87% of Tech Leaders Report AI Agent Use as Visibility Lags
Dynatrace's 2026 survey finds 87% of technology leaders report moderate or extensive AI agent use, but only 27% address visibility and people verify 66% of decisions.
By Grace Kim
3 min read

By Grace Kim
Dynatrace found that 87% of surveyed enterprise technology leaders reported moderate or extensive use of AI agents for observability and IT operations, while only 27% said they address visibility into agent decisions and actions.
The figures came from Dynatrace's State of Observability 2026 findings, published November 10. They draw on an online survey the company commissioned in August 2026 of 1,202 enterprise technology leaders across 12 countries.
Respondents described their own organizations; researchers did not independently verify behavior in production. Dynatrace sells observability software, so the results offer a snapshot of the challenges its respondents reported, not a census of enterprise AI use.
The survey maps reported adoption, not independently proven performance.
How far does agent autonomy extend?
Dynatrace said one organization might use an agent to summarize an incident and propose a response for an engineer to review. Another might let an agent execute a predefined action inside a narrow part of its infrastructure. Both count as agent use despite different levels of authority.
The 87% figure therefore covers monitoring software and running IT systems. It does not mean that 87% of companies let agents make consequential decisions without human involvement.
Dynatrace's survey offered three measures of autonomy:
- 66%: People still verify an average of 66% of AI-powered decisions.
- 12%: Respondents reporting fully autonomous operations.
- 44%: Respondents targeting fully autonomous operations within five years.
Dynatrace described the 44% figure as a measure of ambition rather than a forecast. Human verification remains routine while autonomy targets rise.
What blocks broader autonomy?
Integration complexity was the most frequently cited barrier, according to 31% of respondents. Dynatrace also identified skills, oversight and data quality as additional challenges.
Integration complexity is the most commonly reported obstacle to broader autonomy.
What must teams be able to see?
Dynatrace said standard system monitoring can show an engineer that a service responded and how long it took. It does not reveal whether an agent selected an appropriate action or whether that action resolved the underlying incident.
For an agent that recommends restarting a service, Dynatrace said a useful record would identify the alerts behind the recommendation, whether a person approved it and what happened afterward. Engineers could then determine whether the restart caused recovery or merely coincided with it.
Dynatrace found that 43% of respondents considered visibility into agent decisions an important observability requirement, compared with 27% who said their organizations addressed it. That gap compares stated importance with reported coverage; it does not mean every other respondent had no logs.
Teams need records that explain cause and effect, not merely service uptime.
What do AI workloads cost?
Some 81% of respondents reported higher compute and infrastructure spending from AI workloads. Seventy-seven percent said they needed better visibility into how those costs were allocated.
Dynatrace warned that measuring only the time operators spend on an incident can make automation look cheaper than it is. A team that excludes the extra resources an agent consumes cannot compare the full cost of different approaches to the same work.
Cost comparisons require teams to connect labor time with agent resource use.
When should an agent gain broader authority?
In a November 17 follow-up, Dynatrace outlined a four-step progression:
- Gather evidence.
- Recommend actions.
- Act with human approval.
- Move toward broader autonomy where warranted.
Dynatrace advised teams to test an agent against a specific task rather than expand its permissions because it performed well in unrelated work. Its example: an agent that reliably categorizes alerts has not shown that it can choose when to change a production system.
The survey's practical test is whether a team can reconstruct what an agent did, justify the authority it held and measure whether the action helped. Without those answers, higher adoption may reflect only more agents entering the workflow, not readiness to run it.
For enterprises moving toward autonomy, broader control should wait until teams can audit both the agent's choice and its result.
Source: unite.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,447 articles