G2 pulled every approved review of AI agents left by companies with more than 1,000 employees across the five most-reviewed agent categories — 3,222 reviews covering 231 products, from January 2025 to 23 October 2026. Then it read them for what buyers actually want. The answer is less autonomy than vendors are selling, and more control than most products provide.
Usage first, because it is the good news: enterprise reviews of AI agents more than doubled, from 983 in 2025 to 2,239 in 2026. Fifty-three per cent of reviewers now reach for their agent most of the time or every time, up from 47%, and the share using it every single time rose from 11% to 17%. The work being handed over is mostly unglamorous — process automation leads at 34% of use cases, then customer service at 27% and software engineering at 21%, with sales development at 12% and marketing at 7%. Enterprises prove agents in the back office before trusting them with revenue.
Then the leash. Seventy-seven per cent of enterprise reviewers say they trust their agent "quite a bit" or "very much". Asked how much autonomy they would grant, 60% choose "autonomy with guardrails", 21% "limited autonomy", 11% want a human to approve every action, and 7% would hand over total autonomy — up one point from 6% in 2025. Trust lengthens the leash but rarely removes it: even among reviewers who trust their agent "very much", only 15% would grant total autonomy, and roughly one in ten in that group still wants a human in the loop. The people who carry the risk hold tightest — 5% of admins and 2% of executive sponsors would allow full autonomy, against 10% of everyday users.
What would earn more trust is not accuracy. Ninety-two per cent already rate output accuracy 4 or 5 out of 5, against 82% for how well the agent explains its results. Asked what would raise trust, 79% picked a visibility or control fix: explaining decisions (32%), more guardrail settings (29%), or a warning before a questionable action (18%). Only 21% asked for better reliability.
That caution has a bill. G2's separate 2026 AI Agents Report survey, cited in the same analysis, puts a number on what it calls the Distrust Tax: 31% of organisations spend six or more hours a week overseeing agents, and 45% say oversight and rework consume more than a quarter of the gains.
The reviews also show the early glow fading. Average likelihood-to-recommend slipped from 9.24 in 2024 to 9.10 in 2025 to 9.04 in 2026, and perfect 10s fell below half for the first time, from 55% to 48%. Setup and complexity are the top complaint in both years, named in 25% of 2026 dislikes and peaking for support agents (34%) and SDRs (31%). Praise for integration, once the headline claim, fell from 30% of reviews to 21% — table stakes stop being mentioned.
Adoption is the sharpest number. The average share of intended users who had fully adopted their agent fell from 68% in 2024 to 55% in 2026, and nearly three in ten reviewers say 30% or fewer of the people meant to use it actually do. Roughly a fifth of reviews of agent builders and customer-support agents complain about cost, against 9% for business-operations agents, with several describing usage-based pricing that is hard to forecast.
Caveats: G2's sample is self-selected vendor reviews, weighted toward products that ask for reviews; the likes and dislikes shares come from keyword tagging and should be read as directional rather than exact. But the direction is consistent across categories, and it matches what deployment teams report anecdotally: the hard part of enterprise agents is no longer the model, it is the handover, the audit trail, and the person who has to sign off when it is wrong.
