AI In Production Exposes Gaps In Enterprise Observability And Incident Resolution, Dynatrace Study Finds

Enterprises are rapidly deploying AI into production environments, but the teams responsible for keeping those systems running are finding that traditional monitoring approaches and operational tooling are not keeping pace with the demands of AI workloads.

That is the central finding of The State of SRE and Platform Engineering 2026, a global survey of 919 IT leaders published by observability platform vendor, Dynatrace. The study examines how site reliability engineering (SRE) and platform engineering teams are adapting as AI workloads move from pilot projects to production infrastructure.

The research paints a picture of disciplines that have matured considerably in recent years but are now confronting a new class of operational challenges introduced by large language models, coding assistants and agentic AI systems.

AI Model Monitoring Now The Top SRE Priority

The survey found that monitoring AI models for performance and accuracy has become the primary use case for SRE teams, with 67 per cent of respondents naming it their top priority. That figure places it ahead of traditional SRE concerns such as automation, service-level objectives and quality gates.

The shift reflects how quickly AI workloads have embedded themselves into production environments. Unlike deterministic software, AI systems behave probabilistically, can drift over time and can fail silently, producing outputs that appear valid but are increasingly inaccurate.

Some 58 per cent of SRE teams reported using AI to monitor model performance and accuracy, making it their most common AI-powered capability. Half of SRE respondents now use AI-powered capabilities for automated incident response, signalling a shift toward agentic operations.

Yet only 3 per cent of respondents indicated they are not using AI for SRE at all, suggesting near-universal adoption of AI-assisted tooling across the discipline.

Strong Executive Backing, But AI Falling Short On Key Metrics

The report found that SRE practices enjoy strong executive support, with 92 per cent of organisations reporting leadership backing for SRE initiatives and 55 per cent describing that support as “very strong.”

That institutional buy-in has translated into operational maturity. Some 64 per cent of respondents indicated that most or all of their IT operations processes are fully automated, and 89 per cent use service-level objectives across at least some teams or systems.

However, while AI technologies are generally meeting expectations around system reliability and developer productivity, they are delivering less impact than expected in two critical areas: cost reduction and mean time to resolution (MTTR).

On cost reduction, 45 per cent of respondents reported actual benefits, compared to 55 per cent who had expected improvements. For MTTR, the gap was similar, with 46 per cent seeing actual reductions against expectations of 53 per cent.

The report attributed this shortfall to limited workflow automation and weak system-level orchestration, suggesting that simply adding AI tools to existing environments is not enough to achieve more complex operational outcomes.

Platform Engineering Struggles With Integration And Scale

Platform engineering has also reached broad adoption, with 89 per cent of organisations practising it having implemented an internal developer platform (IDP) and 60 per cent reporting broad adoption across departments.

The primary use case for platform engineering teams is providing developers with access to observability and monitoring dashboards, cited by 74 per cent of respondents. Some 55 per cent named enabling developers with AI-powered tools such as coding copilots and chatbots as their top priority.

But integration remains the primary obstacle. More than a third of platform engineers (37 per cent) reported that integrating with existing tools and systems is their top challenge, followed by difficulty maintaining standardisation across teams (32 per cent) and security or compliance complexity (31 per cent).

The report noted that AI tools exacerbate these challenges because they behave fundamentally differently from traditional application stacks and require their own provisioning and governance methods.

Observability Widely Adopted But Not Deeply Embedded

One of the more striking findings is the gap between observability adoption and observability depth.

Nearly all SRE teams (97 per cent) use observability in some form, whether through minimal logging and monitoring or full observability tools. Some 65 per cent use one or more commercial or open-source observability solutions.

Yet only 40 per cent of platform engineers have embedded observability across all deployment stages. While 77 per cent embed it into at least some services, the uneven coverage means many organisations lack full visibility into their AI and cloud workloads.

Teams rely most heavily on observability during deployment (65 per cent), when they need immediate visibility into performance and risk. But use drops off in later stages, with post-deployment and production monitoring cited by only 55 per cent.

Too Many Data Sources Threaten SLO Effectiveness

The report also identified data overload as a growing barrier to effective SRE operations. Nearly half of SRE respondents (49 per cent) indicated that too many data sources hinder their ability to define and manage service-level objectives, while 45 per cent cited too many metrics as a challenge.

These difficulties are compounded by the introduction of AI workloads, which generate new forms of telemetry and require SLOs to account for probabilistic outputs including accuracy, hallucination, latency and cost. Teams are being asked to define reliability for systems that do not behave deterministically, without established patterns or tooling to draw on.

Some 35 per cent of respondents indicated their monitoring tools do not allow them to easily define and track SLO performance history.

SRE And Platform Engineering Converging

The two disciplines are increasingly intertwined. The survey found that 92 per cent of organisations with SRE programs also have or are initiating platform engineering programs, while 82 per cent of organisations with platform engineering programs also have or are initiating SRE programs.

Some 73 per cent of SRE and platform engineering teams now collaborate and share responsibilities across reliability and platform domains.

Security and compliance practices have also been widely embedded, with roughly 70 per cent of SRE and platform engineering teams embedding security into their platforms as code and more than 60 per cent building regulatory compliance into infrastructure and CI/CD workflows.

Dynatrace Moves To Close The AI Evaluation Gap

Chief Product Officer at Dynatrace, Steve Tack, framed the findings as evidence that existing operational models are insufficient for the demands of production AI.

“SRE and platform engineering laid the groundwork for modern digital reliability, but AI is rewriting the rules,” Tack remarked. “Enterprises need to now move from managing systems to orchestrating them, connecting observability, automation and agentic AI to operate at the speed these initiatives demand, turning insight into action at scale.”

Tack linked the research to the company’s announced intent to acquire AI evaluation firm, Arize, pointing to the disconnect between AI development and AI operations tooling.

“AI engineering teams have been evaluating in one set of tools while operations teams monitor in another, and that gap is no longer sustainable as AI moves deeper into enterprise production,” he added.

Multicloud AI Agent Orchestration

Dynatrace has also moved to address the operational complexity of managing AI agents across multiple cloud providers with a product called Cloud SRE Agents.

The application orchestrates AI agents from AWS, Azure and Google Cloud for automated investigation and resolution of incidents across multicloud environments. When the Dynatrace platform detects a problem, it routes the issue to the appropriate cloud-native agent based on configurable rules, and the agent investigates and recommends or executes remediation steps.

The company cited figures from AWS indicating organisations using the AWS DevOps Agent with Dynatrace reported up to 75 per cent reductions in mean time to resolution.

Customers using the combined approach include Western Governors University, which runs an online learning environment for 200,000 students, and United Airlines, which transports more than 500,000 passengers daily across a hybrid environment that includes more than 500 AWS accounts and 20,000 Lambda functions.

Deterministic Versus Probabilistic AI

Underpinning much of Dynatrace’s positioning is a distinction between probabilistic and deterministic approaches to AI-powered observability.

The company argues that many observability platforms layer probabilistic AI on top of siloed data, using large language models to correlate signals and rank likely causes without being able to determine correctness. This approach can produce different answers for the same problem and is prone to hallucinations, requiring human validation that increases overhead and delays resolution.

Dynatrace contends its platform uses deterministic AI, which traces exact causal changes to identify what happened and why, rather than proposing likely candidates. Vice President of Portfolio and Strategy at Dynatrace, Bob Wambach, argued that when agentic AI systems act on business-critical applications, the cost of being “mostly right” becomes operationally unacceptable.

The Road Ahead

Gartner projects that by 2028, 80 per cent of enterprises will adopt SRE practices across their organisations, up from 30 per cent in 2024. The Dynatrace report suggests the discipline is well-positioned for that growth, but the challenge has shifted from adoption to orchestration.

The survey was conducted by Qualtrics partner, Y2, on behalf of Dynatrace between October 2025 and January 2026. Respondents were senior leaders, decision makers, managers and supervisors at enterprises with annual revenues of $500 million or more, drawn from 16 countries across the Americas, EMEA and Asia-Pacific.

The margin of error was plus or minus 3.2 per cent at a 95 per cent confidence level.

Last Updated on September 5, 2026 by Nick Ross