The industry has opinions about agent reliability. We're building the data.
Every CARA audit contributes anonymized, aggregated results to the CARA Benchmark — the emerging reference for how production agents actually perform, by vertical and by dimension.
What the benchmark collects
Scores by dimension and vertical. Finding classes and frequencies. Remediation velocity between retests. All of it aggregated, all of it anonymized.
What it never collects
Identities. Transcripts. Prompts. Anything reconstructable. Founding clients are compensated for contributing — that's what founding pricing is.
The reading room
The external evidence, curated and cited — the best-organized collection of agent-failure data we could assemble. Sources linked; check everything.
40%+ of agentic AI projects canceled by end-2027; one-third of companies will harm CX in 2026 via premature AI; ~130 of thousands of “agentic” vendors are genuine.
SOURCE ↗ MIT NANDA · 202595% of enterprise GenAI pilots deliver no measurable P&L impact. Externally built tools succeed roughly twice as often as internal builds.
SOURCE ↗ IBM IBV · 202654 AI agent incidents per organization last year. Two-thirds of tech CxOs accountable for AI they don't fully control. 11% feel ready.
SOURCE ↗ LANGCHAIN · 2026Quality is the #1 blocker to production (32%). 89% run observability; only 52% run evals. Human review remains the most-used method for high-stakes judgment.
SOURCE ↗ EY · 202564% of large enterprises lost over $1M to AI-related risks; average loss $4.4M.
SOURCE ↗ AIR CANADA PRECEDENT · 2024A tribunal held the airline liable for its chatbot's incorrect statements — the legal anchor for agent accountability.
SOURCE ↗
Get the first report when it publishes.
You're on the list. One email when it publishes — that's the whole sequence.
One email. No sequence. We'd rather impress you once than drip on you weekly.