On July 10, 2026, Gartner published its "Hype Cycle for AI in IT Operations, 2026." According to the forecast reported by The Register, 40% of infrastructure and operations (I&O) organizations using agentic AI in production at scale will experience a business-impacting service disruption by 2028—a sharp jump from under 1% in 2026. Contrary to hopes that AI would unify operations across tools, in the near term the number of monitoring, orchestration, and permission-control products will actually increase, and that complexity will combine with the speed of automation.
Reducing this forecast to "AI is dangerous, so go back to manual operations" strips it of practical meaning. Traditional AIOps has centered on aggregating alerts, detecting anomalies, and suggesting probable causes. The emerging model of agentic I&O has AI plan actions, select tools, and execute configuration changes and recovery operations itself. What's being crossed here isn't a performance boundary—it's the boundary between suggestion and execution.
The 40% Refers to Organizations Running AI Operations in Production at Scale
The denominator behind the 40% figure isn't every company, or even every I&O department. Gartner's forecast specifically targets organizations running agentic I&O in production at scale. Also, "business-impacting service disruption" is broader than total outage—the figure doesn't mean 40% of failures will be caused by AI, nor that failure counts will rise by 40%. This isn't a measured value; it's a strategic planning assumption looking toward 2028.
Still, the warning matters because adoption and autonomy are expected to advance together. The same report forecasts that companies deploying agentic AI in IT infrastructure operations will grow from under 10% today to 60% by 2029. Furthermore, the share of AI-proposed actions that go through human approval is expected to drop from 80% in 2025 to just 20% by 2029. As the ratio of human-reviewed proposals falls from roughly 4-in-1 to 1-in-1, the number of agents operating in production will keep climbing.
In a separate survey Gartner conducted in July 2025 with over 700 CIOs, respondents predicted that by 2030, 75% of IT work would be handled by humans augmented by AI, while 25% would be handled by AI alone. This Hype Cycle also anticipates that by the same year, half of I&O teams will be restructured around investment in agents handling complex management tasks. This isn't a linear prediction of operators disappearing—it's a forecast that operators' work will shift from individual operations toward permission design, exception handling, and agent oversight.
Consoles Multiply Before They Consolidate, and Context Gets Fragmented
What AI operations tools promise is an integrated experience: querying multiple systems in natural language, correlating logs and metrics, and proceeding through to recovery. But Gartner expects the opposite trend for at least the next two years or so—requiring "more layers, more control points, and specialized observability, orchestration, and management capabilities." The sequence is that tool proliferation will only decrease later, once market and vendor consolidation progresses.
The console sprawl described here runs deeper than the problem of accumulating browser tabs. When cloud, network, and endpoint systems—through to IT service management and FinOps—each have their own separate agents, identity and permissions, policies, and audit logs get generated separately as well. During incident response, teams must reconstruct a timeline of which agent observed what, what information informed its decisions, and which credentials were used to make which changes.
| Operational Dimension | Existing Management Scope | Scope Expanded by Agent Adoption | Symptoms When Fragmented |
|---|---|---|---|
| Observability | Logs, metrics, traces | Prompts, model responses, tool calls, memory | Decision paths can't be reproduced; candidate causes conflict |
| Governance | IAM, change approval, configuration policy | Non-human identities, per-agent permissions, guardrails | Authorized tools get used for unintended purposes |
| Execution | Runbooks, IaC, job management | Natural language plans, delegation across multiple agents | Failures cascade into other domains; stop points become unclear |
| Cost | Cloud resources, licensing | Tokens, models, retries, external APIs | Recovery loops inflate spend and load simultaneously |
Existing observability environments haven't even finished consolidating yet. In a 2026 survey by Grafana Labs, gathering responses from 1,363 people across 76 countries, 38% named complexity and operational burden as their top concern. Meanwhile, 77% said centralizing their observability infrastructure had saved them time or money. While the survey included responses sourced through the company's own community and can't be taken as representative of the entire market, it does show that consolidating context has value even before AI enters the picture.
The agents themselves are proliferating too. Gartner forecasts that the average number of agents in use per Fortune 500 company will balloon from fewer than 15 in 2025 to over 150,000 by 2028, while only 13% of organizations believe they have adequate governance in place. Even as they wait for management tools to consolidate, the number of entities requiring management is expanding by orders of magnitude.
What Amplifies Failure Is the Product of Error Rate, Permissions, and Speed
When considering how AI destabilizes operations, looking only at model accuracy isn't enough. Actual damage is determined by the rate at which erroneous actions occur and how many times they execute, and this gets amplified by the scope of accessible credentials and detection/recovery time. Reduced human approval means faster execution. When an agent with broad permissions spans multiple domains, a single error's blast radius expands accordingly.
While DORA's 2025 survey wasn't a study that directly measured I&O agents, it examined roughly 5,000 technical professionals and reported that AI usage correlates positively with software throughput and product outcomes, while correlating negatively with delivery stability. In the 2024 survey, each 25% increase in AI usage was associated with a 7.5% improvement in documentation quality, 3.4% in code quality, and 3.1% in review speed—but also a 7.2% decline in stability. By 2025, even as throughput turned toward improvement, the stability weakness persisted. DORA explains that when AI increases the volume of changes, downstream processes with weak automated testing, immature version control, or slow feedback become correspondingly more unstable.
This mechanism has been documented in pre-AI automation as well. Google's SRE team has published a case where, when a set of servers slated for decommissioning turned out to be empty, the processing system interpreted the empty set as "all targets" and initiated disk wiping at colocation facilities around the world. The bug itself was small, but because the legitimate automation held broad permissions and machine speed, the blast radius expanded instantly. Google describes automation as a "force multiplier"—one that becomes harmful without clearly defined scope.
Agentic I&O adds probabilistic planning and dynamic tool selection to this old problem. OWASP's "Top 10 for Agentic Applications 2026" counts misuse of legitimate tools and abuse of identity/permissions among the primary risks, alongside compromise of the agent supply chain and unexpected code execution. Deterministic guardrails are effective at restricting authorized operations to a fixed set of rules. But if the policy itself is flawed, permissions are too broad, or stop conditions are missing, the same mechanism becomes a device that consistently repeats the wrong operation.
Fix the Boundary Between Observation and Execution Before Waiting for Consolidation
Consolidating management tools under a single vendor isn't necessarily a solution. When an integrated product places observation, judgment, and execution within the same failure domain, losing that control plane means losing the fallback path as well. What's needed first is a boundary that remains fixed regardless of which agents come and go. Configuration should have its source of truth in Git or a CMDB, identity in IAM, and change history in ITSM—with a design that never lets an AI conversation log become the new source of truth.
Execution permissions are easier to reason about when divided by an operation's reversibility and blast radius. Reads and diagnostics can be allowed to run autonomously. Low-risk, reversible changes can be conditioned on canary deployment and automatic rollback. Configuration changes affecting all of production, data deletion, and credential updates should still require human approval. Irreversible operations should be excluded from an agent's credentials entirely. Following Google SRE's recommendation, changes should roll out in small partial deployments, be evaluated against SLOs, and be immediately reversible. This principle serves as a safety mechanism for AI operations as well.
Unifying observability data is something that can begin before console consolidation. OpenTelemetry is developing GenAI-specific semantic conventions that record model names, token consumption, and call latency in a common format—tracking each tool call and its result as well. However, common specifications across agent frameworks are still under development. Since prompts and tool arguments can contain sensitive information, unless collection scope, retention periods, masking methods, and access permissions are decided together, the observability infrastructure itself risks becoming a new information leakage path.
Gartner itself, in its June 2026 Agentic AI IntentOps Framework, states that without machine-readable operational intent, AI agents can't reliably safeguard business outcomes, and urges organizations to define explicit specifications before expanding beyond troubleshooting. When selecting products, check whether credentials can be scoped down per operation, and whether dry-run capability exists. It's also essential to confirm that repeating the same operation doesn't corrupt state, and that audit logs can be exported externally. Stop and recovery mechanisms need to be executable through a separate path that remains usable even if the product itself goes down.
Avoiding the 2028 40% forecast means not treating AI automation rate as the success metric. Measure the rollback rate of AI actions and SLO violations, and log permission deviations as well. The time required to reconstruct a decision path during an incident needs a baseline established before deployment. As agents proliferate, the goal is to keep the blast radius of a single error small enough that humans can regain control. Only in organizations that meet these conditions does the growth in management tooling become a transitional phase heading toward future consolidation.