AI Workslop: It's Not the AI, It's the Leader

You have read it already this week.

A status update that covers every base and tells you nothing. A strategy memo with five well-formed sections and no argument. A deck that looks like it took a day and reads like it took four minutes. Nothing in it is wrong. Nothing in it is useful. You get to the end and realize you now have to do the actual thinking the document was supposed to contain.

You did not flag it. You did not send it back. You just quietly redid the work and moved on, because naming it felt petty and there was no obvious person to blame.

That reflex — absorb it, don't name it — is why the problem keeps growing.

THE COST NOBODY IS BOOKING

Researchers at BetterUp Labs and Stanford's Social Media Lab gave this a name: workslop. AI-generated output that looks polished and complete but carries no real payload, shifting the burden of thinking downstream to whoever receives it.

Their survey of just over a thousand full-time desk workers found that 40% had received workslop in the previous month. On average, workers estimated that 15.4% of everything arriving in their inbox now fits the description. Each incident consumed roughly two hours to resolve. Run that math across an organization and it lands near $186 per employee per month — about $9 million a year for a company of ten thousand people.

Nine million dollars, and not one line of it appears in any budget.

It is absorbed silently, two hours at a time, by people who assume it is their job to fix quietly.

Here is the number that should stop you.

Managers reported receiving workslop at a rate of 54%. Individual contributors reported 38.5%.

The people most responsible for AI adoption are the people most buried in its failure mode. Which means the reflexive explanation — my team is being lazy with these tools — collapses on contact with the data. If this were an employee discipline problem, the burden would distribute evenly or fall downward. It doesn't. It concentrates upward, at the exact level where standards are supposed to be set.

That is not a workforce failure. That is a signal about how the work is being led.

FOUR FAILURES WEARING THE SAME DISGUISE

Workslop looks like one problem. It isn't. It is four distinct failures that happen to produce identical-looking output, and treating them as one is why most responses to it don't work.

The first is economic. Effort used to be a quality filter nobody had to enforce. Producing something that looked finished was expensive, so most people didn't bother unless they had something to say. That filter is gone. The cost of generating polished output has collapsed toward zero while the cost of verifying it has not moved at all. This is Capability Arbitrage running in reverse — the gap between what you can produce and what you can stand behind, widening every quarter.

Software teams give us the cleanest measurement of this, because their tools log everything automatically — no one has to self-report. When a developer finishes work, a colleague reviews it before it ships, and the system records how long that takes and whether it passes.

The analytics firm LinearB tracked what happened when AI entered that loop. Developers submitted 98% more work for review — nearly double. That work then sat: AI-assisted submissions waited roughly five times longer than human-written ones before a reviewer even picked them up. And acceptance rates collapsed — where roughly 84% of human-written work used to pass review, only about 33% of the AI-assisted work did. Two-thirds came back.

The bottleneck was never reading speed. It was human attention, which did not double just because the output did.

The developers reported feeling about 20% faster. They were running about 19% slower. A thirty-nine point gap between the experience of productivity and the fact of it.

Now hold that against your own organization. Software is the most instrumented work we have. Every submission counted, every review timed, every rejection logged. The gap still opened, and it took automated telemetry to see it.

Most knowledge work has none of that. No acceptance rate on a strategy memo. No review clock on a board deck. No system that flags when a recommendation came back weaker than it went out. The quality of that work is assessed qualitatively, late, and usually by whoever inherited the problem it created. If the gap between felt productivity and real productivity runs thirty-nine points where everything is measured, there is no reason to assume it runs narrower where nothing is.

It is almost certainly worse. We just have no instrument pointed at it.

The second is psychological. When AI drafts something, ownership gets slippery. The work is not quite yours, so it is not quite your fault. Nobody consciously abdicates — they just verify a little less carefully than they would have if they had written every word. Glean's 2026 workforce research puts a number on where that leads: 41% of desk workers admitted delivering work they could not explain if questioned, and 28% admitted blaming AI for errors in work they had personally reviewed and submitted.

Read that second figure again. That is not a tooling gap. That is a person choosing, in the moment, to let the machine absorb a consequence that belongs to them.

The third is managerial. Most organizations are measuring AI adoption — seats activated, queries run, percentage of the team “using AI weekly.” None of those measure whether anything got better. When you measure usage, you get usage. People produce AI output because producing AI output is what is being counted, and the fact that it lands on a colleague as unfinished work is invisible to the metric.

BetterUp's longitudinal research separates leaders into two patterns here. Automators deploy the tools and pull back the human investment — less coaching, fewer standards, faster rollout. Calibrators pair the tools with coaching, explicit quality standards, and clear guidance about when AI is and isn't appropriate. Teams under Calibrators produced 35% less workslop, gained 47 points on performance measures, and came in 26 points lower on burnout.

Same tools. Same models. Same access. Entirely different outcomes, decided by what the leader did after the license was purchased.

The fourth is cultural, and it is the one leaders least want to hear. People ship polished-looking AI output because polish is a place to hide. If you cannot safely say I don't know how to do this well yet or I ran out of time, a document that looks complete is the safest thing you can produce. The gloss is not laziness. It is cover.

BetterUp's data splits the workforce roughly 64/36 between what they call Passengers and Pilots — those who accept AI output passively and those who retain judgment and ownership over it. The variable that moves people between those groups is not training. Teams with high trust — where people can openly discuss their AI use, including where they are struggling — showed 61% less workslop.

Now the part that reorganizes everything above.

When these mechanisms are modeled together, the two structural ones — collapsing cost and diffused ownership — are not the strongest predictors of whether workslop actually appears. The managerial and cultural mechanisms are. Economics and psychology create the conditions for workslop. Leadership behavior and team culture determine whether it happens.

The tools made it possible. Leaders made it likely.

HUMAN OVER THE LOOP IS NOT SELF-EXECUTING

The standard prescription here is keep a human over the loop — retain ownership, verify before you ship, don't let the machine be the last set of eyes on your work.

That is correct, and it is not sufficient, because it quietly assumes the individual is operating in conditions where it's possible.

Ask a person to stay over the loop while you mandate AI use, measure their adoption rather than their impact, and make it professionally unsafe to admit they're struggling with the tool — and you have asked them to hold a standard you are actively taxing them for holding. The reliable individual behavior is not resistance. It is compliance. They ship the polished thing.

And the verification burden does not disappear when they do. It moves. It lands on the next person, and the person after that, each paying two hours to reconstruct the thinking that was supposed to arrive with the document. That transfer is the Supervisor's Tax — the cost of unverified work, invisible to whoever produced it, unbudgeted by whoever measured it, and paid in full by everyone downstream.

So the discipline has to go one level up. The individual stays over the AI. The leader stays over the human. Not as surveillance — as the deliberate construction of the conditions where staying over the loop is the rational choice rather than the expensive one.

That is three specific things, and none of them are a tool.

Measure impact, not adoption — the moment usage becomes the number, the number is what you'll get. Make ownership explicit before the work starts, so no one has to decide mid-draft whose name is really on it. And build enough trust that I'm stuck is a safer sentence than a well-formatted document that says nothing.

AI does not remove accountability. It concentrates it — upward, into the person who set the conditions.

THE STANDARD YOU SET

Every leader reading this has shipped something they hadn't fully verified. The instinct is to promise to be more careful. Being more careful is not a system, and it does not survive a busy Thursday.

What survives is culture — the standard your team holds when you are not in the room, on the deadline you don't know about, in the document you will never see.

That standard is not built by mandate. You cannot instruct people into ownership. You build it by making ownership survivable: by measuring impact instead of adoption, by naming whose work it is before the work begins, and by making I'm stuck a safer sentence than a well-formatted document that says nothing. A team that can admit what it doesn't know produces less that looks finished and isn't.

The tools will keep improving. Output will keep getting cheaper, faster, and more convincing. None of that touches the question of who is answerable for what gets sent, and none of it ever will.

Accountability can never be artificial. It can be diffused, deferred, buried in a metric, or quietly pushed onto the person downstream — but it cannot be delegated to a machine, and it cannot be automated away. Somebody owns the outcome. The only real decision a leader makes is whether that ownership is designed or discovered after the damage.

Build the culture that accepts it before the tools make it easy not to.

When AI becomes abundant — what will make you indispensable?

Scott Wise brings 30 years of transformation consulting experience to the most important leadership challenge of our time. Author of AI4Leaders: Amplify Your Impact and certified in AI by MIT and Oxford, he helps executive teams and organizations move from AI-Curious to AI-Capable. Explore his work at ScottWise.ai.