Over the past nine months, companies gave their AI agents more freedom and trusted them less. The share letting agents act with no human approval nearly doubled, from 4.2% to 7.5%. The share who say they fully trust an agent to run autonomously fell from 37% to 32%. Both numbers come from the same people, asked the same questions, in December and September.
They are not the only numbers that moved. The share who say they’d be “very likely” to switch vendors for stronger agentic capabilities fell from 46% to 38%. And the share who say they’re adopting agents to meet leadership expectations nearly doubled, from 13.5% to 25.3%.
More autonomy, less conviction, more pressure from above. That’s not a category losing interest; it’s a category that has started running agents in production and is adjusting its expectations to what it sees. This post walks through what moved on AI agent autonomy, trust, and buying intent between the two waves of The State of Agentic AI.
The leash got longer
The survey asks how teams govern what their agents can do, with three answers: approval on every action, an intermediate mode where reads are automatic and writes need sign-off, or no approval at all.
Approval-on-everything dropped eight points, from 38% to 30%. Intermediate mode rose from 53% to 59%. No-approval rose from 4.2% to 7.5%. Almost nobody is adding gates; the movement is all in one direction, toward letting agents do more without a person in the loop for each step.
What agents are allowed to do looks stable at first glance. Sending internal reminders (65% to 66%), creating tickets (61% to 58%), scheduling meetings (55% to 56%), and updating records (45% to 46%) barely moved. The two actions that did move are the two with external consequences: sending customer emails fell from 49% to 44%, and changing system settings fell from 30% to 27%. Teams are giving agents more latitude inside the building and slightly less at the edges where a mistake reaches a customer or a production system.
The promises got smaller
Three questions measure how much teams expect from agents, and all three softened.
Full trust in autonomous operation went from 37% to 32%, with the difference moving into partial trust (52% to 60%), not into low trust, which also fell. Respondents aren’t losing faith in agents; they’re getting specific about which operations they’ll hand over.
Full trust in autonomous agents fell five points in nine months; partial trust absorbed the difference.
Vendor switching intent went from 46% “very likely” to 38%, with the difference moving into “possibly” (48% to 55%). The total who would consider switching is unchanged at 93%. What changed is the urgency.
Adoption motives shifted the same way. “Launch differentiated features” fell from 30% to 26%. “Meet leadership expectations” rose from 13.5% to 25.3%. Nine months ago, a quarter of respondents said their company built custom workflows for competitive advantage; now one in four says, at least in part, that the boss asked.
“Meet leadership expectations” nearly doubled as a reason for adopting agents between the two waves.
The combination is hard to read as hype. Teams that were going to be disappointed by agents have mostly had the chance to be. What’s left is a more sober version of the same plan.
What teams want from the platform underneath
The question that moved most in the buying section is what respondents require from an agentic AI platform. Reliability stayed the top answer and rose from 62% to 65%. Security controls rose from 53% to 61%. Observability and debugging rose from 21% to 27%, and compliance readiness from 27% to 32%. Performance, meanwhile, fell from 57% to 50%.
That’s a list written by people who have run agents, not people planning to. Raw speed matters less once the agent works; what matters is whether it keeps working, whether security will sign off, and whether you can see what it did. The September survey’s other headline finding fits here: 54% had an agent make a mistake in the past 90 days, and only 53% could fully audit what their agents did last week. A platform wish list that rises on security and observability is what you’d expect from a group where half have cleaned up after an agent and half can’t fully reconstruct what happened.
The blocker question moved in the same direction. Security approvals remained the top obstacle at 42%, and engineering bandwidth fell from 34% to 24%, which suggests the constraint has shifted from “can we build it” to “will we be allowed to run it.” (September added a “nothing is blocking us” option that 23% chose, so the drops across the board are partly that; the ranking of blockers is the reliable comparison, not the levels.)
What this means if you’re planning for 2027
The December survey described a market deciding whether to adopt agents. The September survey describes a market that has adopted them and is now deciding how much to let them do. The adjustments it’s making are the ones you’d want: narrower trust, more autonomy where the blast radius is small, higher standards for the platform underneath, and a clearer sense that reliability and auditability are what separate an agent you keep from one you turn off.
For product leaders, the practical read is that the window for “we have agentic features” as a differentiator is closing. 70% of respondents say agentic AI is already an expected baseline in their market or will be within a year. What buyers are starting to ask instead is the question the rest of this data points to: when your agent does something wrong, can you show me what it did and under whose authority?
That’s an infrastructure question as much as a product one, and it’s the one we’ll take up in the next post in this series.
Email and Contacts API 
Calendar API & Scheduler 
Notetaker API 
Agent Accounts 
Sales & CRM 
Recruiting & ATS 
Healthcare
Financial
Legal
Real Estate
Education
Field Services
Docs
SDKs
Guides
Integrations
API status
Changelog
Support
Professional Services
