In April 2026, Stanford published the ninth AI Index report, the most comprehensive independent measurement of the field. Buried in its economy and responsible-AI chapters is the cleanest description we have seen of the thing we started Ripcord to fix. We call it the standoff: everyone bought autonomy, and almost nobody dares to use it. Here it is in three numbers.
of surveyed organizations now use AI, and 70% run generative AI in at least one business function. Adoption is not the bottleneck.
Scaled deployment of AI agents, the systems that actually act, sits in the single digits of respondents across nearly all business functions. In most functions, a majority report no agent use at all.
of respondents name security and risk concerns as the primary obstacle to scaling agentic AI, far ahead of technical limitations (38%), regulatory uncertainty (38%), and gaps in responsible-AI tooling and control (36%).
Source: McKinsey and AI Index survey of business leaders, via the Stanford AI Index Report 2026, chapters 3 and 4.
Read that third number again. When companies explain why agents are stuck in pilots, the top answer is not "the models aren't good enough." Only 38% say that. The top answer, by a wide margin, is fear. That is not a model gap. That is a missing control layer.
The same report explains why the fear is earned. On τ-bench, the benchmark that tests agents on following policy constraints while using tools, no frontier model exceeds 71%. On OSWorld, agents complete 66.3% of real computer tasks, within six points of the 72.35% human baseline, which means agents now fail about as often as a junior employee. The difference is that the junior employee works inside approvals, spending limits, and an audit trail, at human speed. The agent works without any of that, at machine speed.
And the incidents are arriving on schedule: documented AI incidents rose 55% in one year, from 233 in 2024 to 362 in 2025, after staying under 100 a year until 2022 (AI Incident Database). The OECD's separate monitor peaked at 435 in January 2026 alone. Both trackers skew toward high-visibility, English-language cases, so treat those counts as floors. Nine of the worst, fully sourced, are in our incident file.
Yolanda Gil and Raymond Perrault, co-chairs, Stanford AI Index Report 2026.
The standoff feels safe because nothing visibly breaks. But the St. Louis Fed measured what AI adopters are banking: 2.2 hours per employee per week on average, with the top 27% of users saving 9 or more and power users reporting 20 plus. For a 500-person company at a $75 loaded hourly rate, waiting costs millions a year, silently. Meanwhile the organizations that did deploy are absorbing hits with falling confidence: among organizations reporting AI incidents, the share hit 3 to 5 times in a year rose from 30% to 50%, while "excellent" self-rated incident response fell from 28% to 18%.
So both doors are expensive. Deploy blind and eat the incidents, or wait and eat the productivity loss. The standoff persists because most companies believe those are the only two doors.
The way out is not a better model. Model capability has converged: Stanford puts the top four frontier models within 25 Elo points of each other, and concludes that competitive pressure is shifting toward reliability. The way out is the thing 62% of respondents are actually asking for: a control layer that makes a 70%-reliable agent deployable today. Score every action before it runs. Route the risky ones to a named human. Keep the destructive ones reversible, so being wrong once is survivable. Freeze the agent that argues with a rejection. Produce the evidence trail your auditor and, eventually, your insurer will ask for.
That is what we built. The budget for it already exists, too: among organizations with at least $30 billion in revenue, 41% expected to spend $25 million or more operationalizing responsible AI, and 22% budgeted $50 million plus. The standoff will not be broken by the next model release. It will be broken by the first company in each industry that can say yes safely, and bank the hours its competitors are still debating.
The calculator on our homepage models what the standoff costs your company per year, unprotected, with the assumptions in the open.
Run your number →