Agentic AI Paper Shows 43-Point Gap

The uncomfortable takeaway is that AI productivity claims keep collapsing when the whole workflow is measured, not just the model output. A 43-point gap is not a rounding error; it's a calibration failure.

• –In the coding study, the time sink is prompting, waiting, and verification, which can erase any raw generation advantage.
• –The clinical documentation examples show this is not just a coding problem: vendor claims still outrun independent measurements.
• –Mature codebases and experienced engineers are the hardest test case, because tacit context and established workflows blunt the advantage AI is supposed to provide.
• –Teams should measure end-to-end cycle time, quality, and review burden, not just suggestion acceptance or self-reported speedups.
• –The real lesson is calibration: use AI where it genuinely shortens work, but don't assume the demo gain survives contact with production constraints.