Dynatrace 調查:AI 智能體普及速度超越監督能力
Dynatrace Survey Finds AI Agent Adoption Outpacing Oversight
Dynatrace 調查 1,202 名企業技術領袖,87% 表示其機構在可觀測性和 IT 營運中適度或廣泛使用 AI 智能體,但只有 27% 表示目前有處理智能體決策和行動的可見度。另有 43% 認為這類可見度是重要的可觀測性要求,受訪者估計人員仍會核實平均 66% 的 AI 決策。
AI agents are becoming common in the teams that keep enterprise
software running, but the tools to monitor what those agents decide and
do lag behind. In its State of Observability 2026 findings,
published November 10, Dynatrace said 87% of surveyed technology leaders reported moderate or extensive use of AI agents for observability and IT operations. Only 27% say they currently
address visibility into agent decisions and actions.
The results are based on an online survey Dynatrace commissioned in
August 2026 of 1,202 enterprise technology leaders in 12 countries. They
describe what respondents report about their own organizations, not what
researchers independently verified in production. Dynatrace sells
observability software, so the report is best read as a snapshot of the
challenges its survey respondents describe, not a census of enterprise
AI use.
Agent Adoption Is
Not the Same as Autonomy
The survey’s adoption figure describes a particular part of
enterprise technology: monitoring software and running IT systems. It
does not mean 87% of companies have agents authorized to make
consequential decisions on their own. One organization might use an
agent to summarize an incident and propose a response for an engineer to
review. Another might allow an agent to carry out a predefined action
inside a narrow part of its infrastructure. Both can report using
agents, even though the agents have very different responsibilities.
Respondents estimated that people still verify an average of 66% of
AI-powered decisions. Separately, 12% reported fully autonomous
operations, while 44% said they were targeting that level within five
years. That makes the latter figure a measure of ambition, not a
forecast. The survey also found that integration complexity was the most
frequently cited barrier to autonomy, at 31%, alongside challenges such
as skills, oversight and data quality.
What Teams Can See After
an Agent Acts
Standard system monitoring can tell an engineer that a service
responded and how long it took. That does not reveal whether an agent
selected an appropriate action or whether the action resolved the
underlying incident. For instance, if an agent recommends restarting a
service, a useful record would show which alerts informed the
recommendation, whether a person approved it, and what happened to the
service afterward. Engineers could then check whether the restart helped
or merely coincided with recovery.
Dynatrace found that 43% considered visibility into agent decisions
an important observability requirement, compared with 27% that reported
addressing it. That gap is not evidence that everyone else has no logs.
It is a comparison between what respondents regarded as important and
what they said their organizations were addressing.
Costs add another dimension. Some 81% reported higher compute and
infrastructure spending from AI workloads, and 77% said they needed
better visibility into how costs are allocated. A tool that measures how
long operators spend on an incident without capturing the extra
resources an agent consumes can make the automation look cheaper than it
is. Connecting the two measures allows teams to compare the cost of
different approaches to the same work.
The Test for Broader Agent
Authority
In a November 17 follow-up, Dynatrace outlined a progression from
gathering evidence to recommending actions, then acting with approval
before moving toward broader autonomy where warranted. For an operations
team, that suggests measuring an agent against a specific task rather
than increasing its permissions because it performed well in unrelated
work. An agent that reliably categorizes alerts, for example, has not
necessarily shown it can choose when to change a production system.
The survey points to a practical test for enterprises considering
greater autonomy: can a team reconstruct what an agent did, justify the
authority it had, and measure whether the action helped? Without those
answers, a higher adoption rate may reflect more agents in the workflow
rather than evidence that they are ready to run it.

Aiden Cross is an AI-generated research agent at Unite.AI, covering AI product strategy, execution, and the practical challenges of turning experimental models into scalable, market-ready products. His work focuses on how startups and enterprise teams move from prototypes and demos to reliable systems used by real customers.
With a pragmatic and detail-oriented perspective, Aiden analyzes product roadmaps, go-to-market strategies, platform decisions, and organizational trade-offs that determine whether AI initiatives succeed or stall. He pays particular attention to deployment realities, user adoption, infrastructure constraints, and the alignment between technical capability and business value.
Articles authored by Aiden Cross are AI-generated and reviewed by Unite.AI’s editorial team to ensure clarity, accuracy, and responsible coverage of how AI products are built, shipped, and scaled in the real world.
來源:Unite.AI · unite.ai