Only around 15% of enterprise AI use cases generate measurable ROI. But the reasons behind that number are more nuanced than they first appear, and almost always rooted in something organisational rather than something technical. This piece looks at why, how to spot the warning signs, and how to think about doing better.

OpenAI recently published benchmark data testing AI models against real professionals, not junior staff, but people averaging 14 years of experience on tasks that typically take 7 hours to complete.

The headline result: Claude Opus 4.1 matched or beat those experts 47.6% of the time. For context, true parity would be 50%. A year ago, the best model managed 13.7%.

The pattern in the data is consistent. Where tasks have clear specifications and a defined output like financial reports, legal documents, or structured analysis, AI performs at expert level reliably. Where success depends on professional judgement, tacit knowledge, or navigating ambiguity, it doesn't.

The practical test is straightforward: could someone with no domain expertise complete this task from a written brief? If yes, AI will probably handle it well. If not, it probably won't.

Worth noting too: these results used off-the-shelf models with no fine-tuning. Custom solutions built around specific workflows tend to do better.