An AI system can be available, responsive and still be wrong. That simple observation sits at the heart of Dynatrace’s new research report, The State of Observability 2026, and it explains why so many technology leaders are rethinking what they expect from their monitoring investments. AI is quickly becoming some of the most mission-critical software enterprises run, yet the evidence that it is reliable, secure, governed and cost-effective has not kept pace with the ambition to deploy it.
The report is based on a survey of 1,202 technology leaders at enterprises with annual revenue of $100 million or more, conducted between August 10 and August 24, 2026, across 12 countries including India, the United States, the United Kingdom, Germany, Japan and Australia. Respondents ranged from the C-suite to mid-level executives, and the research was run with Qualtrics and Y2 Analytics, with a margin of error of plus or minus 2.8 percent at 95 percent confidence. The study was produced by Dynatrace, an observability vendor, so its conclusions should be read with that in mind, but the underlying numbers describe pressures most CIOs will recognise.
AI is the baseline, and the budgets are following
The adoption question is largely settled. According to the report, 99 percent of respondents use AI in their operations, and portfolios are expanding: 49 percent now run between six and 25 production AI projects, up from 42 percent a year ago. Most are building on existing models, APIs and AI features rather than starting from scratch, with 62 percent integrating third-party models or APIs into applications and workflows.
Money is moving accordingly. Nearly half of organisations, 46 percent, spend at least $1 million a year on observability, and 78 percent expect that budget to grow over the next year. Among those expecting an increase, 96 percent say AI adoption is at least a moderate driver. The shift is also visible in buying criteria. AI-powered capabilities lead the list of what leaders want from an observability platform at 57 percent, followed by visibility into AI workloads and large language model behavior at 49 percent. Notably, the report finds that AI is expanding the observability mandate rather than displacing older priorities, with infrastructure monitoring, application security, multi-cloud optimization and generative AI monitoring all drawing similar planned investment.
Complexity is the price of speed
The cost of that expansion shows up in the operating environment. Asked what drives technology complexity, 43 percent of respondents named the growth of AI, large language model and agentic workloads, ahead of multi-cloud management at 37 percent and security threats at 35 percent. Those pressures land on estates that were already stretched: 62 percent of organizations use three or more cloud providers and 56 percent use six or more observability tools.
The resulting blind spots are strikingly concentrated. Only 6 percent of respondents say they have no significant visibility gaps, and the five largest gaps, each cited by 41 to 45 percent, all involve AI, covering AI infrastructure, agent activity, interactions between AI and conventional systems, model behavior in production, and the cost of running it all. Gaps unrelated to AI rank at 36 percent or below.
The bill is arriving as well. Some 81 percent say AI workloads have raised compute and infrastructure spending, 78 percent report higher energy costs, and 77 percent say the need for cost-allocation visibility has grown. Yet 39 percent say they lack insight into energy consumption and emissions, and the same share cannot attribute them to specific workloads. As the report puts it, organisations cannot optimize AI resource consumption they cannot measure and attribute.
The proof gap
The most pointed finding concerns how organisations verify AI itself. Two-thirds, 66 percent, rely on observability tools built into their AI providers’ platforms. Such tools report token counts, latency and API errors, but they leave it to the enterprise to connect AI behavior and quality across providers and the wider technology estate.
Across nine AI observability requirements, the report finds an average 14-point gap between what leaders consider important and what they actually do today. Half say real-time monitoring of output quality and accuracy is an increasingly important requirement, while 38 percent currently address it. The shortfall widens where the stakes may be highest. Some 43 percent consider it important to understand how AI agents reach decisions and take actions, but only 27 percent do so today, and while 39 percent want to measure the business impact of AI behavior, just 24 percent currently measure it. Skills gaps are the leading barrier, cited by 37 percent, followed by fragmented tooling at 34 percent.
Autonomy: big goals, careful hands
Those gaps matter because organisations are pushing AI toward greater independence. According to the report, 38 percent of organisations have reached supervised autonomy, where AI acts within human-defined guardrails, and just 12 percent operate fully autonomously. Yet 44 percent aim to get there within five years, and the ambition rises with scale: half of organisations with revenue of $10 billion or more are targeting full autonomy, against 29 percent of those with $100 million to $499 million.
AI agents are already part of day-to-day operations, with 87 percent reporting moderate or extensive use for observability and IT operations. Oversight remains substantial, however. Respondents estimate that humans verify an average of 66 percent of AI-powered decisions, and organisations use nearly four validation methods on average, from checking outputs against trusted source data, cited by 45 percent, to continuous automated evaluation and audit logging. The brakes are practical rather than philosophical. Integration complexity is the top barrier to autonomous operations at 31 percent, ahead of skills gaps at 29 percent and a lack of trust in unverified AI decisions at 27 percent.
Security: the window has collapsed
The report also argues that AI is shrinking the time defenders have. It describes the window between vulnerability disclosure and weaponized exploitation as having collapsed from months to hours, and in some cases to before public disclosure. Among the 741 respondents responsible for application security, 55 percent experienced a zero-day incident in the past year, including 20 percent who reported significant operational or business impact, even though two-thirds describe their security posture as runtime-connected or fully integrated.
Maturity is not enough on its own, the report suggests, because faster detection does not answer which vulnerabilities pose real risk. About 28 percent remediate critical vulnerabilities within 24 hours, and 45 percent now continuously prioritise using runtime exploitability status. A further 82 percent describe their security and observability approaches as significantly or fully converged, with faster detection of AI misuse the leading motivation. The obstacles to going further are mostly organisational, with silos between security and engineering teams leading at 23 percent.
Open standards as the connective tissue
To keep evidence and context portable across tools, many organisations are turning to open instrumentation. Some 77 percent of respondents have adopted OpenTelemetry in some form, but only 20 percent have made it a strategic standard. Eighty-eight percent of adopters report high or moderate value, though skills gaps, at 33 percent, and rising telemetry costs, at 31 percent, complicate scaling. Supporting observability of AI and large language model workloads is now a primary adoption driver, cited by 37 percent.
Industry voices quoted in the report echo the theme. Deloitte’s Alvaro Martin argues the next phase of enterprise AI is about “scaling autonomy where it creates value”, rather than deploying more agents, while AWS’s Mona Chadha frames the challenge as scaling AI efficiently, with observability, cloud economics and resource optimization treated as one discipline.
The report’s conclusion is that observability must evolve from telling teams whether systems are up to showing whether AI is working, what it costs, and whether it delivers the business outcome it was funded to produce. In practice, that points to a handful of moves: connecting AI telemetry with application, infrastructure, security and business data; evaluating production AI continuously for quality and drift; making consumption attributable to workloads and teams; expanding AI decision authority only as reliability, explainability and governance are demonstrated; and standardizing telemetry so that insight survives changes in tools and providers.
For CIOs, the uncomfortable takeaway is less about technology than about accountability. Budgets are rising, agents are in production and autonomy targets are set. The organisations that can show their AI is behaving as intended, and show it with evidence, will be the ones best placed to give it more responsibility.