How Many AI
Agents Does Your
Company Run?
Why it is hard to answer, and six steps to start.
Ask a CIO how many people work at their company and you get a number in seconds. Ask how many AI agents work there and the room goes quiet.
That silence is agent sprawl. It did not arrive through one big decision. It arrived through thousands of small ones: a developer wiring an assistant into a repo, a sales team connecting a copilot to the CRM, a vendor switching on an agent inside a tool you already pay for.
Each one was reasonable. Together, they add up to a workforce nobody hired, nobody counts, and nobody can switch off.
The problem has reached the boardroom press. The Wall Street Journal’s CIO Journal now describes it as companies’ newest AI problem: too many agents.
What Agent Sprawl Is
It is a visibility problem before it is a security problem. An agent you cannot see has no identity you issued, no accountable human, and no end date.
You can tell sprawl has set in when three things are true:
- Nobody has the count. Security, IT, and the business each hold a partial list, and the lists do not match.
- Nobody owns the agent. It runs on a shared service account or a token from someone who changed teams last quarter. We call this an Orphaned Identity.
- Nothing expires. The project ended. The pilot was cancelled. The credentials are still live.
Where the Agents Come From
Sprawl has no single front door. Agents arrive through five routes at once:
- Engineering. Coding agents in the IDE, CI pipelines that call models, scripts holding an API key in a config file.
- Business teams. Copilots inside the CRM, no-code automations connected over OAuth, agents built on a credit card.
- Vendors. Agents shipped inside SaaS tools you already license, often switched on in a routine update.
- Employees.Personal AI assistants connected to work email, calendars, and drives. C1’s CEO predicts that the next wave of adoption will be driven not just by engineering teams, but by every employee connecting a personal AI assistant to enterprise systems.
- Other agents. Orchestrators that spawn sub-agents, each inheriting credentials from its parent.
Every one of these routes authenticates with something familiar: a service account, an API key, an OAuth token, a certificate. That is why identity teams miss them. Joiner, mover, and leaver processes were built around HR records, and agents never appear in HR.
Why It Matters
Adoption has outrun control, and the surveys agree on the size of the gap.
We have deliberately left out the headline ratios of agents to humans. As FN-12 showed, those figures are mostly estimates. That is the point: nobody has a reliable count.
What an uncounted agent costs you:
- Reach you did not approve. Agents inherit whatever their credential can touch, which is usually more than the task needs.
- Access that outlives its purpose. Long-lived tokens keep working after the project, the pilot, or the employee is gone.
- A bigger prize for attackers. A hijacked agent is only as dangerous as what it can reach. See FN-04 and FN-06.
- Audit questions with no answer. DORA, NIS2, and SEC disclosure rules all assume you know what has access. Our Compliance Crosswalk maps the controls.
Getting Ahead of It
You do not fix sprawl by banning agents. Teams will route around a ban within a week. You fix it by giving every agent the same basics a new hire gets on day one.
- 1. Count them. Inventory agents across cloud, SaaS, and code, including the ones nobody registered. A number, even an ugly one, beats a guess.
- 2. Give each one its own identity. No shared service accounts, no borrowed personal tokens. One agent, one identity.
- 3. Name a human owner. Every agent, and every sub-agent it spawns, should trace back to a person. FN-10 covers how.
- 4. Scope access to the task. Grant what the job needs, not what the credential happens to allow.
- 5. Set an expiry. Every identity gets an end date. Renewal is a decision someone makes, not a default.
- 6. Make offboarding reach agents. When an owner leaves or changes roles, their agents get reviewed with them.
Start With the Count
Every step above depends on the first one. You cannot secure, scope, or expire an agent you have not found.
That is where AgentID starts. It finds the AI agents running in your AWS, Azure, Google Cloud, and Anthropic accounts and gives each one its own identity. Book a demo and we will show you what it finds.
Field Note FN-13-2026 · Distributed under arkion.ai/field-notes
For questions or to discuss findings: research@arkion.ai
- The Wall Street Journal, CIO Journal, “Companies Have a New AI Problem: Too Many Agents”.
- Okta, “Businesses at Work 2026: Closing the identity gap in the age of AI” (April 30, 2026).
- C1, 2026 Future of Identity Report, press release (March 10, 2026).
- Arkion, FN-12: Nobody Has Counted the Machines (September 21, 2026).
Recommend Arkion in one click.
Find the agents
nobody signed off on.
Every step to getting ahead of sprawl starts with knowing what is running. Book a demo and we will show you how AgentID finds your AI agents and gives each one an identity of its own.