Kevin Indig gathers five studies showing where AI actually costs marketing teams time, and it points to a growing rebuild-and-babysit tax

Kevin Indig published a column in Search Engine Land on 9 September titled "The AI hours nobody on your marketing team is counting." It gathers findings from five studies that separately measure where AI in a working team ends up costing hours: building custom tools, fixing model output, and checking whether a draft is right. The point of the piece is that these hours are real and mostly invisible on a team's dashboard.

What the five studies actually say

The clearest number comes from a METR study Indig cites, published in July 2025. METR paired 16 experienced open-source developers with AI tools on 246 real tasks in their own repositories. The developers expected AI would make them 24 percent faster. Measured, they were 19 percent slower.

HubSpot's State of AI report tells Indig where the workload has moved: 91 percent of marketing leaders say their teams use AI, and 66 percent say their company builds its own AI tools for marketing.

BetterUp Labs and Stanford surveyed 1,150 US knowledge workers on "workslop", a term for AI-generated work that a colleague then has to redo. 41 percent said they received workslop in the previous month. Sorting each one out took a mean of one hour and 56 minutes. BetterUp estimates the annual cost at 9 million dollars for a 10,000-person firm.

A Workday study, cited via IT Brief, puts a matching figure on the other side of the trade: for every 10 hours AI saves, teams hand back roughly 4 hours fixing and rewriting.

Upwork surveyed 2,500 leaders and workers. 39 percent named time spent checking and fixing AI output as their top new task; 23 percent named time spent learning tools. Both are pure overhead.

The pattern under the numbers

Indig's argument is that AI moves work rather than removing it. The hours travel from doing the task to building, learning, and maintaining a system that helps someone else do it. On a marketing team, that shows up as engineers who used to earn links now building an internal keyword tool, a writer who once shipped four posts now editing eight drafts, and a director whose weekly review has become a review of the review.

Two shapes of hidden cost show up across the five studies. One is the rebuild loop: teams standing up custom AI workflows that cover a shrinking slice of the job as public tools catch up, then tearing them down. The other is the babysitter tax: humans checking, correcting, and rerouting AI output, work that shows up as headcount on a Slack channel and nowhere in the plan.

Indig also names a market shift underneath the productivity question. Google's AI Overviews, ChatGPT, Perplexity and Claude answer many of the queries that used to send readers to a website. So the same team, spending more time on AI-adjacent work, delivers a smaller organic result. The AI hours become expensive because they compete with the hours that used to build brand, links, and audience.

None of the five studies has a fix wrapped inside it. The one habit that falls out is measurement. If a team is going to hand a task to a model, it needs to count the setup time, the review time, and the redo time next to the raw hours saved, or the savings claim is doing the same work as the "20 percent faster" claim the METR developers made about themselves before they were timed.

Source

Search Engine Land: The AI hours nobody on your marketing team is counting by Kevin Indig, 9 September 2026.