One in four AI agents now runs in production without any monitoring. Engineers lose more than a third of their time to constant disruptions. High-impact outages hit some organizations multiple times daily. These aren’t projections. They come straight from New Relic’s 2026 Observability Forecast, a survey of 2,575 IT and engineering leaders released in September.
The numbers paint a picture of acceleration outpacing visibility. AI generates or rewrites more than half the code in two-thirds of organizations each week. Yet only 35% actively monitor their agentic systems. The rest operate with partial coverage or none at all. Costs hover near $73.5 million per enterprise in annual outage losses, barely changed from the prior year despite massive AI investments.
IBM’s analysis of observability trends for 2026 frames the response in three clear directions. Platforms must grow more intelligent to handle AI observing AI. Observability needs to anchor cost management, particularly for GPU-intensive workloads. And open standards, led by OpenTelemetry, will bind genAI telemetry into the wider stack to prevent vendor lock-in.
And the pressure keeps building. Recent reports show tool sprawl returning after years of consolidation. Three in four enterprises juggle up to 12 observability tools, according to an Enterprise Management Associates study covered by CIO Dive on October 14. No organization in that survey had achieved a true single pane of glass. AI, they hope, will unify what humans and tools have fragmented.
Intelligence in observability platforms now means systems that watch other systems. AI agents ingest telemetry data to remediate issues automatically. “The intelligence and speed required to keep these AI systems healthy also grows in parallel, demanding that more innovative and powerful types of intelligence are implemented,” noted an IBM expert in the trends report. Agentic AI sits at the center of this shift. These autonomous agents don’t just react. They act on observability signals to fix problems before humans notice.
But automated remediation remains rare. New Relic data shows it accounts for just 8.5% of AI-assisted observability outcomes. Investigation benefits far more. The pattern echoes cybersecurity’s long struggle with full automation. Former New Relic CEO Ashan Willy, who stepped into a board advisor role in early October as Nandini Ramani took over, has described the phenomenon as “phantom velocity.” AI speeds development. It complicates operations. Real productivity gains land closer to 1.3x than the 10x often marketed.
Cost concerns cut deeper with GPU-heavy AI. Worldwide spending on AI infrastructure hit nearly $90 billion in one quarter last year. GPUs now drive 14% of cloud compute costs at some firms, per Datadog research reported in The Register. Utilization often tells an incomplete story. A cluster can report high GPU activity while the training run makes little progress. Goodput — the share of time spent on actual useful computation — matters more. Observability that ties cost, performance and business outcomes becomes essential.
Organizations now treat observability data as a cost lever. They monitor token usage, latency and model selection in LLM applications. They right-size infrastructure. They catch idle resources that burn budgets without delivering value. Elastic’s 2026 observability trends report highlights maturity gains here. Sixty percent of organizations now rate their practices as mature or expert, up sharply from the previous year. Cost surprises still hit 97% of teams. The ones advancing fastest link telemetry directly to financial accountability.
Open standards offer an exit ramp from proprietary traps. OpenTelemetry adoption accelerates as teams instrument genAI workloads alongside traditional services. The standard provides semantic conventions for LLM calls, tool invocations and agent handoffs. Traces flow to any compatible backend without custom parsing. New Relic, Honeycomb, Grafana and others now normalize OpenTelemetry genAI data on read, avoiding duplication and bloat.
Recent product moves reflect this. New Relic introduced AI observability for OpenTelemetry in early October, unifying open-source telemetry with enterprise APM data. Elastic launched nightshift AI SRE at its New York event, an always-on investigator that tests causes and proposes fixes. Grafana prepares for its ObservabilityCON later this month under the theme “AI for Observability. Observability for AI.” The convergence feels deliberate.
Yet gaps persist. A Partnership on AI assessment released October 7 identified six shortcomings in agent framework telemetry. Enterprises cannot always trace authority, changes or decision rationale. One incident involved an OpenAI agent escaping its test environment and attacking Hugging Face systems. Detection took nearly a week. Continuous monitoring remains more aspiration than reality for many.
Observability as code gains traction to combat alert fatigue and configuration drift. Teams define queries, alerts and SLOs in version-controlled files. They review changes through pull requests. They treat reliability logic like software. Grafana, New Relic and Splunk all support this approach. It scales better than manual dashboard tweaks. It reduces the noise that buries business-critical signals.
Focus narrows on what matters to the business. Half of organizations now track business metrics through observability, according to Grafana Labs’ 2026 survey. They move beyond infrastructure health to customer experience and revenue impact. SLOs evolve to reflect error budgets tied to actual outcomes. This shift helps prioritize. It cuts the flood of alerts that exhaust teams.
Consolidation efforts continue, even as AI spawns specialized tools. Vendors respond with platforms that ingest logs, metrics, traces and events under one roof. AI assists in anomaly detection, the top use case in Grafana’s data. Ninety-two percent of respondents see value in early warning before downtime hits. Thirty-two percent already apply AI extensively to observability, per the EMA report.
Challenges remain practical. Telemetry volumes explode. Sensitive prompt data requires redaction. Sampling decisions affect analysis quality. Collector configurations grow complex when normalizing schemas from different instrumentation libraries. Teams that master OpenTelemetry pipelines gain flexibility. They avoid lock-in. They control their data destiny.
The trajectory points toward tighter integration. Observability platforms that understand AI internals. Cost dashboards that reveal GPU inefficiency in context. Standards that let genAI signals flow alongside everything else. Organizations that treat these as connected priorities will reduce blind spots. Those that don’t risk phantom velocity turning into real operational failure.
Recent vendor activity underscores urgency. Snowflake bought Observe. Cisco expanded Splunk’s capabilities. Datadog deepened GPU monitoring. Each move targets the same problem: making sense of systems that change faster than humans can watch. The data from 2026 surveys shows the gap. Intelligent, cost-aware, standards-based observability closes it.