Your engineering teams shipped fifteen AI agents last year. By 2028, that number hits 150,000, according to Gartner. Most IT leaders will not see it coming as one big decision. It arrives as a thousand small ones: a support agent here, a data-pull script there, an “experiment” that quietly becomes production infrastructure. The problem is not that agents multiply. The problem is that governance does not multiply with them.
Only 13% of organizations believe they have the right governance in place to manage agents at this scale, per the same Gartner research. This piece breaks down what the numbers actually show, why the gap is dangerous, and what IT leadership can do about it before the next audit finds out first.
The gap by the numbers
Start with adoption speed. Gartner projects the average global Fortune 500 company will run over 150,000 AI agents by 2028, up from fewer than 15 in 2025 (Gartner). That is not linear growth. That is a step change most org charts, budgets, and security reviews were not built for.
Governance has not kept pace. Deloitte’s 2026 State of AI in the Enterprise survey, based on 3,235 business and IT leaders across 24 countries, found close to three-quarters of companies plan to deploy agentic AI within two years. Only 21% report a mature governance model already in place (Deloitte). Adoption is racing ahead. Oversight is walking.
SAP’s research puts a finer point on it: 98% of companies have deployed agents or plan to, yet less than half have visibility into an actual inventory of what they are running (SAP News Center). You cannot govern what you cannot list. Most enterprises currently cannot list it.
Why this gap is more dangerous than it looks
An orphaned script is a cleanup task. An orphaned agent is a standing liability. Agents call tools, touch production data, and trigger business processes on their own. When one goes wrong, it does not throw a silent error in a log nobody reads. It takes action.
The identity layer is where this becomes concrete. VentureBeat Research surveyed 573 enterprise leaders in June 2026 and found 69% of companies let AI agents share credentials instead of using scoped, per-agent identities. Organizations that allow credential sharing reported security incidents at a 63.5% rate, compared with 40.9% for companies that scope identity per agent (VentureBeat Research). That is a 22-point swing tied directly to one access-control decision.
The pattern shows up in Gartner’s forward-looking security data too. By 2028, 25% of enterprise generative AI applications are projected to experience at least five minor security incidents per year, up from 9% in 2025, largely driven by identity and access sprawl as agents outrun the controls around them (Reco.ai, citing Gartner). The agents are not the risk on their own. Ungoverned access is.
Nobody is tracking what agents cost, either
Security is one blind spot. Budget is another. VentureBeat’s same research found only 44% of enterprises rigorously track AI agent compute costs and returns. The rest are scaling agent fleets on faith, discovering the bill after it arrives rather than metering it as it accrues (VentureBeat Research).
Cost sprawl compounds quietly. A single agent running an inefficient loop is a rounding error. A hundred of them, unmonitored, across a dozen teams, is a line item finance eventually asks about — usually after the fact, not before.
A written policy is not governance
Plenty of organizations already have an “AI policy” document. Fewer have the operational controls that make the policy real. Deloitte’s 21% “mature governance” figure sits next to SAP’s finding that fewer than half of companies can even produce a full agent inventory. A policy without an inventory, scoped access, and monitoring behind it is a PDF, not a control.
“As CIOs and IT leaders see an explosion of AI agents across their organizations, many are contending with an ungoverned sprawl of agents that expose their organizations to a range of risks.”
That line comes from Gartner analysts, cited in SAP News Center’s coverage of the same research. It is worth sitting with, because it describes a condition most IT organizations are already in, not one they are heading toward.
Four moves to close the gap
Gartner’s six-step framework boils down to four things an IT leader can start this quarter (Gartner):
- Build a real agent inventory. Every agent gets an owner, a stated purpose, and an entry in a central registry before it touches production.
- Scope identity per agent. Stop sharing credentials across agents. The 22-point incident-rate gap from VentureBeat’s research is the argument for this on its own.
- Manage the full lifecycle. Agents get retired or replaced like any other production system, not left running because nobody remembers who owns them.
- Meter cost per agent. Real-time cost tracking, not the after-the-fact bill review that most enterprises are currently stuck with — only 44% rigorously track agent compute costs and returns today.
None of these require slowing down agent adoption. They require treating agents as production systems from day one, which most of them already are in practice.
Start before the audit does
Fifteen agents can hide in plain sight. A hundred and fifty thousand cannot. The organizations that get ahead of this will not be the ones that deploy the fewest agents. They will be the ones that can answer, at any moment, who owns each agent, what it can access, and what it costs. That answer takes infrastructure most companies have not built yet. Building it now costs less than rebuilding it under incident response pressure later.