Why your AI productivity numbers are probably wrong, in both directions

In short: self-reported AI time savings are unreliable in both directions. In one randomized study experienced developers felt 20 percent faster and were measured 19 percent slower, while a field study of support agents measured real gains concentrated among less experienced staff. Measure observed output and quality against a baseline, expect a possible dip while processes are redesigned, and report the complements that predict the recovery.^1,2,3

Most organizations measure AI productivity the easy way: they ask people. The answer is unreliable in both directions, and the research shows why.