Grund InstituteGenerative AI has made documents, presentations, analysis and code cheap to produce. A draft that once took an afternoon now takes ten minutes. Most organizations read that as a productivity gain, and most AI business cases are built on it.
The reading is incomplete. A draft is only useful when someone can act on it. When the draft looks finished but is not, the time saved by the author is spent again by the reader, who has to work out what it means, check whether it is right, and often rebuild it. The author reports a gain. The organization absorbs a cost. Neither shows up in the same place.
This briefing opens a three-part look at the hidden costs of AI adoption. Workslop is the most visible of them. Briefing 02, The AI Productivity Illusion, asks whether saved time ever becomes business capacity, and Briefing 03, The AI Verification Tax, looks at the checking work that automation creates.
Researchers have given this pattern a name: workslop.
At GrundMind we treat workslop as an adoption quality problem, not an output problem. Better prompts help at the margins. The causes sit in how work is measured, handed off and owned.
The author
Where the saving is measured
- Prompt
- Generate
- Skim
- Share
Counted as time saved
The receiver
Where the cost lands
- Decode
- Verify
- Ask back
- Correct
- Rebuild
Not counted anywhere
§The evidence
Two recent studies put numbers on the pattern. They survey different populations with different questions, so the figures should be read side by side, not added together. What they share is the shape of the problem: work that looks efficient at the point of creation generates real effort further down the line.
BetterUp Labs & Stanford Social Media Lab · 2025
Survey of 1,150 US desk workers, September 2025. The $9M figure is the researchers' extrapolation for a hypothetical organization, not an observed loss.
Atlassian, State of Product 2027
Survey of 1,000 product professionals in the US, Germany and France, August 2026. Each figure answers a different question and is not a workforce prevalence estimate.
All of these are self-reported. They still make one point clearly: AI adoption cannot be judged by usage or output volume alone. A team can double its AI activity and lose ground on the work that matters. Briefing 06, The AI Adoption Trap, examines why usage figures make such a poor proxy for impact.
§Five conditions that produce workslop
Workslop is usually framed as individuals skipping a review step. A 2026 peer-reviewed analysis by Karl de Fine Licht points elsewhere. Diligent people produce it under unclear expectations, tight deadlines and thin support, and management controls most of those conditions. That makes it an organizational design question. These are the five conditions we see most often.
Output is rewarded, outcomes are not
When success is counted in tasks closed, documents produced or prompts sent, people optimize for visible activity. The polished draft wins over the useful one.
- Where it shows
- AI adoption dashboards that track usage and volume but say nothing about whether the work was accepted.
- What to change
- Give every AI-supported deliverable a stated purpose, acceptance criteria and a named owner. Track acceptance, not production.
Production speeds up, the workflow stays the same
Drafting faster does not shorten a process when colleagues still have to verify facts, resolve ambiguity and reconcile decisions by hand. The bottleneck moves; it does not disappear. When the process itself is fragmented, AI makes it run faster without fixing it, the subject of Briefing 05, AI on Top of Broken Workflows.
- Where it shows
- Review queues that grow after an AI rollout. More drafts reach approvers, and approval takes longer.
- What to change
- Redesign the hand-off along with the drafting step. Build verification into the workflow, scaled to the cost of an error. Briefing 03 sets out when oversight costs more than the automation saves.
Nobody owns the draft
AI writes the first version. Who answers for its accuracy, completeness and relevance? Without a clear answer, unfinished work travels between teams until someone is forced to finish it.
- Where it shows
- The receiving colleague becomes the default quality control for work they did not commission.
- What to change
- Make cognitive responsibility explicit. Before sharing, the author states what the output is for, what has been checked and what is still uncertain.
One AI workflow for every kind of thinker
Some people use AI to explore and challenge their own thinking. Others need structure, clear outcomes or strong verification before they can rely on it. Neither is better. Workslop grows when one workflow is imposed on everyone, regardless of cognitive fit or task risk. Briefing 07, Cognitive Fit, explains why the same tool works for one employee and not another.
- Where it shows
- A single prompt library and training program, with very different quality coming out of different teams.
- What to change
- Match structure, explanation and verification to the role, the task and the interaction pattern of the people doing it.
Speed hides the downstream cost
The author reports real time savings. The receiving team experiences extra corrections, clarification requests and mental effort. Without end-to-end measurement, the loss stays invisible and the business case looks healthy. Briefing 02, The AI Productivity Illusion, takes this gap between time saved and value created further.
- Where it shows
- Strong self-reported time savings alongside flat delivery times and rising complaints about quality.
- What to change
- Measure net efficiency: all human effort on the task, including prompting, checking, correction, coordination and rework.
§Where the value leaks
Workslop does not mean AI is lowering productivity. It means AI is redistributing work, and some of that redistribution is a loss. The leakage takes five recognizable forms.
| Leak | What happens | What it costs |
|---|---|---|
| Hidden rework | Time saved in drafting is absorbed by colleagues who correct or rebuild | Capacity the business case already counted as freed |
| Decision friction | Volumes of polished, unsupported material bury the facts a decision needs | Slower, less confident decisions |
| Eroded trust | People stop relying on shared work and check everything twice | A permanent verification overhead |
| Cognitive load | Readers must separate the meaningful from the merely plausible | Fatigue concentrated on the most careful people |
| Phantom ROI | Usage and speed figures show success while outcomes stay flat | Investment decisions made on the wrong evidence |
Two of these leaks deserve their own treatment. Eroded trust is one side of a wider problem covered in Briefing 08, The AI Trust Paradox, where too little trust is as costly as too much. Phantom ROI feeds directly into Briefing 12, The Missing AI ROI, on why companies struggle to prove their AI investments pay off.
§What leaders should measure
The question to ask is not how much AI is being used. It is how much of what AI produces is accepted, and at what total human cost. We propose six operational indicators. They are working measures for diagnostic use, not yet validated GrundMind metrics.
- Workslop incidenceShare of sampled deliverables that needed avoidable, substantial rework.
- Downstream correction effortHours the receivers spent correcting or rebuilding AI-supported work.
- First-pass acceptanceShare of deliverables accepted without significant correction.
- Net time savingsBaseline human effort minus total AI-supported effort, downstream work included.
- Quality preservationErrors, completeness and acceptance compared with the pre-AI baseline.
- Accountability clarityEvidence that each deliverable has a named owner and acceptance criteria.
When workslop keeps appearing at the same hand-off, treat it as a process signal. Ask whether the cause is missing context, weak quality criteria, the wrong tool for the task, unrealistic deadlines or unclear responsibility. Fixing the system works better than blaming the individual.
The same logic applies to AI agents. The Atlassian finding that half of product professionals regularly check agents to keep them on task is workslop in another form: supervision effort that rarely appears in the business case. Briefing 11, The Agentic AI Supervision Problem, looks at who manages that work.
§The GrundMind view
GrundMind treats AI adoption as a system of business value, workflows, human cognition, roles and governance. Seen that way, workslop is a symptom. It marks the places where the way AI is used has drifted away from the way the organization actually creates value.
It also lands unevenly across the four interaction archetypes that underpin Briefing 07, Cognitive Fit:
Move fastest and generate the most drafts. Without validation steps, they are the likeliest source of polished but unfinished work.
Carry much of the verification load. When workslop rises, they spend more time checking and less time on their own work.
Respond to unreliable output by tightening rules, sometimes to the point where useful AI work stops.
See their doubt confirmed. "This creates more work than it saves" stops being an opinion and becomes an observation.
Our diagnostic work traces where AI-generated work moves effort and cost to other parts of the organization. It asks where verification and rework accumulate, whether time savings survive the whole workflow, where leadership expectations differ from what employees experience (the subject of Briefing 04, The AI Perception Gap), and whether each role's tools and practices fit the work it does. For workslop, two GrundMind diagnostics do most of the work. AI Value Leakage locates where productivity gains are lost to downstream rework and correction. AI Net Efficiency tests whether AI reduces total human effort once verification and revisions are counted.
AI productivity is measured at the end of the workflow, not at the point of generation.
Workslop is what happens when an organization counts the speed of drafting and ignores the cost of finishing. The fix is not less AI. It is clearer ownership, verification sized to risk, workflows that fit the people using them, and measures that follow the work all the way to acceptance.
An AI-enabled organization does not succeed by producing more content, code or reports. It succeeds when it reaches better decisions and stronger deliverables with less unnecessary human effort.
The question is no longer how much work AI can generate. It is how much valuable work the organization can complete with it, reliably.
Book an AI adoption diagnostic pilot.
Find out where AI saves time across the whole workflow, where that time is spent again downstream, and which hand-offs need redesigning before you scale.
About this briefing
The GrundMind Adoption Briefings are short, leadership-facing notes from the Grund Institute on the measurement and design of AI adoption in organizations. The series accompanies the longer-form Grund Institute white papers on cognitive fit, the Second Mind framework, and the GrundMind diagnostic.
Sources
- BetterUp Labs and Stanford Social Media Lab, workslop research, 2025.
- Atlassian, State of Product 2027, pp. 18–19, 24.
- Karl de Fine Licht, AI Workslop: The Moral Significance of Withholding Effort, 2026.
- Grund Institute, Bridging the Cognitive Gap in AI Adoption, 2026.