprawl is distinct because the unmanaged thing can act autonomously and chain into other agents, not just sit unused |
| "Blocking agent-building tools solves the risk" | Gartner's and The Register's reporting both point to blocking pushing use into unsanctioned shadow AI, which carries greater risk |
| "This is a future problem, not a 2026 one" | 15% of surveyed leaders were already piloting or deploying fully autonomous agents as of the 2025 survey |
| "Governance just means writing a policy" | The maturity model ties reduction to concrete mechanics — inventory, lifecycle rules, data governance, monitoring — not a document |
| "Shadow AI and agent sprawl are the same thing" | Shadow AI is unsanctioned tools used with zero visibility; agent sprawl includes sanctioned agents too, and ungoverned sprawl is what pushes people toward shadow AI |
Worked example
- Finance builds an agent to auto-approve small expense claims under a cost policy that was active at the time.
- Finance's policy changes months later, raising the auto-approval threshold, but nobody updates or retires the original agent.
- Two other departments, unaware the finance agent exists, each build their own near-identical approval agent rather than reusing or asking about one — illustrative, but exactly the duplication pattern the maturity model names.
- The original agent keeps running, approving claims against an outdated policy, and over time its access gets broadened by whoever last touched its configuration, without a corresponding review.
- Nobody owns it. No registry entry means no one is prompted to check whether it should still exist, what it can now access, or whether its outputs still match current policy.
- A registry step, had it existed from the start, would have surfaced the duplicate builds at step 3 and flagged the policy mismatch at step 2 — this is the gap the maturity model's "centralised inventory" dimension is built to close.
Should you use it?
The question isn't whether to allow agents — Gartner's own data suggests restricting access correlates with lower reported value, not more safety. The real question is whether a registry, an owner, and a retirement rule exist before the next agent gets built, not after the hundredth one does.
Start now if agents are already being built across more than one team without a shared list of what exists. Hold off on heavier tooling if you're still at a handful of agents built by one team that already reviews its own access — but build the habit of logging each one anyway, because the curve from 15 to six figures doesn't announce itself gradually.
The cheapest first experiment: list every agent currently running across the organisation, who owns it, and what it can access — before building any new governance tooling at all. If that list is hard to compile, that difficulty is the finding.
Say it in the room
"Allow agents broadly, but only inside a registry with an owner and a retirement date — blocking them just makes the risk harder to see, not smaller."
Lines to carry
- "Creation is easy; tracking is the hard part."
- "An agent without an owner never retires itself."
- "Blocking agents just makes them harder to see, not fewer."
- "Sprawl isn't the agents — it's the missing registry."
- "Governance maturity isn't a policy document, it's a measurable number."
FAQ
Is agent sprawl just a bigger version of shadow IT?
Not quite. Shadow AI — tools used with zero organisational visibility — is one slice of agent sprawl, but sprawl also includes agents built with full sanction that still go untracked, duplicate each other, or outlive their purpose. The distinguishing feature is that agents act autonomously and can chain into other agents, which plain unused software never does.
Does blocking agent-building tools actually make things safer?
The evidence points the other way. Reporting on Gartner's survey found organisations that sanctioned and trained staff on agent use, rather than restricting it, saw broader adoption under governance and higher reported value — 3.3 times more likely, in that survey. Blocking tends to push the same need toward tools nobody can see.
How many agents does a typical company actually have right now?
Gartner's figure for the average global Fortune 500 enterprise was fewer than 15 as of 2025, expected to pass 150,000 by 2028. Smaller organisations will sit on a different curve, but the shape — rapid, decentralised growth with no central count — is the pattern worth watching for, not the exact number.
Is the 94.6% sprawl reduction a real, measured result?
No — it's a modelled figure from an academic maturity model, built to test a proposed framework rather than observed across real companies that have climbed from the lowest to the highest level. Treat it as a strong directional signal, not a guarantee.
What's the first concrete step, if we're starting from nothing?
Build the inventory before anything else: every agent, its owner, its data access, and whether its original purpose still exists. Gartner's six-step guidance and the academic maturity model both start from the same place — you can't govern what you haven't counted.
Glossary
- Agent sprawl — the uncontrolled growth of AI agents across an organisation, built without centralised tracking, oversight or lifecycle management.
- AI agent — a software system that can take autonomous, multi-step actions on behalf of a user or process, often without approval at every step.
- Shadow AI — AI tools or agents used by employees without the organisation's knowledge or sanction, bypassing official IT controls.
- AI TRiSM — AI Trust, Risk and Security Management, a category of tools for discovering, governing and securing AI systems including agents.
- Shadow agents — agents deployed outside IT governance without being registered or approved, creating security blind spots.
- Orphaned agents — agents whose original task or purpose has ended but which continue to run, consuming resources and posing ongoing risk.
- Permission creep — the gradual, often unnoticed accumulation of access rights by an agent beyond what it was originally granted.
- Functional duplication — multiple agents built separately by different teams that end up performing the same or very similar tasks.
- AI governance committee — a centralised group, typically including the CIO, CISO and legal leaders, that sets organisation-wide policy for AI use.
- Agent lifecycle management — tracking an agent from creation through active use to eventual retirement, so it does not persist unnecessarily.
- Governance maturity model — a framework that ranks an organisation's AI governance capability across levels, from ad-hoc to fully optimised and automated.
- Least-privilege access — a security principle of giving an agent or user only the minimum permissions needed to do its job.
- Agent inventory/registry — a centralised catalogue listing every AI agent deployed in an organisation, including its owner, purpose and permissions.