Skip to the cards
Filed
Loans
916

AI Agent Sprawl: The Cost, Access Risks and Controls IT Teams Need to Track

Untracked agents can multiply token, compute and licence costs while retaining data access. Inventory, ownership and lifecycle reviews help teams find the overlap.

AI Agent Sprawl: The Cost, Access Risks and Controls IT Teams Need to Track

AI agent sprawl can quietly turn a useful pilot into an ongoing bill and an access risk. An agent may keep consuming tokens and compute, and retain permissions to business data, after the team that created it has moved on. For IT and finance teams, the challenge is finding those agents across cloud and SaaS platforms and connecting each one to an owner, a purpose, its costs and the data it can reach.

/articles/90/7jW3auOkoX7HfSPsuZ7Bof95GsIzLgOQ.webp

The projected scale makes manual tracking a poor fit. Gartner projects the average Fortune 500 enterprise will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025. Yet only 13% of organizations believe they have adequate governance for that growth. Those figures point to a practical problem: spending can rise long before an organization has a reliable count of what is running.

There’s a second cost to account for when agents operate beyond IT’s view. IBM’s 2025 Cost of a Data Breach Report found that breaches involving shadow AI cost $670,000 more on average, and that 97% of organizations with breaches involving their AI models or applications lacked proper AI access controls. The operational response is to connect discovery, cost tracking and access review, then make it routine to retire agents that no longer have a clear purpose.

Why agent sprawl differs from unused software

AI agent sprawl is the unchecked accumulation of agents, chatbots, copilots and autonomous workflow tools across an organization’s cloud and SaaS environments, without a shared inventory, a named owner or recurring access reviews. It shares some features with ordinary software sprawl, but an agent can keep taking actions after its original project has lost momentum. It may continue making model calls, using compute and reaching data under permissions nobody has reviewed recently.

An unused software seat can continue to cost money while sitting idle. A forgotten agent can also carry permissions and make calls to a model. That difference changes what IT teams need to track: a licence count alone won’t show which agents are still running, who is responsible for them, or whether their access remains appropriate.

Consider three teams building an agent for similar work. A support team creates a Copilot Studio agent to triage tickets; a data team builds a Vertex AI agent for the same task using another data source; and a sales operations team deploys an Agentforce agent for a related workflow. If the teams don’t know about each other’s projects, the organization may pay for overlapping capabilities, tokens and compute while maintaining separate access paths to customer data.

That overlap can be difficult to spot in ordinary licence reports. Agent usage may appear in different platform consoles or cost categories, and the reports may not identify the business process each agent serves. The practical unit of review is therefore the agent itself: its owner, purpose, platform, usage, permissions and review date.

How growth creates a budget and governance gap

Gartner’s projection of more than 150,000 agents at the average Fortune 500 enterprise by 2028, compared with fewer than 15 in 2025, describes a sharp change in scale. Gartner also projects that agents pursuing misaligned goals or acting outside their intended scope will be the top AI concern for 40% of Fortune 1000 companies by 2028. In both cases, a complete inventory is a starting point: without one, teams can’t reliably match costs and access to the agents that generate them.

The financial effect may first show up as an unexplained increase in token, compute or licence spending. The operational cause can be less obvious: parallel projects, pilots that never ended, custom agents layered on top of existing subscriptions, or tools employees adopted without IT’s knowledge. In each case, the organization needs to establish whether an agent is still needed before deciding whether to keep its spend or access.

Deployment delays offer another indication that governance and data management are unresolved. AvePoint’s State of AI 2026 report found that 86% of organizations delayed AI agent deployments by an average of nearly six months because of data security and data management concerns. That finding doesn’t mean every delay has the same cause, but it illustrates how gaps in ownership, access and data readiness can affect adoption as well as risk management.

Three sources of sprawl

Unassigned identities and lifecycles

Agents are often created faster than teams can assign them an owner, define their permission scope or set a retirement date. A pilot intended to run for three months can remain active much longer if nobody is responsible for reviewing it. Its original access may remain in place even when the work that justified that access has ended.

AvePoint’s State of AI 2026 report found that roughly one in three employees already has access to sanctioned or unsanctioned agent-creation tools, and that the figure is expected to pass one in two within 12 months. The report also found that 21.1% of organizations cannot say whether unsanctioned agent-creation tools are in use in their environment. For IT teams, that uncertainty limits the value of an inventory built only from platforms they already manage.

Accumulating access to data

Context and data-access sprawl occurs when agents retain permissions to more data sources than they need, or when nobody checks whether those permissions still make sense. Each additional source can widen the consequences of an unwanted action. Broader context and extra retrieval calls can also increase token use, so data access and operating costs can move together.

Gartner’s six-step framework for managing agent sprawl includes AI information governance: review agent access, keep it current and archive access when it is no longer required. This turns permissions into part of the agent’s lifecycle instead of a one-time setup task.

Rachel Simon headshot

AvePoint’s State of AI 2026 findings report that 88.4% of organizations experienced at least one AI-agent-related security breach in the past year. Data leakage was the most common cause, ahead of malicious manipulation of the agent itself. These figures make access reviews a practical control for limiting exposure, especially when agents can retrieve information from several sources.

Shadow AI outside IT’s view

Shadow AI refers to tools, applications or agents employees adopt or deploy without IT’s approval, oversight or visibility. It is a particularly hard part of sprawl to manage because there may be no central record, owner or established review process. Until teams discover the deployment, they may not know what it costs or what data it can access.

AvePoint’s State of AI 2026 report found that the share of organizations unable to say whether employees use unsanctioned generative AI tools nearly tripled year over year, from 6.3% to 17.6%. Gartner separately projects that by 2030, more than 40% of enterprises will experience a security or compliance incident tied directly to unauthorized shadow AI. Those findings make discovery an operating task, not just an exercise for the next audit: an agent that isn’t on the list can’t be assigned an owner or reviewed.

What the incident cost figures show

Agent sprawl has ongoing costs even when no incident occurs: token consumption, compute, licence seats and the staff time needed to support overlapping systems. The incident figures provide another way to understand the potential expense, though they don’t measure the day-to-day cost of every ungoverned agent.

IBM’s 2025 Cost of a Data Breach Report found that incidents involving shadow AI accounted for 20% of breaches, seven percentage points higher than incidents involving sanctioned AI. Organizations with high levels of shadow AI had breach costs that were $670,000 higher on average than those with low or no shadow AI. In these incidents, personally identifiable information was exposed in 65% of cases, while intellectual property was exposed in 40%.

The report also found that 97% of organizations whose breaches involved their AI models or applications lacked proper AI access controls. These figures describe reported breach findings; they don’t provide a per-agent cost estimate. They do show why organizations should put agent access and ownership into their regular operational reviews instead of treating agent spend as a separate finance issue.

Build an inventory across platforms

A useful inventory covers every environment where employees can create or run agents, including platforms outside IT’s usual deployment path. For each agent, record a named owner, a business purpose, its data-access scope, its usage and costs, and a date for its next review. Include unsanctioned deployments once discovered so the inventory reflects what is actually running.

  • Search across platforms. Include Microsoft 365 products such as Copilot Studio, Microsoft Foundry and SharePoint agents; Google Cloud’s Vertex AI; and Salesforce’s Agentforce. A review limited to one platform can miss duplicate work elsewhere.
  • Identify unsanctioned agents. Record agents created outside IT’s usual process, then assign an owner and review their access. They need to be found before they can be managed.
  • Require an owner for every agent. A named person or team should be accountable for the agent’s purpose, access and continued operation.
  • Keep access and spend in the same record. A cost report alone doesn’t show whether an agent should still be allowed to reach its current data sources. Review the two together.
  • Set a recurring review date. Ask owners to confirm on a fixed schedule that the agent remains necessary and its access is still appropriate.

Discovery should be continuous where possible, with a full review on a recurring schedule. A manual list can become incomplete as teams create new agents. The more quickly an environment is growing, the less an organization can rely on a one-time inventory or on employees remembering to report every deployment.

/articles/90/QblGF3SoZshpSTxXuBCOVc89wT31AX1M.webp

Keep governance from becoming a deployment bottleneck

Governance can support deployment when it is part of the agent lifecycle. Teams should know who can create agents, how owners are assigned, which data sources can be connected and when reviews happen. That gives project teams a working path to launch while helping IT maintain an accurate picture of what is active.

  1. Set identity and lifecycle rules before launch. Define creation rights, ownership, data connections and review timing as part of the build process.
  2. Automate discovery where available. Continuous discovery can help keep an inventory current as agents are added across platforms.
  3. Alert owners to usage against expected budgets. The source recommends alerts at 70 to 80% of expected token or compute spend, giving owners a chance to investigate before costs exceed expectations.
  4. Plan retirement when creating the agent. A review date gives the owner a prompt to justify keeping the agent and its access.
  5. Review cost and access together. Separate finance and security reviews can leave teams with partial records of the same agent.

A hard usage cutoff can disrupt a workflow that is still needed. The suggested 70 to 80% alert threshold is a prompt for an owner to investigate usage; teams should set expectations based on the agent’s role and workload rather than apply a flat cap to every agent.

What different tool types cover

Identity tools, security posture tools and governance platforms can each contribute to agent management, but their capabilities differ. Identity tools may register agents built within their own ecosystem, with more limited coverage across other platforms. Security posture tools may flag risk signals without maintaining a live inventory. The comparison below reflects the capabilities described for each category.

CapabilityIdentity toolsSecurity posture toolsGovernance and recovery platforms
Agent discoveryMay register agents within their own ecosystem; third-party coverage can be limited.Can scan for risk signals but may not maintain a living inventory.May provide continuous discovery across Microsoft 365, Google Cloud and Salesforce.
Cost visibilityNot typically a native capability.Not typically a native capability.May show token, compute and licence spend per agent.
Access governanceMay register basic permissions, with limited recertification.Can offer risk scoring, with limited lifecycle enforcement.Can bring ownership, access reviews and recertification into one workflow.
Backup and recoveryNot typically a native capability.Not typically a native capability.May restore agent configuration or data affected by an unwanted change.

These categories aren’t interchangeable, and organizations should check the capabilities of the products they already use. The operational requirement remains the same: teams need to identify agents across their environment, know who owns them, understand their access and spend, and have a process for removing them when they’re no longer needed.

Why cross-platform reviews matter

Microsoft 365, Google Workspace and Salesforce each offer ways to build or run agents. Their management consoles don’t automatically give teams a single view of agents in the other environments. That makes it possible for separate departments to build agents for similar work without seeing the combined cost or access footprint.

In Microsoft 365, teams may need to account for Copilot Studio agents, Microsoft Foundry projects and SharePoint agents created by site owners, alongside Copilot licences. In Google environments, the inventory may include Vertex AI agents and Gemini-based workflow automations. In Salesforce, it may include Agentforce agents created by sales operations teams. The exact inventory depends on what an organization has deployed, but reviewing each platform in isolation can hide duplicated workflows across the SaaS estate.

Licence management has a related blind spot. An organization may track the cost of a Copilot, Gemini or Agentforce seat while missing usage-based costs from agents built on top of the platform. It may also continue paying to run an agent after the project that created it has ended. Retirement and recertification give owners a regular chance to remove agents that no longer have a business purpose.

Set a governance cadence that fits the environment

A small number of agents may be tracked informally by the teams that built them, but that approach provides little visibility outside those teams. As the count rises, organizations may begin tagging agents by team or project and reviewing costs periodically, while ownership and access reviews remain inconsistent. At larger scale, a reliable inventory needs automation and a fixed process for reviewing ownership, access, cost and retirement together.

The tiers sometimes used to describe this progression are illustrative benchmarks, not the results of a single-sourced study: early environments are described as having roughly fewer than 50 agents; scaling environments as roughly 50 to 5,000; and enterprise sprawl risk as 5,000 or more, trending toward Gartner’s projection of 150,000 agents by 2028. Organizations should use their own environment to set thresholds and review frequency.

For teams managing more than a few dozen agents, a quarterly full inventory review can provide a recurring checkpoint, with continuous discovery between reviews. Fast-growing environments may need a tighter cadence while the count is rising. The goal is to ensure an owner can confirm that each agent is still needed and that its access matches its current purpose.

Connect inventory, cost and recovery

Useful agent governance treats inventory, ownership, cost and recovery as connected operating work. Start by finding the agents, including shadow deployments, then require a named owner and business justification. Review platform overlap across Microsoft 365, Google Workspace and Salesforce together. Set recurring checks for usage and access, and establish how important configurations and related data can be restored after an unwanted change.

Recovery deserves attention because an agent can alter data or configuration as part of its work. If an organization cannot restore an important agent configuration or recover data affected by an unwanted change, it may have to rebuild the agent and its connections. The source describes agent backup and recovery as a way to treat configuration and the data the agent touched as recoverable assets; teams should assess what their existing tools actually protect.

Governance maturity is less about reaching a particular agent count than being able to answer basic operational questions: What agents are running? Who owns each one? What data can it reach? What is it costing? When was it last reviewed? If those answers require teams to reconcile separate spreadsheets and platform consoles, the inventory process needs attention.

Common questions from IT teams

How does agent sprawl differ from shadow AI?

Agent sprawl is the wider accumulation of agents, whether sanctioned or not, without adequate inventory and ownership. Shadow AI is the subset that operates outside IT’s visibility or approval, making it harder to discover and manage.

What share of agents should be governed?

The source recommends aiming to govern all agents in production, with no unknown or unowned agents. Even a small number can retain access to live data, so a low count doesn’t remove the need for ownership and review.

What should Microsoft 365 teams track beyond Copilot licences?

Track custom agents built with Copilot Studio and other Microsoft 365 tools, along with their usage, owners and permissions. Seat costs alone won’t show the usage-based costs or data access associated with custom agents.

How often should teams review their inventory?

The source recommends at least a quarterly full inventory review for many organizations scaling beyond a few dozen agents, with automated discovery in between. Teams with faster growth may need a more frequent review until the environment stabilizes.

What happens if an agent was never backed up?

Without a suitable recovery path, teams may need to rebuild an agent’s configuration after an unwanted change and assess how to restore affected data. The available recovery options depend on what the organization’s tools protect.

As agent creation spreads across business teams and platforms, the first practical step is a current inventory that links each agent to its owner, data access and spend. Regular reviews can then surface duplicate work, stale permissions and agents that no longer justify their cost.

Filed alongside