Essay · Technology

Agent Sprawl Is the New Shadow IT

Enterprises have been here before. The answer last time was not a ban or a tool. It was ownership.

Devendra KumarNovember 20266 min read

It is happening again

A decade ago, most large organisations discovered they were running far more software than their IT departments knew about. Teams had bought cloud tools on company cards. Data sat in services nobody had assessed. Contracts renewed without review. We called it shadow IT.

The same pattern is now forming around AI agents, and faster.

Agents are being built in low-code tools, switched on inside platforms the organisation already licenses, assembled by developers for their own workflows and bought as features in vendor products. Many of them can act, not just answer. They read data, call systems, raise tickets, send messages and update records.

Research published by OutSystems in April 2026, based on a survey of about 1,900 IT leaders, found that 96% of organisations already use AI agents in some form, and 94% are concerned that AI sprawl is increasing complexity, technical debt and security risk. Only 12% had a centralised platform to manage it. OutSystems sells an agent platform, so the exact numbers deserve some caution. The direction is hard to argue with.

Why agents are harder than shadow IT

Shadow IT was mostly a problem of data and cost. A team stored information somewhere it should not, or paid for the same thing twice.

Agents add a third problem: action.

An unapproved cloud tool holds data. An unapproved agent can do things with it. It can hold credentials to several systems at once, act at machine speed and keep running after the person who set it up has moved on.

A January 2026 survey of 228 IT and security professionals by the Cloud Security Alliance and Aembit shows how weak the foundations are. 68% of organisations could not clearly distinguish AI agent actions from human activity. 74% said agents routinely receive more access than their tasks require. 43% run agents on shared or generic service accounts. Asked who owns AI agent identity and access, no single function dominated, and 9% said nobody did.

Read that list as an executive rather than a security specialist. It says: we do not know how many agents we have, what they can reach, what they have done or who answers for them.

That is not a security gap. It is an accountability gap.

What we learned last time

One lesson from enterprise technology governance is worth revisiting whenever an organisation tries to regain control of its technology estate.

Finding an unauthorised application is relatively straightforward. Changing the behaviour that led someone to adopt it is harder.

Consider a familiar situation. Employees adopt an online tool because the approved process takes too long, the existing platform lacks a useful feature, or a team needs to solve an immediate business problem. The tool spreads because colleagues find it useful. By the time IT becomes aware of the application, people have built working practices around it.

The clean-up begins with an inventory. Applications are identified, risks assessed, owners contacted and decisions made about retention, replacement or retirement.

But removing an application addresses the symptom. It does not necessarily address the cause.

If employees return to the same slow approval process or the same capability gap, another tool will fill the vacuum.

The more sustainable response combines visibility with a better route to adoption. Employees need a straightforward way to request a tool, receive a timely decision and understand the conditions for approval. IT needs enough visibility to assess security, data protection, integration and support requirements.

The behavioural shift comes when the approved route becomes the practical route.

The lesson is simple: technology governance works best when doing the right thing is easier than working around the process.

A successful clean-up should therefore measure more than the number of applications removed. It should examine whether business teams use approved alternatives, whether recurring requests become easier to resolve and whether the original reasons for bypassing IT have been addressed.

Agents raise the same question, with higher stakes.

Ban agents and you lose sight of them

Banning agents will fail for the same reason it failed for cloud tools. People use them because they help. A ban moves them to personal accounts, browser extensions and features buried inside approved software, where nobody can see them at all.

Buy a control plane and you get visibility, not control

Agent management platforms are arriving quickly, and some will be useful. They can inventory agents, manage identities, enforce policies and log actions.

But a platform does not decide which agents should exist, what they are allowed to do, who is accountable when one gets it wrong, or when one should be retired. Those are management decisions. A tool can record them. It cannot make them.

An organisation that buys a platform before answering those questions will have a very good inventory of an ungoverned estate.

Give every agent a manager

Treat every agent that can act as part of the workforce, and give it a manager. In practice that means five things:

  • An inventory. Every agent that can read sensitive data or take action is registered, including those switched on inside vendor products. If it is not registered, it does not run.
  • A named business owner. Not the developer who built it and not the team that hosts it. The person accountable for the outcome it affects.
  • Its own identity. Each agent has its own credentials, scoped to its task. Never a shared service account or a borrowed human login, so its actions can be traced and its access withdrawn.
  • A boundary. What it may do on its own, what needs [human approval before it acts](/journal/before-an-ai-agent-touches-a-real-process), and what it may never do.
  • A retirement rule. When the owner moves on, the process changes or the value is not there, the agent is reviewed and switched off. Agents should not outlive their purpose.

Keep the controls proportionate to the risk

Giving every AI agent an accountable owner does not mean routing every action through a committee.

The goal is to establish clear boundaries before the agent starts working.

A practical approval model has three levels.

Level 1: Pre-approved actions

An agent works within a defined scope using approved data, tools and permissions. Routine, reversible actions proceed without individual approval. The system records what happened and provides a route for human review.

Level 2: Human confirmation

Actions affecting business records, external communications or operational workflows require confirmation from the designated owner before execution. The reviewer sees the proposed action, its supporting evidence and the expected consequences.

Level 3: Restricted or escalated actions

Actions involving sensitive data, financial commitments, privileged access, production changes or other high-impact decisions require stronger controls. Depending on the risk, the action is escalated for specialist approval or prohibited altogether.

Every level has a named accountable owner, defined permissions, an audit trail and a process for handling exceptions. Access is limited to what the agent needs, and permissions are reviewed as the use case evolves.

This structure keeps routine work moving while reserving human attention for decisions where judgement and accountability matter most.

The operating principle is straightforward: approve the boundaries in advance, monitor routine execution, and require explicit approval when an action crosses a defined risk threshold.

Governance should make safe execution repeatable. It should not force every low-risk action through the same approval process.

The question to take into the room

How many AI agents can act on our systems today, and could we name the owner of each one?

If nobody can answer the first half, start with the inventory. If nobody can answer the second half, the inventory will not help.

Agent sprawl is not a technology problem waiting for a better tool. It is shadow IT that can act. The fix last time was ownership. It will be again.

Sources

1. OutSystems (2026). Agentic AI Goes Mainstream in the Enterprise, but 94% Raise Concern About Sprawl. Press release for the 2026 State of AI Development report, 7 April 2026. Survey of about 1,900 IT leaders. Accessed 29 October 2026. https://www.outsystems.com/news/enterprise-ai-agent-report-2026/

2. Cloud Security Alliance (2026). AI Agent Identity Sprawl: The Enterprise Authorization Crisis. Research note, 20 June 2026, citing the January 2026 CSA and Aembit survey of 228 IT and security professionals. Accessed 29 October 2026. https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-agent-identity-sprawl-authorization-202/

TopicsAI AgentsGovernanceShadow ITAccountability

Follow the thinking.

New essays on enterprise technology and leadership decisions, about twice a month. No noise.

Your address is used only for these essays. Unsubscribe at any time.

Related

Read next.

Explore more perspectives →