Someone Has to Decide What the Machine Stops Making. Change Futurist | September 24, 2026
Most marketing teams can tell you how many assets they shipped last week. Far fewer can tell you which of those assets a customer would have missed.
On Monday I wrote about the handoffs between workflow steps and how they cap the return on marketing AI. Clear the handoffs and output goes up. Four pieces published in the last two days deal with what happens after that, which is that the extra output carries a cost on the customer side.
Emily Weiss at Gartner has survey data on how consumers react to the volume of AI content. Sir John Hegarty, interviewed by Josh Stephenson at Marketing Week, blames the decline of creative work on people who treat data as an answer. In HBR, David L. Rogers and Krishnan Sankaranarayanan examine why AI spending has failed to show up in financial results. Kyrsten Halley, also in Marketing Week, describes what happens to a marketer who walks into a finance conversation carrying a single number.
Across all four sits a decision that in most marketing organizations belongs to no one: what to stop producing.
Count volume as a cost against brand trust
Weiss published a Gartner Q&A on September 22 drawn from a survey of 1,006 US consumers conducted in May and June. Sixty-five percent of respondents said brands produce too much AI-generated content in their marketing. Fifty-seven percent said that content has made them less trusting of brand messaging overall.
The effect reaches past owned channels. Thirty-five percent said they rely on influencers less for shopping information because of AI, and 43 percent said they rely more on real people. Weiss tells marketing leaders to point AI at relevance, creativity, and customer experience instead of raw output, and to write explicit AI rules into brand-influencer agreements.
Look closely at what the 57 percent figure covers. Those consumers report lower trust in brand messaging as a category, not lower trust in the specific brands that flooded their feeds. Your team absorbs part of the cost of a competitor's production decisions.
Economists call the mechanism costly signaling. Consumers have long read production effort as evidence that a brand means what it says. Produce a hundred variants for the cost of one and the effort behind any single variant stops carrying information. The 43 percent leaning on real people are looking for effort they can still verify.
It reminds me of what Toby Coulthard, Chief Product & Growth Officer at Jacquard, said when I interviewed him on The Agile Brand podcast: "We can all kind of feel AI-generated content when you see it... There's this convergence of different people sounding the same."
Commentary: Weiss supplies the consumer-side evidence for capping production. Put the 57 percent figure in front of anyone who argues for doubling output on the grounds that the marginal asset costs nothing.
Assign taste to a named person
Stephenson's interview with Hegarty ran on September 23, ahead of Hegarty's keynote at the Festival of Marketing in October. Hegarty told the same publication nearly a decade ago that creativity was receding from the profession. He says it has gotten worse since, and he blames organizations run by people who read a spreadsheet and assume it contains an answer. He states it plainly: "Data doesn't have the answer. Data is knowledge."
He backs the argument with a rough figure. Something like 95 percent of new product launches fail, by his estimate, and the rate has held for 20 years while marketers acquired far more tools and data. He also argues that everyone who knows about a brand helps build it, including people who will never buy, which puts him at odds with teams that trim reach down to the efficient audience. He wants marketers to write a manifesto stating what the brand believes, and to treat taste as something the organization defends.
I draw a version of this distinction in the framework I am building on top of Coherence at Velocity. Intelligence tells you what is and what is likely. Deciding what the brand should say is a normative act, and it stays with a human who answers for it. Scoring which variant earns the most clicks is work a model does well. Whether the brand should run any of those variants is a separate question with a person's name attached to the answer.
The author test from my AI capability framework applies here too. Remove the AI and ask whether a skilled person could have produced the asset, slower and at smaller scale. If yes, the work sits in Augmentation, where a human author carries responsibility for quality. Most marketing content passes that test, so most marketing content needs an author with the standing to reject it.
It reminds me of what Jason Ippen, Vice President of Brand Strategy and Content at Georgia-Pacific, said when I interviewed him on The Agile Brand podcast: "Where humans stay essential is deciding whether an idea is right. I mean, taste really matters. Brand stewardship matters."
Commentary: Hegarty has made this argument for years from inside the agency business, largely on conviction. Weiss's respondents describe the effect he predicted, which gives marketing leaders an evidence base for the same position.
Tie each AI investment to a line on the income statement
Rogers, faculty at Columbia Business School, and Sankaranarayanan, who leads AI transformation at Eastman, published Closing the Gap Between AI Investment and Financial Return in HBR on September 23. They identify two low-value patterns that consume AI budget: features customers do not care about, and employee time savings the organization never captures as measurable savings.
Their four corrective actions are to define your value drivers, focus on your position in the industry value chain, find opportunities in your income statement, and define your measures of financial return. Firms reporting measurable value from AI, in their account, aimed it at problems central to strategy and profitability.
Content volume lands in both patterns. Hours saved per asset stay on the team's calendar unless someone reallocates them, so finance records nothing. The additional assets rarely trace back to a value driver anyone wrote down.
In my latest book, Marketing Operations 3.0, I describe this as a change in how leaders evaluate the function. They will measure performance less by the volume of campaigns launched and more by the quality, governance, and business value of the work the team supervises. Rogers and Sankaranarayanan put the same test in finance terms.
Commentary: Most marketing AI business cases I review skip straight past the first of these four actions. Without a named value driver, finance has no way to score the return, which is one reason these programs stall at renewal.
Bring your CFO a range with a downside case
Halley, a former Aldi marketing director, published Stop learning to speak the language of the CFO on September 22. Marketers who collect finance vocabulary without learning how finance teams think, she argues, come apart at the first follow-up question. Her commercial finance colleagues taught her to present a range people could stand behind, and to state a confidence percentage in place of false certainty. As she puts it, a number with no reasoning behind it is "just a guess wearing a suit."
Run that standard against an AI content program. The usual business case shows cost per asset falling, output rising, and one projected revenue lift. Weiss's numbers supply the missing downside: a trust decline that reduces what each additional asset returns. Expect that question from any CFO who reads the same research, and bring the answer to the first meeting.
Commentary: Halley is writing about boardroom credibility in general. Her point about ranges carries particular force for AI programs, where the second-order effects on trust resist precise measurement for several quarters.
Run the numbers on a content program you recognize
An $800 million specialty home goods retailer runs email as its largest owned channel, with 2 million active subscribers, each worth about $38 a year in attributable revenue. Before generative AI, the team of 25 sent 12 campaigns a month. Revenue per send averaged $210,000, for $2.52 million a month. Unsubscribes ran 0.05 percent per send, costing the list about 12,000 subscribers a month.
The team deploys AI production and doubles cadence to 24 sends a month, with more variants per send. Revenue per send falls to $120,000, and monthly email revenue rises to $2.88 million. The dashboard shows a gain of $360,000 a month, and leadership praises the program in the quarterly review.
Unsubscribes climb to 0.08 percent per send. At 24 sends, the list loses about 38,400 subscribers a month, an extra 26,400 against the old baseline. At $38 each, every month of the new cadence removes roughly $1 million in annual revenue capacity. After one quarter the team has booked $1.08 million in added email revenue and permanently lost about $3 million a year in subscriber value.
No one on that team made an obviously bad call. They tracked monthly revenue because monthly revenue sat on the dashboard, and the question of what to stop sending had no owner.
So the retailer names a brand editor with authority to cut any send that fails a written standard. The team drops to 14 sends a month and retires the lowest-performing variants. Marketing operations starts reporting subscriber value lost alongside revenue gained. When the CMO takes the revised plan to finance, she brings a range: revenue per send recovering to between $180,000 and $200,000, list erosion between 13,000 and 16,000 a month, and a stated confidence level on each.
The AI tools stay in place. The team aims them at 14 sends a month in place of 24.
Take these four steps into 2027 planning
Name the person who decides what stops. Give one senior person authority to cut output that fails a written brand standard, and publish the standard so teams know what fails before they build it.
Put the downside case on the dashboard. Report subscriber loss, trust metrics, and engagement decay next to output and revenue for every AI-scaled channel. Use the Gartner figures as your starting assumption for the size of the effect.
Map each AI content program to a value driver. If a program cannot name the income statement line it moves, pause it until someone can.
Forecast in ranges with stated confidence. Take Halley's advice into your 2027 budget. A range with a downside case survives the CFO's second question.
I majored in photography in college, back when that meant film and a darkroom. A roll gave you 36 frames. You developed it, printed a contact sheet, and went over every frame with a loupe and a grease pencil. On a good roll I printed two. Deciding which two took longer than the shooting did, and I have never found a tool that does that part for you as well as a human can do it.
Go back to last week's list of assets. Whoever has the authority to cut it should have their name is on the standard.
Featured this cycle:
Gartner Marketing Survey Finds 35% of Consumers Rely on Influencers Less Due to AI. Q&A with Emily Weiss, Gartner, September 22, 2026
'Repetitive, cliché, boring': Sir John Hegarty on marketing's failure of creativity. Josh Stephenson, Marketing Week, September 23, 2026
Closing the Gap Between AI Investment and Financial Return. David L. Rogers and Krishnan Sankaranarayanan, Harvard Business Review, September 23, 2026
Stop learning to speak the language of the CFO. Kyrsten Halley, Marketing Week, September 22, 2026