Marketing teams have acquired a second job nobody budgeted for. They are building software. Prompt chains, automations, internal agents and the maintenance those things demand now sit inside teams whose headcount was approved for campaigns, and the hours come out of work that used to be the job.

That is the argument Kevin Indig made in a column published on his Growth Memo newsletter and republished by Search Engine Land on 9 September. Indig, a growth adviser who previously led SEO and growth at Shopify, G2 and Atlassian, calls it meta work: working on how to get work done. The claim is worth taking seriously, and the evidence behind its most-quoted supporting study is weaker than it was a year ago.

What does the column actually claim?

Indig’s case is that homebrew AI work is invisible in a way ordinary work is not. A campaign has a brief and a deadline. A half-built internal tool has neither, so the hours spent on it are absorbed rather than recorded, and the bill is paid from content production, digital PR, community presence and third-party review campaigns.

Two numbers support the exposure. HubSpot reports that 91% of marketing leaders say their teams use AI, and that 66% say their company builds its own internal AI tools for marketing. The second figure is the one that matters here: two thirds of companies are not simply buying AI features inside existing software, they are constructing things that then need maintaining.

Indig’s own example is a confession rather than a case study. Leading engineering at Shopify, he spent too much of his team’s capacity building tools the company could have bought, at the cost of higher-impact work. That is an anecdote from one company, offered as one, and it is the kind of evidence a reader should weigh accordingly.

The HubSpot figures carry their own interest. HubSpot sells marketing software, including AI features, and benefits from a narrative in which marketing teams adopt AI heavily. That does not make the numbers wrong, but a vendor survey on AI adoption is not a neutral instrument.

How strong is the evidence underneath it?

The study that anchors most arguments of this kind is METR’s randomised controlled trial, published on 10 July 2025. Sixteen experienced open-source developers completed 246 real tasks in repositories they already knew well, each task randomly assigned to allow or forbid AI tools. The developers predicted AI would make them 24% faster. Measured, they took 19% longer. Afterwards, they still believed they had been 20% faster.

That 39-point gap between perception and measurement is the finding worth carrying into a marketing context, and it is the one the column uses. The 19% figure itself has moved.

In February 2026 METR published new data and changed its experimental design, after finding that between 30% and 50% of invited developers declined to take part if it meant working without AI access. That is a selection effect: the original sample skewed towards developers willing to give AI up, who are plausibly the ones it helps least. The newer cohort, 57 developers across more than 800 tasks, measured a 4% slowdown with a confidence interval running from 15% slower to 9% faster, which does not exclude a real speed-up. METR’s own conclusion was that AI likely provides productivity benefits in early 2026. The organisation labels the 2025 result historical.

The original figure also carried a confidence interval of 2% to 39% slower, which is a wide band rarely mentioned by the people quoting the midpoint.

None of this rescues the “AI makes everyone faster” position, and none of it makes the column wrong. It does mean that a marketing lead who walks into a budget conversation armed with “a study proved AI makes people 19% slower” is quoting a number its own authors have superseded, and will deserve the correction they get.

There is a larger gap behind all of it. No controlled trial has measured AI’s effect on marketing productivity. The evidence base is developers, customer support and general knowledge work. Applying a developer result to a content team is an inference, and the column presents it as one.

What can a marketing team actually measure?

The most useful number here was not measured on developers. BetterUp Labs and Stanford surveyed 1,150 full-time US workers about what they term workslop, meaning AI output that looks finished and is not. In the previous month, 41% had received some. Each incident took an average of one hour and 56 minutes to resolve. At a company of 10,000 people, the researchers put the annual cost past $9 million.

The asymmetry is the point. The sender saves roughly 20 minutes. Somebody downstream spends two hours working out what is wrong with what they were sent. Both halves are real, only one is visible to the person who felt the time saving, and neither appears in a tool’s usage dashboard.

Two things follow that a team can act on this quarter. Log maintenance hours on internal AI tooling for one quarter as a named line, in the same way agency time or design time is logged, since the argument cannot be had without a number. And before building, price the commercial alternative honestly against total cost of ownership: licence fee against build time, prompt tuning, quality assurance, integration and the hours spent when it drifts.

No controlled trial of AI’s effect on marketing productivity has been published, so the strongest evidence available remains a developer study its own authors have since labelled historical.