Ninety-five percent of organizations have faced at least one AI agent-related incident in the past 12 months. Seventy percent of the infrastructure and operations leaders responsible for those agents believe the agents are operating within safe boundaries. Both numbers come from the same Gartner survey of 298 I&O leaders in the U.S. and U.K., released on November 16, 2026, and the distance between them is the most important finding in the release.
Gartner’s headline prediction is blunt: without strong governance, 40% of I&O organizations scaling AI will suffer a business-critical service disruption by 2028. The survey behind it explains why. Fifty-seven percent of respondents said their current governance is weak, 55% said their existing IT risk tools are completely outdated, and only 15% have automatic circuit breakers in place to stop runaway agent errors or infinite logic loops. Agents are already in production, already causing incidents, and mostly running without the AI agent guardrails that would stop a bad run before it becomes an outage.
Confidence is not a control
The 70% figure deserves a second look. It does not mean seven in ten leaders are careless. It means most leaders are judging agent behavior by the absence of visible damage, which is how you would judge a system you cannot fully observe. If an agent quietly loops through a task three times, pulls data it did not need or burns through a week of token budget in an afternoon, nothing in a traditional IT risk tool is designed to flag it. From the outside, that agent looks like it is operating safely, right up until it is not.
Nathan Hill, Distinguished VP Analyst and Chief of Research at Gartner, called it “a massive operational readiness gap.” His prescription starts with the basics: I&O leaders “must immediately inventory active AI agents and classify them all as nonhuman identities in their identity platforms.” That instruction assumes something many organizations cannot yet do, which is produce a complete list of the agents running in their environment. Agents now arrive through vendor platforms, low-code workflow tools and business teams that never filed a ticket. We covered the same pattern in our look at how far behind even the largest banks are, where only 8% of the institutions BCG benchmarked could fully monitor their AI agents.
What Gartner says AI agent guardrails should include
Gartner recommends that I&O leaders work with the CIO or AI leader on a governance framework built around dynamic runtime controls, and it names seven pillars. They are worth reading as a set, because together they describe an operating capability rather than a policy document.
- Identity: treat every agent as a machine identity tracked in a centralized registry, so nothing runs in the shadows.
- Guardrails: build automated safeguards into runtime that can slow a misbehaving agent before it disrupts operations.
- Observability: track the agent’s actual reasoning chain to understand why it did what it did.
- Agentic FinOps: route intelligently between private and public frontier models and set session-level spending caps to keep token costs under control.
- Escalation: set boundaries by risk level, with human sign-off still required for high-impact actions.
- Containment: use gating and identity kill switches to find and isolate runaway agents before failures cascade.
- Data perimeters: create safety perimeters around data to prevent context poisoning and leaks.
Notice how many of these pillars are measurement problems wearing a governance label. A registry is an inventory. Observability is a record of what happened. Agentic FinOps is cost attribution with a budget attached. Containment depends on detecting that an agent is running away in the first place, which you can only do if you know what normal looks like for that agent. Guardrails are only as good as the data that tells them when to fire. This is the gap Olakai was built for: a vendor-neutral record of what every model, agent and AI tool is doing, what it costs and what it returns, so the controls Gartner describes have something reliable to act on.
The cost of a runaway agent shows up in two places
Gartner frames the risk as service disruption, and for I&O teams that is the right frame. But a runaway agent shows up on the finance side too. An infinite logic loop is not only an operational failure; it is an uncapped token bill. The fact that Gartner puts session-level spending caps inside its governance framework is a signal that the line between AI governance and AI cost management is disappearing. We saw the same convergence when Apple cut some engineers’ daily AI allowance to $200, which we covered in our analysis of Ramp’s AI spend data, and when stolen AI keys ran up model charges on someone else’s bill.
Forrester’s latest data shows why the two sides need to be joined. In its State of AI Survey, 2026, 82% of AI decision-makers report at least one generative AI application in production and 77% at least one agentic AI application, yet only 20% believe their organization has the technology in place to scale AI initiatives. Even among high-readiness organizations, only 42% report that they can link their AI initiatives to P&L. Forrester’s recommendation lines up closely with Gartner’s: establish controls for agent access, keep the ability to trace what agents do and intervene when necessary, and document where each initiative delivers value and where it does not yet.
Put the two studies side by side and a single requirement emerges. The same data that tells an I&O leader an agent has crossed a boundary should tell a CFO what that agent cost and what it produced. When those answers live in separate tools, owned by separate teams, neither question gets answered well, and leaders fall back on the kind of confidence the 70% figure reflects.
Why switching agents off is getting harder
One line in the Gartner release should concern every operating leader: “As dependence on agents grows, I&O leaders have less flexibility to switch them off.” Hill’s advice is that a governance framework must identify which agents are critical to keep running, even if they are expensive or difficult to maintain. That decision cannot be made on intuition. Knowing which agents are critical requires knowing which ones carry real business volume, which processes depend on them and what value they deliver against what they cost. In other words, it requires AI ROI measurement at the level of the individual agent, not a portfolio estimate in a quarterly deck.
This is also where the governance conversation stops being purely defensive. An organization that can see every agent’s activity, cost and outcome can make the kill-switch decision quickly and with evidence. It can also make the opposite decision, expanding the agents that are earning their place, with the same confidence. Olakai’s AI governance and analytics and custom KPIs run on one dataset for exactly this reason: risk and value are two readings of the same record.
What I&O, finance and AI leaders should do now
Start with the inventory Gartner asks for, and make it continuous rather than a one-time census. Every agent should have an identity, a named owner and a business purpose, including agents running inside vendor platforms and workflow tools. Agents adopted outside formal channels belong in the same inventory, which is why shadow AI detection and agent monitoring should sit in one program rather than two.
Next, instrument agents before you add more of them. Record what each agent does in production, including actions, data access, model calls, policy exceptions and escalations, and baseline what normal behavior and normal spend look like for each one. Circuit breakers and spending caps only work when they are tuned against that baseline. Olakai provides that layer as a system of record for AI, giving I&O the observability to set guardrails, finance the cost attribution to set caps and AI leaders the outcome data to decide which agents to scale.
Finally, review risk and value together. A monthly agent review that covers incidents, cost per completed task and outcome against the business case will surface the agents that need tighter guardrails and the ones that deserve more investment. That review is what turns Gartner’s seven pillars from a framework into a habit, and it is how the 15% with circuit breakers today will become the organizations that avoid the 2028 disruption Gartner is forecasting.
Where to start
Pick the five agents your business depends on most and answer four questions for each: who owns it, what it did last week, what it cost and what it returned. If any of those answers requires a manual data call to several teams, your AI agent guardrails are running on assumptions, and you are likely closer to the 95% than the 70% would suggest. Olakai’s complete AI monitoring gives you those answers for every agent, model and AI tool from one platform, so the confidence your teams have in their agents is backed by evidence.
Talk to an expert about putting measurable guardrails around every AI agent in your enterprise.
