The study, conducted across 47 enterprise AI agent deployments spanning customer service, sales automation, research synthesis, and internal operations, tracked agent behavior over a combined 180,000 task executions. The headline finding "',"Å",¾ that 73% of task executions contained at least one error, redundant step, or hallucinated context element "',"Å",¾ is consistent with what practitioners have been reporting anecdotally for the past 18 months.

The error taxonomy from the study is revealing. The most common failure mode wasn't catastrophic hallucination "',"Å",¾ it was subtle contextual drift: agents that started a task with the right information but gradually incorporated incorrect assumptions as the task progressed, eventually producing outputs that were plausible in structure but wrong in specific details. These failures are hard to detect without explicit verification steps, and most production deployments don't include those steps.

Why Testing Environments Mislead

Benchmarks and pilot environments systematically overstate agent reliability because they test on task distributions that are known to be tractable. The tasks that get selected for testing are the ones where the agent performs well "',"Å",¾ either because they're well-suited to the agent's capabilities, or because the testing team had enough context to set up the agent appropriately. The failure modes that emerge in production are disproportionately the ones that weren't anticipated during testing.

Error accumulation is the other structural problem. Multi-step agent tasks compound small error rates into large reliability problems. An agent that has a 95% reliability rate per step will successfully complete a 10-step task only about 60% of the time. Production agent deployments typically involve 15-30 steps for complex tasks. At those lengths, even high per-step accuracy produces failure rates that make the system unreliable without extensive human oversight.