A new Optimizely study puts a number on something marketing teams have been feeling for two years: the time AI saves on the front end is showing up as a bill on the back end. In a survey of 2,003 marketing leaders across seven countries conducted by Savanta between May and June 2026, 76% said they spend at least three hours a week editing, fact-checking, or correcting AI-generated output. Just 4% said AI saves them time at every stage of the process. Fact-checking and hallucination review were the single biggest driver of that extra work, cited by 48% of respondents, ahead of the 40% who pointed to time lost moving information between disconnected AI tools.
The pattern matters more than any single statistic in it. Marketing organizations adopted generative AI on the promise that it would compress content production, freeing up time for strategy. Instead, as AI outputs multiply across channels, teams are absorbing a new category of labor: reviewing, correcting, and standardizing machine output before it goes out under the brand’s name. The study found only 19% of marketers work from a single integrated AI platform, meaning most are stitching output together from multiple disconnected tools, which compounds the review burden. More than half said their tools capture facts but miss the brand’s emotional tone, and 54% said leadership underestimates how much human effort AI actually requires to use safely.
The original insight buried in the data is a gap between the C-suite and the people doing the work: 69% of C-suite leaders told Optimizely their AI adoption is fully aligned with strategy, versus just 27% of individual contributors and analysts who agreed. That gap is a governance problem as much as a productivity one, and it lines up with the broader push toward guardrails on what AI agents are allowed to do unsupervised. Sixty-five percent of marketing leaders said they would consider pausing their AI rollout for 90 days to rethink their approach, a signal that the “move fast” phase of generative AI adoption in marketing is giving way to a harder conversation about process, oversight, and what AI output actually needs to be trusted at scale.
Source: Optimizely
