Research report 2026
The pulse ofagentic AI
Building trust in autonomous operations from pilot to production
©2026 DynatraceIntroduction
Agentic AI for autonomous operations in 2026
Agentic AI is leaving the lab. The report tracks how far projects have got, where they stall, and why observability has to become the control layer as they scale.
Autonomous operations are moving from experimentation into production as enables systems to reason, learn, and execute complex tasks with minimal human intervention.
The goal of autonomous systems is to boost resilience, scalability, and customer experiences. But enterprises are discovering that autonomy without control can create more risk and reliability concern than value.
In this inaugural report on the pulse of agentic AI, we focus on project maturity, challenges, and scale to clarify:
- How organizations are adopting, prioritizing, and deploying agentic AI
- Where teams are making progress and hitting barriers
- When and why organizations keep humans in the loop
- How organizations measure AI success
- Why observability is becoming a critical control plane for operationalization
, the real-time understanding of system behavior through connected data, has always played a role in system performance. But with the variability of generative models, observability must shift to a foundational control layer as organizations scale projects from pilot to production.
Why it matters
With a clear understanding of how organizations are advancing agentic AI deployments, the challenges they face, and where they're finding value, this report provides a data-driven resource for leaders looking to make strategic decisions, enabling them to advance and scale trustworthy autonomous operations using real-time observability.
Before chapter one
Choose your track
Same findings, two depths. You can switch any time from the bar at the top; the page keeps your place.
Chapter one
Agentic AI is rapidly expanding beyond IT operations
IT operations is where agents prove themselves (70% use them there). Customer-facing uses are the fastest-growing, and three-quarters expect budgets to rise.
Agentic AI has gained the most traction in mission-critical proving grounds such as IT operations, software engineering, cybersecurity, and support—applications that can least tolerate failure. This importance emphasizes the need for foundational control, governance, and comprehensive observability of agentic systems.
An operational best practice is to prioritize the most measurable, repeatable use cases with the greatest expected ROI. The findings show that organizations are prioritizing foundational internal functions like IT operations, data processing and reporting, and cybersecurity.
Repeatable customer-related functions, like services and personalized engagement, rank as the top external applications. Nearly half (49%) use agentic AI for both, although 33% currently emphasize internal use cases.
Customer-facing applications and budgets are rapidly rising
While internal use cases currently dominate, the fastest-growing implementations over the next five years are externally exposed, demonstrating an imperative to realize agentic AI value in customer-facing applications.
AI automation for legal, which is currently the least common use case, is projected to triple in the next several years, indicating strong growth potential for non-IT use cases.
Production readiness is growing fast across all use cases
Consistent with current investment patterns and expected ROI, IT operations and system monitoring use cases show the largest share of production-level readiness. For most functions, about half of those projects are in production or being operationalized.
Agentic AI budgets remain robust
With current spending averaging $2–5M, the majority expect to increase their budgets, and nearly half expect spending to increase by another $2–5M in the coming year. This is a clear signal of sustained priority.
Respondents' top answer for where agents will pay back most. Pick yours in the widget below.
Recommendation
Control starts with goal-based automation
Agentic systems represent a quantum leap in productivity, efficiency, and scale, but also complexity and often unpredictability. Gaining visibility into agent interactions and decisions helps organizations maintain control.
- Develop clear goals, guardrails, and oversight for agentic projects to avoid disorder and maintain reliability.
- Design systems so agent-to-agent communications remain accountable to humans.
- Manage the nexus of autonomous, real-time interactions with observability.
Deep dive · editor's notePicking a first production use case
The report's best practice is "measurable, repeatable, greatest expected ROI." In practice that means a task that already has an owner, a runbook and a metric, such as alert triage or log summarisation in IT operations, where 46% are already in production.
Write the success measure before the agent: time to resolve, tickets closed without escalation, or cost per incident. If you can't name it, the project will struggle with the "lack of clear business case" barrier that 33% report in chapter two.
Chapter two
Autonomous operations is making progress—and hitting barriers
Most organizations run 2–10 agentic projects, and half have something in production. What holds them back is trust: knowing when an agent may act alone, and seeing how it decides.
Although agentic AI is still an emerging technology and adoption is relatively shallow, organizations are quickly moving toward maturity. But organizations face technical barriers to governing and monitoring agentic AI behavior, such as limited ability to set automation rules and limited visibility, which are hindering progress.
Agentic AI adoption is still growing as organizations evolve small automation projects into autonomous operations. With adoption still in early phases, nearly three-quarters (72%) report having between 2-10 active projects, and about a quarter have 11-21 or more projects.
Most of these projects are in some phase of pilot or limited production, suggesting that organizations are feeling the benefits of agentic AI and looking to expand and perfect its implementation.
Production criteria prioritize security and reliability
Security and data privacy and AI accuracy are the top criteria for moving projects from pilot to production. Ethics and fairness rank lowest in priority, indicating that for many, technical accuracy currently outweighs ethical concerns.
Criteria for moving agentic AI from pilot to production
Technical barriers to production include observability concerns
The top technical barriers to advancing projects relate to establishing trust, such as setting clear rules for when an AI agent should act autonomously versus when it requires .
Other technical challenges include limited visibility, difficulty monitoring, coordinating troubleshooting efforts, and difficulty tracing downstream effects, indicating that organizations are struggling with observability and traceability of agentic AI behavior.
Barriers to production
Recommendation
Build trust in autonomous operations incrementally
While expectations ride high for agentic AI projects, building trust in autonomous operations is best done incrementally. By implementing functionality in phases, organizations can reliably prepare the environment for fully autonomous operations.
- Start with preventive operations and recommendation-driven workflows.
- Using a human-in-the-loop model, progress toward supervised autonomy by hardening services, data sources, and contextual signals.
- Graduate to full autonomy using advanced observability data to understand each component in context and verify its performance in real time.
One incident, three phases of trust
Deep dive · editor's noteWriting the "act alone or ask" rule
The top barrier (45%) is the rule itself: when may an agent act autonomously, and when must a human approve? A workable rule usually combines three things the control plane can check at decision time: the blast radius of the action (one pod or a whole region), its reversibility (restart versus delete), and the agent's evidence (does the root cause match deterministic telemetry?).
Start every action type at "recommend." Promote it to "supervised" once its approval rate is consistently high, and to "autonomous" only when its outcomes are traceable end to end.
Chapter three
Agentic AI trust and automation: Human oversight is intentional and enduring
Over two-thirds of agent decisions are still checked by a person, and that's by design. Most teams build both autonomous and supervised agents, and expect a human-AI partnership, not a handover.
AI-powered automation is the clear direction of travel, which raises the stakes for trust, governance, and control. But organizations are intentional about human involvement across the board and foresee a future of human-AI partnership.
Agentic AI powered by generative models is fundamentally probabilistic in nature and can produce in the process. For more than two-thirds of respondents, humans validate agentic AI decisions using sophisticated systems—data quality checks, AI reviews, intelligent logs and traces—creating the primary trust mechanism that enables scale. These findings are consistent for organizations of all types, regardless of company size, region, and revenue.
A mix of autonomy and human supervision
Nearly two-thirds of respondents (64%) are currently building both fully autonomous and semi-autonomous agents supervised by humans.
Only 13% are building solely autonomous AI agents, and about a quarter (23%) are building only human-supervised agents.
Which one describes your team? Tap a card.
Validation measures vary
Lower-revenue organizations ($100-$499M) are more likely to use security validations to verify agentic AI decisions (55%), whereas higher-revenue organizations ($10B or more) are more likely to validate these decisions by recording comprehensive logs and traces (31%).
Automation goals favor humans-in-the-loop
Data processing, cybersecurity, and customer service/support functions currently show the highest rates of AI autonomy, but organizations foresee building greater autonomy for knowledge management and customer-facing digital products in the future.
In general, respondents expect a higher percentage of human-involved automation (60/40 on average) for business-related applications than IT and routine customer-support-related functions (50/50 on average).
The future is human-AI collaboration
Future automation goals indicate that organizations consider agentic AI as a crucial partner and enabler of human productivity, not a replacement.
Observability data is the crucial intelligence layer for human architects to govern and control the collaboration required to produce safe and reliable autonomous operations.
Recommendation
Redesign roles to optimize the human-AI partnership
As organizations adapt to the shifting priorities and demands of integrating agentic AI into workflows, leaders can be strategic about how AI facilitates human productivity.
- Human judgment guides the system by setting goals, defining boundaries, and ensuring accountability.
- AI performs execution for repeatable or time-sensitive tasks.
- An observability-based control plane helps manage the human/AI interface.
Deep dive · editor's noteWhat "verified by a human" can mean
The 69% covers a wide range: a person clicking "approve" before every action, a reviewer sampling yesterday's decisions, or an on-call engineer reading an agent's summary before closing a ticket. They cost very different amounts of human time.
As volumes grow, review shifts from every decision to the exceptions. That only works when logs and traces show why the agent acted, which is why the report pairs human oversight with observability throughout.
Chapter four
Reliability and resilience are key benchmarks of success
Success is measured by technical performance first. Real-time decisions are the top goal, yet 44% still review agent-to-agent traffic by hand, and that can't keep up as faults spread across connected agents.
Organizations measure agentic AI success in terms of reliability and resilience. Real-time decision-making is a top priority, but monitoring methods still rely heavily on manual reviews, which limits the ability to scale.
Agentic systems are composed of a variety of specialized AI technologies (see appendix A). Because these systems are distributed and interconnected, a small flaw can quickly propagate through applications and environments, making reliability, resilience, and trust a driving concern. As a result, leaders are prioritizing real-time decision-making, internal efficiencies, and system performance above regulatory compliance, addressing labor shortages, and reducing mundane tasks.
Priorities and key performance indicators are converging
In terms of how they measure agentic AI success and ROI, organizations are focusing less on daily use rates and sustainability than on technical performance, operational and developer efficiency, and customer satisfaction. Reducing risk is a fast-follow, suggesting that for systems to remain available and secure, reliability and security are becoming inseparable.
The larger an organization's size and revenue, the greater the likelihood it uses compliance and security benchmarks for measuring agentic AI success. Conversely, lower revenue organizations ($100M-$999M USD) focus more on increased productivity/efficiency and time-to-value.
KPIs used to measure agentic AI ROI
Monitoring methods focus on outcomes
Methods organizations use to monitor and analyze agentic AI systems are largely outcome-driven (about 60%), with a heavy reliance on logging, metrics, and traceability tools. While they're trying to automate anomaly and drift detection, 44% are still reviewing communication flows manually, indicating a need for automation inputs that are accurate, real-time, and context-aware.
Methods for monitoring/analyzing agentic AI systems
Recommendation
Measure success in resilience and reliability
Successful agentic systems must be able to absorb disruption, recover quickly, and maintain consistent performance under peak loads. Detecting problems early and rapid recovery are essential, because faults spread faster through interconnected agentic systems.
- Measure resilience by how systems respond in real time under stress.
- Treat reliability and security as equal requirements to keep systems available and secure.
- Leverage factual observability signals to uplift all KPIs and build trust for automation.
Chapter five
Observability is a control plane for autonomous operations
Teams already use observability across the lifecycle, most of all in implementation. But tools built for deterministic software miss how probabilistic agents behave. The fix is a control plane that grounds agents in facts.
Organizations are already using observability across the AI delivery lifecycle, especially during implementation. But existing tools that rely solely on are not designed for probability-based, goal-driven systems.
With conventional software, organizations can test quality during development. However, generative AI models introduce variability that can't be fully verified in pre-production. This variability requires real-time systems to monitor performance in production, identify unexpected behaviors, and implement corrective actions quickly.
Most respondents report using observability during at least one phase of the agentic AI life cycle. Significantly, more than two-thirds use observability during implementation for integrating with existing systems, detecting anomalies, and validating outputs. These indicators reflect the value of observability data at every stage, from identifying pre-deployment risks to unexpected AI outcomes in production.
Nearly all organizations use observability at some point during the agentic AI lifecycle
Traditional monitoring and analysis tools leave gaps in understanding
However, respondents also report clear gaps in their current methods for understanding how agentic AI systems behave, the risks they expose, and drawing a throughline to business outcomes. This recognition suggests that current practices lack sufficiently structured inputs to ensure trustworthy outputs over time.
Respondents describe the following types of blind spots:
"There's a need for improved capabilities to assess the ethical implications and unintended behaviors of agentic AI systems."
"The lack of a transparent, real-time view of the agent's decision-making logic makes its behavior resemble a 'black box'."
"Current tools lack real-time insights and proactive anomaly detection."
"Existing tools lack sufficient monitoring capabilities for the long-term behavior of AI agents."
"It's difficult to directly link technical monitoring metrics to business results."
Observability as a control-plane approach
Using observability data for monitoring is a good start, but to build trust in AI, observability needs to function like a control plane that can orchestrate deterministic data and context with generative models.
An intelligent observability framework needs to detect and take real-time actions to detect hallucinations, understand downstream implications, and take immediate action to limit impacts. A precise fact basis for reliable decision-making produces autonomous operations that can scale.
Recommendation
Establish observability as the control plane for autonomous operations
Because generative AI models are based on a probabilistic approach to decision making, building trust in automation requires that AI agent actions be grounded in deterministic, real-time facts.
- Ensure agents act on the same factual signals and remain accountable to the human operator.
- Establish deterministic guardrails that enable agentic systems to behave safely, act predictably, and collaborate using reliable real-world signals.
- Instrument thoroughly and standardize common semantic conventions, metrics, and trace attributes.
Deep dive · editor's noteWhy "standardize semantic conventions" matters
When every team names its spans, model calls and tool invocations differently, an agent's decision can't be followed across the stack, and the "challenges coordinating troubleshooting across multiple tools or teams" barrier (40%) appears. Shared attribute names for model, prompt version, tool, token count and outcome let one trace run from a GPU to a business KPI.
Instrument the agent layer first. It is the least monitored today: only 40% track and in operations.
Appendix A
Additional findings
Currently used AI technologies
The agentic systems that fuel autonomous operations are made up of many AI technologies that perform specialized functions. The most used types are conversational and generative AI, such as chatbots and virtual assistants, as well as predictive and analytical AI. These findings support the value organizations expect from AI technologies and the priorities to which organizations are actively applying them.
Agentic AI infrastructure
Hyperscaler platforms, such as Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure, host 82% of agentic AI infrastructure, followed by GPU-as-a-service platforms. 42% of respondents host their agentic AI infrastructure using on-premises assets.
Appendix B
Global data summary
Methodology overview
This report is based on a global survey of 919 senior leaders and decision makers directly involved in or responsible for agentic AI development and implementation in large enterprises with annual revenues of $100 million or more. It was conducted and analyzed by Qualtrics partner Y2 on behalf of Dynatrace during November and December 2025.
Industry verticals: Financial Services, Retail, Government, Utilities, Transportation & Fulfillment, Software & Digital Services
Respondent roles: Vice Presidents and directors of engineering leading AI projects; platform engineering leaders scaling AI systems; executive and IT decision-makers responsible for AI investments and risk; AI/ML and automation Leads; SRE and observability stakeholders
All findings were analyzed to uncover success patterns, competitive advantages, and enterprise strategies for scaling agentic AI through intelligent observability across global markets.
The margin of error is ±3.2% at a 95% confidence level.
Respondent countries: tap one to find it.
Signal 12 · Checkpoint
Five questions. Take the pulse.
Instant feedback, with the figure behind each answer.
Dynatrace
Dynatrace: The observability control plane for advancing agentic AI from pilot to production
We hope this report has provided valuable insights about how agentic AI projects are advancing to autonomous operations and how observability provides vital intelligence at every stage.
Dynatrace is committed to providing enterprises the data and intelligence they need to be successful with their enterprise cloud and digital transformation initiatives, no matter how complex.
About Dynatrace
Dynatrace is advancing observability for today's digital businesses, helping to transform the complexity of modern digital ecosystems into powerful business assets. By leveraging AI-powered insights, Dynatrace enables organizations to analyze, automate, and innovate faster to drive their business forward. Learn more at www.dynatrace.com.
Interactive edition of “The pulse of agentic AI · Research report 2026” (©2026 Dynatrace, 01.23.26 bae22697-ebk-cs). Report text and figures as published; gists, editor's notes, deep dives, glossary, simulations and quiz written for this edition. Edition design: studio-vektor.com