A study of over 500,000 software developers found that autonomous coding agents increased individual code commits by 240%. At the project level, the gain dropped to 80%. At the release level — finished software that actually ships — it was 30%. Every metric on the dashboard was green. The organization was producing a fraction of what the activity numbers implied.
A completion metric tells you what happened inside the task boundary. It tells you nothing about where the unfinished residue went — to another team's queue, to a reconciliation process, to the customer, to next quarter.
A study of enterprise agents found that among runs passing every final-output check, 92% still had process deviations: wrong tools selected, steps taken out of order, scope boundaries crossed. The output looked fine. The organizational path that produced it was not what anyone assumed.
The information that would reveal whether an organization is actually adapting well to its agents — whether people can still intervene meaningfully, whether exceptions are concentrating in predictable ways, whether recovery capacity is quietly eroding — tends not to be metric-shaped. It lives in the texture of how work is actually flowing, which is precisely what dashboards are designed to abstract away.
The adoption-impact gap: 69% of firms report using AI; 89% report no measurable productivity effect over three years. Tracking adoption and tracking impact require different instruments.
Hidden process drift: Correct final answers can mask wrong tools, wrong sequencing, scope violations — deviations that compound over time but never trigger an alert.
Shifted difficulty: When agents handle easy cases, remaining human work gets harder. One telecom study found human call length rose even as automation metrics improved. The agents looked efficient; the workforce absorbed the complexity.
Nominal oversight: If stopping an agent means stopping a business process, a reviewer has formal authority but no real alternative. Human-in-the-loop rates can look healthy while the review itself becomes perfunctory.
75% of production agents lack formal benchmarks. Most organizations evaluate agent output by having people look at it — which works until agent volume, speed, or complexity outpaces the people watching.