“Every single workflow your team builds creates a small permanent job to maintain. Those jobs accumulate over time.” Kevin Indig and Amanda Johnson wrote that in Growth Memo this month, and the rest of their piece is an inventory of jobs nobody put on an org chart.
The dashboard that justified the AI budget measures the other side of the ledger. It counts the hours the model took off a task. It has no column for the hours the same model handed back: the hour spent checking a draft, the afternoon spent learning a new tool, the recurring Friday spent patching a pipeline someone built months ago. Those hours still get worked. They are booked against a person instead of against the project, so the project looks like a win and the person looks slow.
The same shape shows up in marketing, in finance and in analytics this month, in three pieces written from three different desks.
The hours marketing is not counting
Growth Memo aggregates numbers from other studies, so each one below belongs to its original author. I have not gone back to the primary sources here, and the piece is Part 1 of 2.
BetterUp Labs and Stanford surveyed 1,150 US full-time workers. 41% had received “workslop” in the previous month, AI output that passes as finished work. Sorting out each piece took an average of 1 hour and 56 minutes. At a 10,000-person company, the figure lands “past $9M a year.”
Workday’s figure is the cleanest statement of the thesis: for every 10 hours AI saves, about 4 are handed back in fixing and rewriting. Upwork asked 2,500 leaders and workers where the added workload comes from. Among employees who say AI increased their load, 39% point to checking and fixing output, 23% to learning the tools, and 21% to being handed more work.
HubSpot reports that 91% of marketing leaders say their teams use AI, and 66% say the company builds its own internal AI tools for marketing. Two thirds of those leaders are describing, in effect, an internal software practice. How many of those practices have a maintenance budget, an on-call rotation, or a person whose job description says “keep the thing running”? None of the sources asks, and I would expect the answer to be few.
Then the developer number, which Growth Memo borrows from METR’s late 2025 study: 16 experienced developers, 246 real tasks. The developers expected to be 24% faster. Measured, they were 19% slower. If the people whose profession is building tools lost time building with AI, a marketing team building its own tooling on the side should expect at least the same. Growth Memo’s summary: “Building your own in-house AI tools is addition, not necessarily efficiency.”
The finance pipeline with one owner
Brandon Sovran gives the accumulation a name. “Slop-creep is the gradual accumulation of systems, abstractions, features, and infrastructure that never needed to exist, made possible because agents have made them cheap enough to build before anyone seriously asks why.”
His mechanism is friction, or its absence. “Implementation used to have enough friction that stupid ideas often died before someone spent three weeks building them. Now you can just build the stupid idea.” The cost that used to kill a bad workflow before it existed is gone. What replaces it is a cost that arrives later, spread thin, paid by someone else.
His finance scenario is the one I would put in front of a CFO. A pipeline gets built outside Git. The person who built it leaves. The successor inherits a folder of scripts without understanding them. When something breaks, the successor asks an agent to patch it, and each patch takes the pipeline further from anything the rest of the company would recognize as the golden path. My addition to his scenario: no single patch is expensive enough to trigger a rebuild conversation, which is exactly why the rebuild conversation never happens.
Sovran cites a Microsoft Research and Carnegie Mellon study of 319 knowledge workers across 936 real-world uses: “higher confidence in the AI was associated with less critical-thinking effort.” It is one study, and I quote it as he reports it. It does explain why the successor accepts the patch. The agent sounds sure, so the human stops checking, and the pipeline drifts one more step.
The volume trap in analytics
Benn Stancil quotes Anthropic’s own account: “95% of business analytics queries are automated via Claude, with ~95% accuracy in aggregate,” and “the result is a 10-20x increase in the number of questions people ask.” These are vendor claims, relayed by Stancil, and should be read as such.
Take them at face value and do his arithmetic. A team that used to field 5 questions a day now fields 100. The model answers 95. The analyst still answers 5. The analyst’s queue did not shrink. What grew is the pile of 95 answers that someone in the business is now acting on, at roughly 95% aggregate accuracy, which at that volume works out to about 5 wrong answers a day that nobody is assigned to catch (derived from Stancil’s numbers, not a figure he reports).
Analytics did not lose its hard questions. It gained an unowned QA function.
Three functions, one shape
Marketing built tools and inherited maintenance. Finance built a pipeline and inherited a stranger’s scripts. Analytics automated the easy questions and inherited a volume of answers with no reviewer. In each case the work that disappeared from the dashboard reappeared somewhere the dashboard does not look, attached to one person, with no budget line.
We have written about pieces of this before. Agent-created code that nobody maintains covers the engineering version. The workflow as the unit of value covers the majority who never redesigned a workflow at all. A skills library with no garbage collector covers rot when a shared asset has no owner, and slop as an attention cost covers what the flood does to the reader. This piece is about the accounting. The removed hours are measured. The moved hours are not, and unmeasured work always lands on whoever is closest to it.
Do this now
Four steps, in order, and the first one is small.
Inventory the internal AI workflows. Every prompt chain, agent, script or pipeline someone built to do part of their job. HubSpot’s 66% says most marketing teams have at least one. Expect finance and analytics to have more than they list. List them the way you would list SaaS subscriptions.
Name one owner per workflow. Not the team. A person. If the answer is “the person who left,” you have found Sovran’s scenario in your own building, and it should go to the top of the list.
Put maintenance hours on a line item. Workday’s ratio is a defensible starting estimate: for every 10 hours a workflow claims to save, budget about 4 for checking, fixing and relearning. If the owner reports fewer, good. If the owner reports none, nobody is checking.
Install a buy-versus-build gate before the third week of fixing. Sovran’s three weeks is where friction used to kill bad ideas. Recreate it deliberately. Any homebrew workflow that has needed fixing in three separate weeks goes to a review: keep it and fund it, replace it with something someone else maintains, or delete it.
Growth Memo puts the cost of skipping this in marketing in one line: “the 9 months of citations, mentions, and reviews that were not earned this year cannot be backdated.” The hours a team spent maintaining its own tools were hours the market did not see. That is the invoice for the work that moved.
This analysis synthesizes The AI hours nobody on your marketing team is counting (Growth Memo, Kevin Indig and Amanda Johnson, September 2026), Slop-creep: when building gets cheaper than thinking (Awaiting Input, Brandon Sovran, September 2026), and How to find insights (Benn Stancil, September 2026).
Victorino Group helps operating teams inventory their internal AI workflows, assign owners, and budget the maintenance the dashboard is not counting. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation