AI Observability Forecast: Gartner Sees 40% Adoption by 2028

The forecast

Gartner has put a number on something AI engineering teams have been saying for a while: once AI systems go into production, they need their own monitoring. The firm predicts that by 2028, 40% of organizations deploying AI will run dedicated AI observability tools to track model performance, bias and outputs.16 Gartner analysts presented the forecast at the firm's IT Infrastructure, Operations & Cloud Strategies Conference in Sydney on May 11–12, 2026, and the research behind it is a Gartner client report titled "Predicts: AI Will Unlock Observability at Scale."3

Gartner treats observability as a property of a system: you can understand what it is doing from its outputs and answer questions about its behavior. For AI, the firm says this takes dedicated tools that assess model drift, bias and the logic of large language models.6 Padraig Byrne, a VP Analyst at Gartner, said AI is now widespread but most organizations still have no reliable way to monitor or trust it. He called that a visibility gap that makes scaling AI risky, because AI decision-making is hard to inspect and its errors can lead to financial losses, reputational harm and regulatory attention.2

The most useful part of the announcement is Gartner's explanation of the cause. Byrne said the move to specialized tools is being driven by executive worry about risk in complex models and agentic AI, not just by concern for infrastructure or application health.5 He also warned that without standardized model telemetry, infrastructure and operations (I&O) teams will take longer to resolve AI incidents and will have to trace opaque model behavior by hand.4

What Gartner tells I&O leaders to do

Gartner gives four recommendations, and every outlet that covered the forecast repeated them the same way:

• Make monitoring mandatory for every production AI deployment, with continuous tracking of fairness, drift and data-quality metrics.1
• Use one monitoring framework across data science, MLOps and engineering teams so that silos don't slow down troubleshooting.2
• Invest in infrastructure that can handle large volumes of model telemetry, especially tools that support distributed tracing of AI inference calls.6
• Plan ahead for monitoring AI platform performance, catching unsanctioned "shadow" AI use and controlling costs.4

The third point is the most concrete for practitioners. Distributed tracing of inference calls is how teams debug agentic systems: they record each model call, retrieval step and tool invocation so a failure can be traced to where it started. Gartner is endorsing that approach as an enterprise standard, not leaving it as a habit of AI-native startups.

The 40% figure looks low next to practitioner surveys

The coverage diverges most on how widespread AI observability already is. Gartner's 40% describes a point two years out. One industry guide cites Gartner as estimating that only about 15% of generative AI deployments were instrumented in early 2026.11 A practitioner survey of more than 1,300 professionals, reported in a market-research release, found that 57.3% already had agents in production, about 89% had some form of agent observability, and 62% used detailed step-level tracing.12

These figures are not really in conflict. The survey sample leans toward teams already building agents, and "some form of observability" can mean anything from log files to a full evaluation platform. Gartner's forecast is about dedicated tooling across all organizations deploying AI, including enterprises whose AI use is mostly purchased software and copilots. One analysis that cites the same survey data notes that fewer than a third of teams are satisfied with the observability they have.14 Our read is that many builders have adopted something, but enterprise-wide, policy-driven adoption is still in an early phase. Gartner's figure is best read as a forecast of that second phase.

Money has moved ahead of the forecast

Vendors and acquirers acted on this market well before Gartner published its number. ClickHouse bought Langfuse, the open-source LLM tracing, evaluation and prompt-management platform, in January 2026 as part of a $400 million Series D that valued ClickHouse at $15 billion. Langfuse reported more than 2,000 paying customers at the time of the deal.11 In February, Braintrust raised an $80 million Series B at an $800 million valuation, led by ICONIQ.1113 Cisco bought Galileo, a platform for AI observability, evaluation and runtime guardrails, and plans to combine it with Splunk Observability.13 Accounts of the Galileo timeline vary. One outlet dates the acquisition to April, another reports it closing in mid-May, and a May 12 report still described the deal as planned.13152

Other deals followed the same pattern. OpenAI acquired the evaluation tool Promptfoo, and Mintlify bought the logging proxy Helicone, whose standalone product is now in maintenance mode.15 One review of the category reports that Palo Alto Networks bought the AI gateway Portkey in May 2026, after its $3.35 billion acquisition of Chronosphere.1612 Datadog acquired Adaptive ML, a reinforcement-learning operations startup, in June.13 Fiddler raised $30 million in January, and Coralogix raised $200 million in June.12

One editorial tracker counts eight acquisitions of LLM eval and observability companies between March 2025 and May 2026. It sorts the buyers into three groups: foundation-model labs that want to own the testing workflow, infrastructure vendors adding an application layer on top of compute or databases, and adjacent developer-tools companies.15 That model explains the deals better than any one-off story. Labs such as OpenAI and Anthropic want evaluation built into their platforms. Database and observability companies want AI telemetry, which is becoming one of the most valuable workloads they can host.

Some deal figures should be treated with caution. Prices for Langfuse range from analyst estimates of $100–200 million to a separately cited $800 million valuation.1520 Market-size estimates also vary widely depending on how the category is defined. One puts the LLM observability market at about $2.69 billion in 2026.11 Another values the narrower AI agent observability segment at $0.9 billion and projects $14 billion by 2035.12

Evaluation, prompts and agents are converging

The products themselves are absorbing neighboring tools. Langfuse covers tracing, evals and prompt management in one open-source stack, and added monitors and alerts in June 2026.16 Braintrust puts evaluation at the center: it clusters production failures into topics and turns them into new datasets and scorers.16 Portkey combines a gateway to over 1,600 models with observability, guardrails, prompt management and access control for agents.16 Separate prompt-management and offline-eval tools are being folded into platforms that connect pre-release testing to live production traffic.

The agentic case raises the requirements. Commentary collected by APMdigest notes that AI observability for agents means being able to reconstruct a decision: what context the agent retrieved, which tools it called and what the model saw at each step. Teams use that reconstruction both to debug and to answer for outcomes.10 Grafana Labs' Mat Ryer summed up the governance argument: a system whose actions can't be seen shouldn't be given more autonomy.10 At the standards level, OpenTelemetry's generative AI semantic conventions now include tracing for Model Context Protocol tool calls, but the specification is not yet marked stable.11

Our take

Gartner's forecast confirms a trend more than it predicts a new one. Capital and acquisitions had already declared AI observability a necessary layer of the stack before the Sydney presentation. The forecast matters because of where it is aimed: I&O leaders who control budgets and policy. Gartner describes monitoring as a governance requirement, not a developer convenience, and that is the kind of framing that moves spending from pilot projects into standard procurement.

There is a practical risk alongside the opportunity. With so many vendors being acquired, teams that wire their evaluation pipelines to one proprietary SDK could have to rebuild when their vendor changes owners or terms. One commentator points to exactly this risk, citing the year's run of acquisitions.18 Gartner's advice to standardize frameworks and prioritize distributed tracing makes the most sense paired with open telemetry standards. For the 60% of organizations the forecast expects to still lack dedicated tools in 2028, the risk is the one Gartner names: incidents that take longer to diagnose and governance gaps that show up only when something has already gone wrong.