AI Workslop and the Leadership Failure Behind It
Across our member community, the same theme keeps surfacing in conversation. People with experience and judgment are spending more of their week reviewing other people’s AI output — work that lands as if it’s finished thinking, then turns out to need substantial fixing. Sometimes hours of fixing. The senders span the org chart. Senior VPs shipping strategy documents that read like a competent stranger drafted them, because, well, a competent stranger did. I’ve been hearing about a marketing partner who built an HR agent over a weekend and is now sending HR-shaped outputs to actual HR leaders. (Marketing made an HR agent. We’re really doing this.) Direct reports submitting single GPT responses as deliverables.
The output and the cleanup deserve separate names because they are separate problems.
- The output is AI Workslop, the term Stanford and BetterUp researchers gave to plausible-sounding AI-generated work that creates more work for the recipient.
- The experience of cleaning it up is AI Brain Fry, the term BCG researchers used for the cognitive overload of overseeing your own and others’ AI work.
Workslop looks finished and creates work nobody planned for
The work that workslop creates comes as a surprise. It gets produced faster and arrives looking finished. The person who receives it expects completed work, then has to discover that it isn’t, and only then can start to review it, rewrite it, fact-check it, or restore the context that AI stripped out.
Workday’s January 2026 study of 3,200 business leaders quantified the pattern: while 85% of workers report saving time using AI, roughly 40% of those savings are immediately lost to rework.
i4cp’s Rob Cross has documented the underlying pattern for years. His work on collaborative overload shows how a small number of capable, generous people tend to absorb a disproportionate share of collaborative demands. AI adds another layer. The people who say yes to collaboration now also get unfinished AI work from colleagues who don’t know it’s unfinished. Nobody sends workslop to take advantage of a generous coworker, so the receiver not only has to recognize workslop but find a way to tell the sender it sounds like workslop.
The costliest workslop is the kind leaders act on
When workslop informs decisions about strategy or the workforce, the cost goes beyond an afternoon of rework to the decision itself.
Brain fry costs trust and retention
The fixing is only part of what drains people. After a few polished-looking deliverables turn out to be unfinished, nothing that lands can be taken at face value. Every document has to be read with suspicion before it can be trusted, and that constant oversight is the cognitive overload of brain fry.
BCG’s analysis found that 34% of workers reporting brain fry showed active intention to quit, compared with 25% of those not experiencing it, a 39% relative increase. Stanford and BetterUp also found damaged perceptions of colleagues who sent workslop, including lower trust and less willingness to work with the sender again.
Prompting taught us to treat AI as a transaction
Four years into the AI revolution, the dominant explanation for workslop is still that people need better prompts. But the real problem is that we taught people to think of AI as something to “prompt” in the first place. Prompting implies a transaction: compose request, receive answer, ship answer. One and done. Maybe we’ve advanced to two or three prompts, then done.
Language models are very good at putting words together that sound about right but really aren’t, because they’re based on an average of the parts of the internet each AI model has access to. To address that, I’ve stopped talking about prompting AI and started talking about giving AI the context it needs. Because working with AI is a conversation. Give it everything you know about the task. Challenge what it gives back. Question when it sounds confident. Keep going until the work can stand up to your own professional judgment.
Neuroscientist Vivienne Ming ran an experiment that shows the difference. She asked teams to predict real-world events and compared three kinds: people working alone, AI working alone, and people working with AI. Most of the human-AI teams took AI’s answer as their own and did no better than AI by itself. Only a small group, about 5 to 10 percent, outperformed AI on its own. They challenged it and asked it to argue against itself. Essentially, they said to AI: Don’t tell me why I’m right. Tell me why I’m wrong.
Professional judgment must be built – here’s how
Telling employees to challenge AI is not enough. People need practice recognizing what weak AI work looks like, diagnosing why it is weak, and deciding what to do next.
Professional judgment in AI-assisted work is the ability to frame the work well, recognize when an answer is inadequate, diagnose the problem, verify what matters, improve or reject the output, and remain accountable for the final result. It is visible in behavior. Did the employee give AI the context needed to do the work? Did AI improve the employee's thinking, or merely accelerate production? Could the employee identify weak logic, unsupported claims, missing evidence, false confidence, or the wrong audience? Did the employee revise, verify, or discard the output before passing it to someone else?
Philippa Hardman has assembled a research-informed methodology for interrupting common patterns of AI overreliance. Her practices include separating production from critique, independently checking factual claims, deciding how work will be divided between human and AI, and preserving independent thought before AI establishes the framing. The larger lesson is that judgment develops through repeated acts of comparison, challenge, verification, and reflection. It cannot be built through prompting advice alone.
Organizations can turn that principle into practice by helping employees:
- Define the audience, purpose, constraints, evidence, and success criteria before using AI
- Produce an initial human point of view before requesting a draft or recommendation
- Ask AI for alternatives, objections, risks, assumptions, and missing evidence
- Review important output separately from the conversation that created it
- Verify consequential claims against authoritative sources
- Compare AI output with an established quality standard or strong example
- Diagnose specifically what is wrong rather than merely calling the output generic
- Revise, reject, or escalate when AI output cannot meet the required standard
- Reflect on what improved the work and what they will do differently next time
1. Rewrite training and AI usage guidelines around judgment
If your AI training centers on writing prompts and your usage guidelines center on data security, neither one tells people what to do with the answer AI gives back. Update both so the practices above are described as the expected way of working.
2. Give people a private place to practice
It’s hard to realize you’re not as good at using AI as you thought. That can be especially true for senior leaders, who may feel pressure to model confidence while privately lacking a method for evaluating their own use.
i4cp's AI Fluency Coach gives individuals a private way to examine their AI-assisted work. A user can summon the coach into an AI-assisted work session and receive developmental feedback. The coach looks for observable behaviors such as context-setting, iteration, verification, critique, output judgment, quality control, and learning. It does not evaluate job performance, intelligence, productivity, or technical sophistication. It also helps the well-meaning sender who can’t tell that their own output is workslop.
The coach organizes the feedback around six dimensions: output judgment, context-building, thinking improvement, workflow integration, taste and quality standards, and learning orientation. Its purpose is not to certify that someone is good at AI. It helps the person identify the strongest behavior visible in the selected work, the most important missed opportunity, and a concrete practice to try next. Results are treated as a developmental snapshot rather than a fixed capability label.
The coach needs to remain private and voluntary. It is not designed for hiring, promotion, performance review, employee monitoring, ranking, discipline, or other employment decisions. Organizations can set expectations for responsible AI-assisted work without turning an employee's learning environment into surveillance.
3. Give managers a simple coaching framework
Managers coaching an employee need a simpler approach. i4cp’s companion guide, How to Coach for Effective AI Use, groups the same behaviors into three things a manager can see in the work itself:
- how the person sets up the work,
- how they check it, and
- whether they own the result before passing it on.
Each comes with a question to ask in a 1:1, so the coaching conversation stays grounded in real examples of the person’s professional judgment.
4. Make sure performance management goals focus on judgment
Performance management teaches what the organization expects by making some behaviors visible, consequential, and rewarded. If AI judgment matters, the review framework should say so.
- If it rewards efficiency above all, e.g. shipping as much AI output as you can, as quickly as you can, expect more of it. (One company actually had a short-lived leaderboard for most token use per employee!)
- If it rewards exercising judgment over AI output, people have a reason to build that capability.
To support this, I’ve been crafting a set of SMART Goals for Effective AI Use that i4cp members can tailor.
Work you can trust on arrival
A finished-looking document used to tell you something. Someone with expertise had come up with the ideas and checked them, because it took knowing the subject to write about it that well. AI removed that requirement. Anyone can now produce work that sounds expert on a subject they don’t know.
That work fools the sender first. They ask “Is this good?” and it sounds right, so they say yes. It takes professional judgment to know credible work from work that only sounds credible, and AI sounds credible on every subject.
Organizations can no longer assume that judgment is there. They have to teach judgment, give people a way to practice judgment with AI, coach for judgment, and expect judgment.
Then colleagues can take each other’s work at face value again.