Continues AI at work is evolving (partner in the process vs walk-away employee), Personal AI is in Uber time (meter before the habit hardens), and Intelligence spend is the new cloud spend (own the bill the way you own cloud).
Workslop is a review tax. BetterUp Labs and Stanford's Social Media Lab surveyed 1,150 full-time U.S. desk workers in August and September 2025 and published the findings in Harvard Business Review that September. Forty percent said they had received workslop in the prior month. Resolving one incident took an average of 1 hour 56 minutes. The researchers priced that burden at about $186 per affected employee per month, which they scale to roughly $9 million a year in a 10,000-person organization after applying the 40% prevalence rate. That $9 million line is their extrapolation, not a ledger total. Treat it as a floor.
Workslop, in their definition, is AI-generated work content that masquerades as good work but lacks the substance to advance the task. It looks finished. It is not. The cost lands on the person who has to use it.
CNBC Make It returned to the same research on 20 August 2026, in a piece on how AI is warping workplace interaction. The study still sits in the open: polished AI output that forces rework, strained trust, and a productivity bill that never shows up on the sender's dashboard.
What the number actually measures
I buy and operate intelligence for a company north of 3,000 people. I also read AI decks, summaries, and "done" docs every week. The BetterUp × Stanford figure matches the pattern I see. The send looks fast. The receive is slow.
Three numbers from the study matter for operators. Use the public sources, not vibes.
1/ Prevalence. Forty percent of desk workers reported receiving workslop in the last month. Managers were more exposed (54%) than individual contributors (38.5%). Peers send most of it sideways, but it also moves up and down the chain.
2/ Time per hit. The BetterUp Labs PDF puts resolution at 1 hour 56 minutes per incident. Their blog writeup rounds a related self-report to 1 hour 51 minutes. Public coverage usually says nearly two hours. I use the PDF figure and call the round number an estimate of the same self-reported cost.
3/ Trust damage. About half of recipients rated the sender less creative, capable, or reliable. Forty-two percent trusted them less. Roughly one in three said they would be less likely to want to work with that person again. Two hours is recoverable. Quiet reclassification is not.
The mechanism is cost transfer. The producer books a productivity win. The reader absorbs a larger loss. Org dashboards can still look green while the ledger nets negative. That is one candidate explanation the researchers float for why so many generative AI pilots show little measurable return. Candidate, not proof.
Continuity with how people actually work
In late August I argued that AI at work is splitting into two habits. One treats the model as a partner in the process. You stay on the draft, correct it, and own the signature. The other treats it like an employee you dispatch and walk away from. Workslop is what the walk-away path dumps on the next person when nobody sat with the second read.
Personal AI is still in Uber time. Cheap stickers train the habit of asking the model for volume. Intelligence spend is starting to look like cloud spend. Workslop is the desk-level invoice that shows up before Finance ever opens a token report. If you only meter seats and tokens, you miss the review tax.
Shadow AI is next on this series. Unsanctioned tools and personal accounts create a different risk surface. This essay is not that essay. The control here is simpler: what leaves a teammate's keyboard has to survive a second read against a real source.
Show the cut
Kill vanity AI output that creates review tax. Keep AI that shortens the path to a decision someone can sign.
What dies under that rule:
1/ Decks and memos that read finished but fail a source check. Citations that do not exist. Numbers that do not match the export. Summaries that omit the material risk while polishing the immaterial.
2/ "I used AI" as a status signal. Mandates without a definition of done push people into low-effort compliance. Forty-one percent of employees in the study said leadership encouraged AI use without explaining how or why.
3/ Throughput theater. Volume of artifacts is legible to a manager. Whether the artifact advanced the task is legible only to the person who has to build on it. Reward the second metric.
What stays:
1/ Partner loops. Draft with the model, then verify against the contract, the dataset, the codebase, or the customer record before you hit send.
2/ Provenance over prohibition. "What did you verify?" is answerable. Blanket bans push usage underground and make workslop harder to catch.
3/ Measured rework. Hours spent fixing AI output, per team, per month. If that number is invisible, the fake productivity gain keeps looking real.
Team trust cut workslop by 61% in the BetterUp × Stanford research. An agentic "pilot" mindset cut sending workslop roughly in half. Neither fix is a better model. Both are operating habits.
What I'd do Monday
Prove it on one team. Expand on pull. Operationalize so it survives you.
1/ Prove it on one team. Pick one workflow that already ships AI-assisted artifacts to other humans: weekly status, customer notes, research briefs, or internal decks. For two weeks, require a second-read checklist before send: claim checked against source, numbers matched to the export, open questions labeled as open. Count hours spent fixing inbound AI output on that team. That count is your baseline review tax.
2/ Write the cut in one page. Show what dies (vanity send, throughput theater, unverified polish) and what stays (partner loop, provenance, rework hours). Same quality bar you would put on a signed pack. No company-wide AI manifesto.
3/ Expand on pull, not slideware. Widen the practice only when a second team asks with a named owner and a workflow. Do not launch a "workslop zero" banner to the whole company. Pull beats theater.
4/ Operationalize so it survives you. Name one owner for review-tax metrics the same way you name an owner for intelligence run rate. Put inbound rework hours next to token spend in the weekly ops review. Document the escalation when a sender keeps dumping unverified polish: return the work, do not silently absorb it.
5/ Rehearse the exit. If AI volume doubled next month and quality stayed flat, which artifact types get deleted from the send path, which stay draft-only, and which require two-person signoff? Write that list while the habit is still soft.
The operators who win do not ban the model. They refuse to let polished emptiness move the bill onto the next person. Prove the cut on one team. Expand when people ask. Leave a process that still works when you are out of the room.