The deployment funnel nobody talks about: 60% evaluate, 20% pilot, 5% ship. MIT tracked 300 real AI implementations against profit metrics.

Late 2025, MIT researchers measured something the industry had avoided looking at directly. Not projections or pilot numbers. Documented outcomes from 300 AI deployments in real businesses, tracked against profit metrics. The funnel breaks down like this. Sixty percent of companies evaluated AI tools. Of those, twenty percent ran a pilot. Of those pilots, only 5% reached full production deployment on the service line. Ninety-five percent of AI investment dissolved before it produced a measurable outcome. The companies that made it to production had a clear pattern. They didn't ask AI to substitute for judgment. They identified bounded tasks: specific inputs, defined outputs, failure modes that were contained. They measured success criteria before deployment, not after. Content drafting. Code review. Data summarisation at volume. The 95% that didn't make it: haste, no defined success metrics, and the assumption that efficiency gains would be obvious once the tool was in the workflow. There's a line from the research worth sitting with. "We replaced X employees with AI" isn't an efficiency metric. It's a headcount metric. Those are not the same thing. Klarna is already in the reversal phase, rehiring humans after the AI efficiency numbers didn't hold up at scale. What's the clearest signal you've found for whether a deployment is actually working, before it's too late to course-correct?