AI is now embedded across enterprise technology, resetting what organizations need from observability. Dynatrace State of Observability 2026 research shows where complexity, AI behavior, cost, autonomy, security, and scale create new requirements. The challenge now is building an AI observability strategy that connects development and production with the wider environment to improve reliability, manage risk and cost, and support greater automation.
Dynatrace State of Observability 2026 research shows that AI has widened the observability mandate. Closing the execution gaps requires visibility in two directions: across the technology environment and across the AI lifecycle, from development and evaluation through production.
That visibility requires both observability of AI to evaluate AI behavior and AI-powered observability to interpret complexity and improve operations. Six connected actions turn that model into an AI operating strategy.
Six actions for building reliable AI
| Action | Why it matters | What to do |
| Define the outcomes AI must deliver | Observability improvements span every measured business outcome. | Connect AI availability, behavior, cost, and outcomes. |
| Create shared operational context | AI compounds complexity across already fragmented environments. | Connect AI, cloud, applications, security, dependencies, and business services. |
| Instrument and evaluate continuously | AI observability execution trails perceived importance by 14 percentage points on average. | Instrument early and evaluate AI from development through production. |
| Increase AI authority progressively | Integration, skills, validation, data quality, and oversight constrain autonomy. | Expand authority as reliability, explainability, and governance are demonstrated. |
| Apply continuous evidence to security | 45% continuously use runtime exploitability to prioritize vulnerabilities. | Connect pre-production testing with actual runtime exposure. |
| Manage telemetry as a lifecycle | OpenTelemetry beyond evaluation; 20% have made it strategic. | Standardize, filter, enrich, route, optimize, and contextualize telemetry. |
How do you know whether AI is working and worth scaling?
Start by defining the outcomes AI needs to deliver, then connect that outcome with the technology environment responsible for producing it. Reliable AI requires context across cloud infrastructure, applications, models, security, resource consumption, and business services.
That reliability matters in already fragmented environments: 62% operate across at least three cloud providers, and 56% use six or more observability tools. AI success criteria expand what teams need to measure:
- Define what AI should change. Identify the reliability, customer, productivity, resilience, or business outcome you expect.
- Measure whether AI is working as intended. Is it up? Is it working? Is it worth it? Add quality, behavior, reliability, and cost to conventional service health.
- Use that evidence to guide investment and authority. Expand AI where results justify additional resources or responsibility.
AI-powered observability can accelerate that work through causal analysis and automation, reducing manual correlation while clarifying causes, consequences, and cost across AI and conventional workloads.
How do you instrument and evaluate AI accuracy and quality from development through production?
Shared context connects AI horizontally across the technology environment. Lifecycle observability carries that visibility vertically from development through evaluation into production. AI observability therefore needs to start while AI is being built.
Half of respondents identify real-time AI output quality and accuracy as increasingly important, while 38% currently address it. Instrumenting observability from inception provides a consistent basis for tracing AI workloads across the operational lifecycle.
- Instrument early. Capture prompts, model versions, agent handoffs, tool calls, tokens, and evaluation scores in traces using OpenTelemetry and OpenInference.
- Set quality and cost targets. Track quality, relevance, faithfulness, hallucination, latency, and cost. Set monitor thresholds so teams are alerted when a metric crosses one.
- Evaluate continuously and connect results to action. Run online evaluations on production traces and offline evaluations on datasets before release. Define how teams respond out-of-range results. Turn failures into dataset examples to compare against baselines.
- Connect AI and operational context. Relate AI behavior and evaluation results to the applications, services, infrastructure, and dependencies that surround them to determine if the failure came from an agent or supporting system.
The goal is continuity through the AI delivery lifecycle: build, evaluate, release, and operate. Teams need to know what AI did, whether the result met expectations, what it depended on, and what changed in production.
How should AI authority expand?
Building autonomy on fragmented operational context compounds risk. AI authority should expand progressively as observability demonstrates reliability, explainability, integration, governance, and appropriate oversight. Observability should show what AI did, what informed the decision, and what happened as a result.
Integration complexity leads reported barriers to autonomous operations at 31%, underscoring why AI authority depends on the knowledge of the environment that surrounds it.
What should AI be allowed to do? Build authority progressively: start with trusted observability data, recommend, act with approval and within guardrails, then advance to broader autonomous authority for appropriate workloads only when conditions support it.
- Integrate the operating environment. Standardize interfaces and ground decisions in trusted data.
- Validate AI actions with shared context. Oversee AI actions to validate decisions and expose exceptions.
- Expand autonomy incrementally. Give AI more responsibility as observability demonstrates that systems are operating reliability within the required controls.
How do you prioritize actual security risk?
The same continuous-evidence model for AI quality applies to security. Forty-five percent continuously monitor runtime environments for exploitability or active targeting. Teams should connect pre-production testing with runtime context to move from potential weakness to actual exposure and remediation priority.
Real-time awareness and response starts with observability that spans from pre-production to runtime.
- Test before production. Use AI debugging and red-teaming to expose weaknesses early.
- Maintain continuous runtime context. Detect changing behavior and risk after deployment.
- Prioritize actual exposure. Focus remediation on exploitability, reachability, and business impact.
- Converge security and observability. Give teams shared evidence for detection and response.
How do you standardize and manage telemetry at AI scale?
The first five actions depend on context remaining usable as AI systems, providers, and environments change. Open standards help preserve that continuity from end to end. Use open standards to make telemetry portable, then manage it strategically throughout its lifecycle.
77% of organizations have moved OpenTelemetry beyond evaluation, but only 20% have made it a strategic standard. Open standards provide the instrumentation foundation that makes telemetry portable and scalable across environments and technologies.
- Standardize instrumentation and policies at the source. Apply consistent approaches across cloud, conventional, and AI environments.
- Manage telemetry before ingest. Filter, enrich, route, and optimize signals to improve quality and control cost.
- Maintain lifecycle continuity. Connect AI development and evaluation signals with production behavior.
- Add context and analytics. Connect telemetry with dependencies, AI behavior, and business outcomes.
Open instrumentation makes telemetry transportable. Telemetry management improves signal quality and control. Contextual observability turns those signals into intelligence teams can use.
Build observability for the full AI lifecycle
Reliable AI requires visibility from development and evaluation through production:
- AI engineering teams need to trace and evaluate agent behavior, understand where AI goes wrong, and verify improvements before and after release.
- Operations teams need to connect that behavior with the applications, services, infrastructure, costs, and dependencies around it and the business outcomes it must deliver.
Arize expands Dynatrace’s ability to address AI engineering requirements for deep evaluation, investigation, and workflow improvement across the agent lifecycle. Dynatrace provides the wider context and production-level awareness operations teams need to understand reliability, latency, dependencies, cost, and business impact.
Open instrumentation provides portability. Telemetry management creates usable signals. Observability of AI evaluates behavior and quality. AI-powered observability applies causal and agentic intelligence to turn that context into action. Together, those capabilities give teams an operating model that connects AI with the systems around it.
FAQ
What is an AI-optimized observability strategy?
An AI-optimized observability strategy combines observability of AI to evaluate behavior with AI-powered observability to improve operations using causal analysis, automation, and agents across development and production. Together, they connect agent behavior to applications, infrastructure, security, cost, and outcomes.
When should teams instrument AI for observability?
Start during development. Capture prompts, model versions, agent handoffs, tool calls, evaluation results, and other AI-specific signals before systems reach production.
Why is continuous AI evaluation necessary?
Pre-production evaluation cannot capture every behavior that emerges under real production traffic. Continuous evaluation helps teams detect quality changes, drift, and other deviations as operating conditions change.
How does observability support autonomous operations?
Observability supplies trusted data, operating context, auditability, and evidence of AI actions and consequences, helping teams determine where greater decision authority is appropriate.
Why are open standards important for AI observability?
Open standards such as OpenTelemetry make telemetry more portable across technologies and lifecycle stages. Telemetry lifecycle management and contextual observability then make those signals usable for analysis and action.
Looking for answers?
Start a new discussion or ask for help in our Q&A forum.
Go to forum