Trust and visibility are the top technical barriers to agents in production
A survey of 919 enterprise leaders finds agentic projects held back by governance rather than by model capability.
What happened
Dynatrace published its first Pulse of Agentic AI report in 2026, surveying 919 senior leaders at large enterprises. It asked how agentic projects move from pilot into production.
Most coverage framed the answer as a story about stalled pilots, though the report says something rather different. Its own wording is that organisations are "quickly moving toward maturity".
What the source says
On maturity the prose gives three figures. Half of respondents have projects in production for limited use cases, while another 44 percent have them in broad departmental adoption. A further 23 percent have reached mature enterprise-wide integration. The stage chart adds 36 percent still running early pilots. Those figures sum to more than 100 percent, so they overlap rather than partition the sample, and the report doesn't state how the question was put.
On barriers the key-insight line is that "trust and lack of observability are top technical barriers to production". The supporting prose offers one example, which is setting clear rules for when an agent acts alone. Other challenges include limited visibility into agent decisions. The barrier chart also carries security, skills shortages, budget and stakeholder buy-in.
The barrier chart's percentages are not quoted here. Those figures are in the report, but the ranking above rests on what it states in words, and chart items are named as chart items.
On oversight, 69 percent of agent decisions are currently verified by a human. Only 13 percent build solely autonomous agents, while 23 percent build only human-supervised ones and 64 percent build both. The report calls that intentional rather than a passing stage.
Methodology worth knowing: fieldwork ran across 15 countries in late 2025, with a stated margin of error of 3.2 percent. An observability vendor published it and recommends observability, so read the framing accordingly.
What we think
Coverage of this report has been consistently worse than the report itself. The reading that travelled said half of agentic projects never leave the pilot stage, which is close to the opposite of the finding. Half of respondents have projects in production, and the report describes maturity as growing fast.
What the report establishes is more useful than a league table, because the obstacles at the top are about trust rather than capability. Can you set rules for autonomy, and can you see what the agent did afterwards? Somebody who knows the business can answer both of those.
The 69 percent human-verification figure is the one worth sitting with, because it's not a sign of immature technology. Set against the 13 percent building fully autonomous agents, it reads as a choice by people who have priced an unattended mistake.
This matches how we build, in that research, scoring, drafting and CRM updates all run unattended. Anything that sends, spends or commits you to an answer waits for a person, because that line sits where being wrong stops being recoverable.
The obstacles at the top are about trust and visibility. Can you set rules for when an agent acts alone, and can you see what it did.
What to do about it
Write your autonomy boundary down before you build anything. Keep two lists, one for what runs unattended and one for what waits for a human, since most teams settle it by implication and find the disagreement late.
A useful test is whether you could fix the mistake quietly, or whether it's already reached somebody outside the company. Anything in the second group belongs behind a person until the layer feeding it has earned some trust.
When you read a statistic about AI adoption, go to the primary source before you repeat it. Then check the sentence around the number, because removing a figure and leaving the claim it supported still standing is how a correction quietly fails.
Writing that boundary down, and agreeing which side each action sits on, is the first hour of the Diagnostic.
Sources
How we choose and label these: our sourcing standard.
-
The pulse of agentic AI in 2026: building trust in autonomous operations from pilot to production →
Survey of 919 senior leaders at enterprises with revenues above 100 million dollars. Fieldwork ran November to December 2025 across 15 countries, with a margin of error of plus or minus 3.2 percent. Published by an observability vendor whose recommended fix is observability, so read the framing accordingly. The underlying figures are reported transparently.
-
The pulse of agentic AI 2026 (full research report, PDF) →
The full 33 page report. Every figure we cite is read from this document rather than from coverage of it.
-
What businesses must fix before letting AI agents act →
Third-party coverage of the Dynatrace report. We originally cited this alone and stated in print that the underlying report couldn't be located. That was wrong: it's public, and reading it reversed our conclusion. Kept here as the record of the coverage, not as evidence.