Jesse Jacoby
Follow a single AI-drafted document through an ordinary company. A program manager asks an assistant to summarize three vendor proposals ahead of a steering committee. The summary is clean, well organized, and confident. It also carries forward a pricing assumption from last year’s proposal, because that version was sitting in the same folder. Nobody catches it. The summary goes into the deck, the deck goes into the shared drive, and two months later a line from it appears in the business case.
Six months after that, a finance analyst asks the company’s internal AI assistant what the vendor contract is expected to cost. The assistant retrieves the summary, because it is the most recent and best-written document on the subject, and answers with complete confidence. The error now has a citation.
That is the part of the workslop problem most leaders have not priced.
The Cost of Catching It Is the Small Number
The term workslop, coined by researchers at BetterUp Labs and the Stanford Social Media Lab, describes AI-generated work that looks finished but lacks the substance to move the task forward. Their research put a price on the rework: close to two hours per instance, and millions of dollars a year at enterprise scale.1 That figure got the attention, and it deserved it.
But it measures only the workslop someone noticed. In the same research, workers estimated that about 15 percent of the work they receive is workslop, and 53 percent admitted that at least some of what they send probably qualifies. Recognized workslop costs an afternoon. Unrecognized workslop gets saved, forwarded, cited, and built upon. The rework hour is the visible cost. The expensive cost is everything that passed inspection because it looked right.
Debt, Not a Tax
When Ward Cunningham introduced the idea of technical debt in 1992, he described shipping imperfect code as borrowing against the future.2 A little debt speeds delivery, as long as it is paid back. Left alone, it accrues interest, and eventually the interest consumes the team’s capacity to do anything new.
Workslop behaves the same way. A tax is paid once, at the moment of the transaction. Debt sits on the balance sheet and grows. Every unchecked deliverable is a small loan against the organization’s future attention, and the interest payments show up in places nobody connects back to the original document: the meeting spent relitigating a number, the decision made on a summary that misstated the risk, the new hire who learns a process from a generated procedure document that describes a workflow nobody actually follows.
The difference matters for how leaders manage it. A tax can be estimated and absorbed. Debt has to be tracked, serviced, and paid down, or it eventually sets the terms for everything else.
The System of Record Is Where the Interest Compounds
Two years ago, a mediocre document in a shared drive mostly sat there. Today it is input. Enterprise AI assistants answer questions by retrieving from the organization’s own repositories, wikis, ticket histories, and meeting notes. Whatever gets filed today becomes the context for tomorrow’s answers, and generated content is often the most fluent and recent material available, which is exactly what retrieval favors.
The research community has a name for the extreme version of this loop. A 2024 study in Nature found that AI models trained repeatedly on content produced by earlier models progressively lose the rare, unusual parts of the original data, a process the authors called model collapse.3 An enterprise knowledge base is not a training run, but the organizational parallel is close enough to take seriously. The hard-won caveat an expert wrote into a procedure, the exception that only applies to one region, the reason a policy exists at all: those are the tails. Summaries of summaries average them away.
Gartner reported that 63 percent of organizations either lack or are unsure whether they have the data management practices AI requires, and predicted that through 2026 organizations will abandon 60 percent of AI projects that are not supported by AI-ready data.4 Most leaders think of data readiness as a legacy problem: old systems, inconsistent fields, missing records. Workslop is a new source of unready data, and the organization is manufacturing it itself, at scale, every day.
The Trust Interest Rate
The debt also accrues in relationships. In the BetterUp and Stanford research, roughly half of recipients said they viewed the sender of workslop as less creative, capable, and reliable, 42 percent saw the sender as less trustworthy, and nearly one in three said they were less likely to want to work with that person again.1
The downstream effect is a verification tax that spreads beyond the original offense. Once a colleague has sent work that did not hold up, everything from that colleague gets checked, including the good work. Multiply that across a team and collaboration slows to the speed of mutual audit.
The direction of travel matters too. The research found that most workslop moves between peers, but 16 percent flows down from leaders to their teams. That share is the most dangerous, because it arrives with authority attached and is the least likely to be challenged. A director’s AI-drafted strategy memo with a flawed premise does not get returned with comments. It gets executed.
Why Nobody Is Keeping the Ledger
Workslop goes untracked for a structural reason: the person who creates it captures the benefit immediately, and the cost lands on someone else, later, in a different budget. The sender saves an hour today. The recipient, the next team, or the next quarter’s decision pays for it. Nothing in the typical operating model connects the two.
Most AI programs make this worse without meaning to. Adoption targets reward volume. Mandates to use the tools arrive without standards for what acceptable output looks like, a gap the original researchers flagged directly. And the quality of the organization’s knowledge assets is rarely anyone’s named responsibility. Documents have authors and owners in theory, and in practice they have neither once they are filed.
Paying It Down
Retiring this debt does not require slowing AI adoption. It requires four disciplines that most organizations can put in place within a quarter.
Make authorship non-transferable. Whoever sends the work owns it, regardless of what drafted it. The standard is simple enough to state in one sentence: do not send anything you would not sign. Teams that adopt a habit of noting what was checked, and what was not, give recipients the information they need to calibrate trust.
Separate drafts from records. Not everything generated deserves a place in the repositories that AI assistants retrieve from. Define which document types are records, such as policies, procedures, decision logs, and contract summaries, and require a named human review before anything is promoted into them. Label provenance so future readers, human or machine, can tell the difference.
Make returning workslop normal. Most recipients quietly fix what they receive, which hides the debt from the people creating it. A standard, unembarrassing reply, along the lines of “not ready yet, here is what is missing,” turns invisible rework into visible feedback and gives senders a reason to change.
Audit the ledger. Once a quarter, sample the twenty most-retrieved documents in your core repositories and check them for accuracy, currency, and ownership. Retire what is dead. Correct what is wrong. Track how often a decision had to be revisited because the material behind it did not hold up. That number is the interest rate, and it is worth knowing.
Know What Your Organization Believes
Every organization runs on what it believes to be true: what a contract says, how a process works, why a decision was made. For most of corporate history, that body of belief degraded slowly, through turnover and neglect. AI has made it possible to degrade it quickly, at scale, in prose polished enough that nobody thinks to question it.
The organizations that come out ahead will be the ones that treat their knowledge base as an asset with a balance sheet, not a storage folder. Try one test this week. Ask your internal AI assistant five questions you already know the answers to, about your own contracts, policies, and recent decisions. Count how many it gets right, and trace where the wrong answers came from. That count is your opening balance.
References
- Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano, and Jeffrey T. Hancock, “AI-Generated ‘Workslop’ Is Destroying Productivity,” Harvard Business Review, September 2025; and BetterUp, “The Hidden Costs of Workslop,” September 29, 2025, reporting research by BetterUp Labs and the Stanford Social Media Lab with U.S. full-time desk workers. Respondents estimated that 15.4 percent of the work they receive is workslop, and 53 percent said at least some of the work they send may be. About half viewed senders of workslop as less creative, capable, and reliable, 42 percent as less trustworthy, and nearly one in three said they were less likely to want to work with the sender again. Workslop moved between peers 40 percent of the time, up to managers 19 percent, and down from leaders to teams 16 percent.
- Ward Cunningham, “The WyCash Portfolio Management System,” OOPSLA ’92 Experience Report, 1992, the origin of the technical debt metaphor.
- Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal, “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631 (July 2024): 755 to 759.
- Gartner, “Lack of AI-Ready Data Puts AI Projects at Risk,” press release, February 26, 2025, based on a 2024 survey of data management leaders.
