The signal: Enterprise AI can spread faster than an organisation’s ability to inventory, secure and govern it. Employees may create assistants, connect automation tools or deploy agents to solve local problems before security teams understand what those systems can access. The result is not inevitable AI failure; it is a visibility and control problem that becomes more consequential as agents gain tools and permissions.
What agent sprawl actually means
Agent sprawl occurs when an organisation accumulates AI agents, assistants and automated workflows without a reliable record of their purpose, owner, connected tools, data access, permissions or lifecycle. It can emerge through formal product rollouts as well as employee-led experimentation. The key issue is not the number of agents alone; it is whether their capabilities and responsibilities are understood and controlled.
A simple summarisation assistant with access to a single approved document library has a different risk profile from an agent that can search email, update customer records, open support tickets and trigger financial or production actions. Yet when each integration is approved in isolation, teams can miss the combined authority an agent gains across systems.
Microsoft’s guidance on least-privilege agent design identifies recurring problems: ambiguous identity, permission creep, overly broad tool access, incomplete audit trails and slow revocation. These are practical governance failures that security and identity teams can measure—not evidence that every unregistered agent will cause an incident.
Source: Microsoft guidance on least privilege for AI agents.
Why the risk changes when agents can act
Traditional chat interfaces primarily return information to a user. Agentic workflows can also invoke tools, chain tasks and make changes in connected systems. That creates a different control problem: a mistaken interpretation or malicious instruction may lead not only to a poor answer, but also to a data export, unauthorised update, destructive action or sequence of tool calls.
Prompt injection is one example. Instructions embedded in an email, webpage, ticket or document can try to influence an agent that is processing that content. A secure design treats such material as untrusted data and enforces permissions outside the model. The agent should not be able to perform an action merely because a prompt or retrieved document tells it to do so.
The OWASP Top 10 for Agentic Applications 2026 highlights risks including goal hijacking, tool misuse, identity and privilege abuse, and failures in oversight. It is a useful threat-model starting point, not a claim that these risks will materialise in every deployment.
The inventory is the foundation of governance
An organisation cannot review what it cannot discover. Start with a practical inventory that covers internally built agents, vendor-provided assistants, low-code automations and workflows created through approved enterprise platforms. Include the business owner, technical owner, use case, model or service, connected systems, data classifications, identity, effective permissions, environment and current lifecycle status.
Do not rely on a spreadsheet as the only long-term control. Use available identity, endpoint, cloud, SaaS and application logs to discover usage, and establish a process for teams to register new agents before they receive sensitive data or write access. Where discovery is incomplete, document the uncertainty rather than declaring the environment fully governed.
Five controls that reduce agent sprawl
- Give every production agent an owner: name a business sponsor and technical owner responsible for purpose, access reviews, incidents and retirement.
- Use distinct identities and least privilege: avoid shared employee credentials and broad service accounts. Scope access to the specific resources and actions needed, and review combined permissions across tools.
- Restrict tool access: expose only approved integrations and actions. Require separate approval or time-bound elevation for high-impact operations such as deletion, bulk export, permission changes or payments.
- Log actions end to end: record agent identity, initiating user where relevant, tool call, target resource, effective authorization, approval state and result. A transcript of the model’s final answer is not a sufficient audit trail.
- Test revocation and retirement: confirm that disabling an agent also invalidates credentials and tokens, removes downstream access and stops scheduled workflows. Reassess permissions when tools, models, data sources or business purposes change.
Make risk-based rules—not a blanket ban
Blocking every AI tool may drive experimentation out of sight and prevent useful, low-risk work. Unrestricted access creates the opposite problem. A better approach is to classify use cases by data sensitivity, autonomy, reversibility and potential impact, then match controls to the risk.
- Low risk: summarising public material or drafting internal text without sensitive data or system write access. Use approved services, clear data rules and basic monitoring.
- Moderate risk: retrieving internal documents or drafting customer responses. Apply authentication, source permissions, data boundaries and review before external delivery.
- High risk: updating customer or employee records, changing production systems, making payment-related decisions or taking actions with legal or safety consequences. Require strong identity controls, narrow permissions, evaluation, explicit approval gates and a tested rollback path.
Measure control, not just agent count
Useful governance indicators include the share of production agents with a named owner and dedicated identity; the share with documented data and tool scopes; the proportion of high-impact actions protected by enforceable approval policies; the completeness of action logs; the time required to revoke access; and the number of overdue access reviews. These metrics reveal whether controls are operating. A raw count of agents, by itself, does not tell leadership whether the environment is safe or unsafe.
Organisations can map this work to the NIST AI Risk Management Framework resources, which support a lifecycle approach to governing, mapping, measuring and managing AI risks. The right implementation will depend on the organisation’s systems, threat model and risk tolerance.
A 30-day starting plan
- Week 1: identify approved AI platforms and likely sources of agent activity; agree on what qualifies as a production agent.
- Week 2: inventory the highest-impact agents and integrations, including owners, identities, data access and write permissions.
- Week 3: remove unnecessary privileges, restrict high-risk tools, improve logging and define approval thresholds.
- Week 4: test credential revocation, incident response and rollback; assign owners and dates for unresolved gaps.
This connects to FutureIsNow’s broader analysis of shadow AI and non-human identity governance, AI agent identity and enforceable limits and controlled execution in enterprise vertical AI.
Recent evidence reinforces why discovery must be treated as an operating control. Gartner’s April 2026 forecast says the average Fortune 500 enterprise could have more than 150,000 agents in use by 2028, compared with fewer than 15 in 2025; Gartner also reported that only 13% of organisations believed they had the right agent-governance arrangements in place. These are Gartner forecasts and survey findings, not a measured count of agents already deployed. Its recommended response—central inventory, named ownership, scoped identities and permissions, information governance, monitoring and retirement—closely matches the control sequence above. See Gartner’s six steps for managing agent sprawl.
The FINA view: visibility before autonomy
Agent sprawl is not solved by counting every bot or banning experimentation. It is solved by making each meaningful capability discoverable, owned, permissioned and observable. Start with the agents that can reach sensitive data or change important systems, then expand governance as the use of AI matures.
Bottom line: The enterprise goal is not zero agents. It is zero unmanaged high-impact authority. Give teams a safe path to build and deploy useful workflows, and make that path easier than operating outside the controls.



