50% of AI agents fail in prod after passing tests
50% of firms have shipped agents that passed tests but failed customers, with only 5% fully trusting automated evals as 66% move to zero-human deployments.
venturebeat pulse research of 157 enterprises reveals a widening 'evaluation gap' where 50% of organizations have shipped agents that passed internal tests and then failed in production. trust in these tests is low, with only 5% fully trusting automated evaluations and 29% citing poor alignment with real-world outcomes as the primary weakness. despite these reliability risks, 66% of companies are actively pushing toward fully autonomous, zero-human-in-the-loop deployment pipelines. this data suggests that enterprise ai teams are prioritizing shipping speed and autonomy over robust assurance, risking a wave of production failures as agent reliance outpaces evaluation maturity.