Agentic AI edges closer to everyday production use
Ops teams have moved past "should we" and are now focused on "how do we run this safely at scale." A new report shows agentic AI already embedded in IT operations, cybersecurity, data processing, and customer support. 70% of respondents use AI agents in IT operations and system monitoring, with nearly half running them across both internal and external use cases.
Budgets match the momentum. Most organizations expect spending to climb this year, and many are already investing $2-$5 million annually. Funding is gravitating to use cases tied to reliability and operational performance.
From pilots to limited production
Adoption is uneven, but progress is obvious. Half of organizations have agentic AI in production for limited use cases; another 44% report broad adoption within select departments. Most teams run between two and ten active projects.
IT operations, cybersecurity, and data processing lead in production readiness. About half of projects in these areas are live or being operationalized. The gate to production is technical: security and data privacy first, then accuracy and reliability of outputs. Many teams treat observability as a prerequisite for wider rollout.
Observability gaps slow progress
Security, privacy, and compliance remain common blockers. A similar number cite difficulty managing and monitoring agents at scale. Limited visibility into agent behavior-and trouble tracing downstream effects of autonomous actions-shows up across regions and industries.
As systems connect more tools, models, and data sources, teams need real-time insight into decisions and execution paths. Without that, it's hard to diagnose odd behavior or link technical signals to business outcomes. Observability is becoming the control layer that holds everything together.
Nearly 70% already use observability tools during implementation, and more than half rely on them during development and operations. Common patterns include monitoring training data quality, real-time anomaly detection, output validation, and compliance checks.
Humans remain part of the loop
Autonomy is rising, but oversight is still the norm. More than two thirds of agentic AI decisions are verified by a person. Data quality checks, human review of outputs, and drift monitoring are standard.
Fully autonomous agents are rare. Most organizations mix autonomous and supervised agents based on task and risk. Business-facing applications carry higher levels of human involvement than infrastructure-focused use cases.
Measuring success through reliability
Reliability and resilience lead the scorecard. 60% say technical performance is the top success metric, followed by operational efficiency, developer productivity, and customer satisfaction.
Monitoring is still a blend of old and new. About half rely on logs, metrics, and traces; nearly half still review agent-to-agent communication manually. Automated anomaly detection and dashboards are common, but many teams combine automation with manual review. The outcome that matters: systems that hold up under stress and recover quickly from faults. Early detection and rapid response are non-negotiable.
Scaling with tighter controls
The next phase is governance and control. Teams want shared factual signals, standardized metrics, and consistent guardrails to guide autonomous actions. Observability ties these elements together across the lifecycle.
"Organizations are not slowing adoption because they question the value of AI, but because scaling autonomous systems safely requires confidence that those systems will behave reliably and as intended in real-world conditions," said Alois Reitbauer, Chief Technology Strategist at Dynatrace.
Agentic AI expands the operational attack surface and increases dependence on monitoring, validation, and oversight. As more projects hit production, trust becomes an operational requirement-delivered by tooling, process, and human judgment working in concert.
What operations leaders should do next
- Prioritize high-impact, lower-risk use cases first (IT ops, system monitoring, incident response). Tie budget to reliability and performance outcomes.
- Define decision boundaries and approval thresholds. Map human-in-the-loop to risk levels and data sensitivity.
- Instrument agents end to end: logs, metrics, traces, plus prompts, tool calls, model versions, and I/O tags. Consider OpenTelemetry for consistent telemetry.
- Require causal tracing for autonomous actions. Create event timelines and IDs that follow decisions across agents, tools, and services.
- Set SLOs for agentic systems (accuracy, latency, change failure rate, MTTR) and wire alerts into on-call.
- Enforce least privilege, secrets management, and data minimization. Add audit trails and run privacy/compliance reviews early.
- Build kill-switches, safe modes, and rollback playbooks. Use shadow tests and canaries before full rollout.
- Validate data pipelines continuously: training data quality, PII scanning, feedback loops, and drift monitoring.
- Automate anomaly detection but keep spot audits for agent-to-agent communication.
- Adopt a shared risk framework for AI. If you need a starting point, review the NIST AI Risk Management Framework.
Level up team capability
Strong platforms are necessary, but skills close the gap. If you're building an AI operations playbook and training path, see our role-based options here: Complete AI Training - Courses by Job.
Your membership also unlocks:
