I have seen this movie before. The first time, it was called programmatic.
Around 2014 the pitch was identical to the one marketing teams are hearing about AI today. Automate the manual work, cut the waste, let the machine optimise. Efficiency was the headline on every deck.
Six years later, a PwC study commissioned by ISBA traced UK programmatic spend end to end and found publishers received about 51 percent of each advertiser pound. Around 15 percent could not be attributed to anyone at all.
The labour did not disappear, it moved
Greg Jarboe made this comparison explicitly in a piece for Search Engine Journal this week, and the parallel holds up better than I expected.
Programmatic did not remove work. It converted visible labour, people negotiating and trafficking insertion orders, into invisible labour: platform fees, verification vendors, data management and supply path audits. The work was still there, but it no longer appeared on a project plan.
AI is doing the same to content and campaign production. The visible labour, writing the first draft, shrinks. The invisible labour of prompting, checking, fixing and maintaining the system grows, and nobody budgets for it until someone asks why output tripled and results did not.
The numbers on that invisible layer are now public. Workday research found that for every 10 hours AI saves, roughly 4 are clawed back through redoing weak output, verifying results and learning the tools.
That is a 40 percent tax on the headline saving, before anyone has measured whether the output performed.
People cannot feel the slowdown
The more uncomfortable data point comes from METR. The research group gave 16 experienced developers 246 real tasks. They expected AI to make them about a quarter faster.
They finished roughly 20 percent slower. Afterwards, they still believed AI had helped.
That gap between felt speed and measured speed is the core problem for marketing leaders. Your team will tell you, sincerely, that AI makes them faster. They are reporting the drafting step, which genuinely is faster, and not the revision loop that follows it.
BetterUp Labs and Stanford put a price on that loop. Across more than 1,000 workers, each instance of fixing AI-generated "workslop" consumed close to two hours. At large organisations that compounds to more than $9 million a year.
I made a similar point when a 68 percent productivity claim did the rounds a few weeks ago. The vendor figure measured a task. Your business runs a process, and processes include the checking.
Why marketing is especially exposed
In software, bad AI output usually breaks something. A test fails, a build stops, and the cost surfaces. I wrote about how AI ships code fast and hides the cost, and even there the bill arrives eventually.
In marketing, bad output rarely breaks anything. A mediocre email still sends. A generic landing page still loads, and a blog post that says nothing still gets indexed.
The cost appears months later as lower conversion, weaker brand recall and a content library nobody trusts. By then it is impossible to attribute to the decision that caused it, which is exactly how programmatic waste hid for years.
There is also a seniority problem hiding in the maths. Drafting was usually junior work, and checking is usually senior work.
When AI multiplies the number of drafts, it multiplies the review load on the most expensive people in the department. Saving an hour of a junior copywriter's time while adding forty minutes of a creative director's time is not an efficiency gain, even though the hours column says it is.
I see this pattern in agency teams constantly. Heads of department become full-time editors of machine output, and the strategic work they were hired for quietly stops happening, which is a cost no timesheet will ever show.
There is a structural twist too. HubSpot data shows most marketing teams now use AI, and a majority of companies build internal AI tools rather than buying them. Every internal tool is a system someone has to maintain, and that person is almost never in the efficiency calculation.
What an honest AI business case looks like
None of this argues against using AI in marketing. We use it heavily at difrnt., and the gains are real when the whole process is measured rather than the flattering step.
The fix is accounting, not abstinence. Before claiming a saving, count four things: drafting time saved, review and correction time added, tool and prompt maintenance time, and the performance of the output against the human baseline.
Then measure at the process level. Time from brief to live campaign, not time from prompt to draft. Revenue per asset shipped, not assets shipped per week.
Finally, give the invisible work an owner. If nobody is responsible for prompt libraries, evaluation checks and model updates, that work happens anyway, just badly and unbudgeted, spread across people who think they are doing something else.
Programmatic taught the industry that a cheaper unit cost can coexist with a more expensive outcome. The teams that remember that lesson will get the AI gains. Everyone else will get more output and a very confident explanation for why it did not work.