A new G2 survey puts a number on something that's been mostly anecdotal: the gap between how many AI agents companies run and how well they can actually see them. Among B2B organizations with agents in production, 38% operate more than 100 of them. But only 27% maintain a central agent registry, just 18% use automated discovery to find sanctioned and unsanctioned agents, and a mere 14% have real-time observability into unusual behavior, costs, or risks. G2 calls this the "Distrust Tax" — the cost of running agents under constrained trust rather than full confidence. The numbers back up the framing: 31% of buyers say employees spend at least six hours a week "bot-sitting," and 45% report losing more than a quarter of expected productivity gains to oversight and rework. 46% say agent operating costs came in higher than expected. What's notable is where accountability actually sits. Ownership is still mostly informal — 47% say agent ownership and controls remain department-led rather than centrally governed, versus 23% that are centrally managed. Yet central IT and AI/data/ML teams already carry 49% of agent operating costs and 60% of governance accountability, a mismatch between who pays for oversight and who controls the sprawl generating it. The report also shows vetting has gotten more rigorous before deployment: 52% of buyers require human quality review, 48% require security testing, and 41% need legal, risk, or compliance review before an agent goes live. After deployment, 77% re-evaluate agents at least monthly. That's a meaningful amount of process — but it's manual gatekeeping layered on top of weak registries and near-absent real-time monitoring, not a substitute for either. This matters for the broader infrastructure conversation because it quantifies the thing every identity and attestation proposal is trying to fix. Protocols for agent IDs, signed receipts, and audit logging only solve a problem organizations have already demonstrated they can't solve with policy alone: most don't know how many agents they're running, who owns them, or what they're doing right now. The 14% real-time observability figure in particular suggests that even companies with mature review processes are flying blind between reviews — exactly the window where an agent could misbehave for weeks before the next monthly check catches it.