Workslop gives companies a name for the problem many of them created themselves: AI output that looks finished, reads smoothly, and quietly leaves someone else to repair the damage.
The pitch was simple. Put generative AI into your workflows, watch productivity rise, and leave slower competitors behind. Billions of dollars in software spending followed that logic. Now the receipts are coming in, and they don't match the forecast.
In September 2025, Harvard Business Review published research from BetterUp Labs and Stanford Social Media Lab that named what a growing number of offices were already seeing: workslop. The researchers weren't talking about obvious nonsense that gets rejected at first glance. They meant AI-generated work that looks polished but lacks the substance needed to move a task forward. A memo that sounds useful but says little. A summary that misses the point while keeping the tone of competence. A spreadsheet note that makes the next person spend half an hour checking what should have been checked before it was shared.
According to Axios's report on the study, the researchers surveyed 1,150 US desk workers in August and September and found that about 40% had received workslop in the previous month. BetterUp and Stanford put the cost at roughly two hours of extra work per incident, or about $186 per employee each month. For a company with 10,000 workers, Harvard Business Review said that can run close to $9 million a year.
You can see why this is dangerous. Workslop doesn't announce itself as a broken tool. It arrives with clean grammar, neat formatting, and the exact tone companies have taught workers to accept as finished work. That makes it easy to pass along. The damage starts when the bad summary becomes a source, the weak draft becomes a template, and the half-true note gets pasted into a wiki that someone else trusts six months later.
The knowledge base starts rotting from the inside.
This is where the enterprise AI story gets uncomfortable. The issue isn't only that a worker saves 20 minutes and hands two hours of checking to a colleague. The bigger problem is that shared documents, internal reports, customer notes, training materials, and meeting summaries become the raw material for the next round of AI-assisted work. If your records are full of smooth errors, your next AI output will be trained by the same bad habits. Errors don't just survive. They get cited.
MIT's 2025 GenAI Divide report made a related point from another angle. As Fortune and other outlets reported, the MIT analysis drew on 150 interviews, a survey of 350 employees, and 300 public AI deployments, and found that 95% of enterprise generative AI pilots weren't producing measurable profit-and-loss impact. The researchers didn't describe that as a simple model failure. They pointed to poor integration, weak workflow design, and companies buying tools before deciding how the work itself should change.
Buying software is not the same as building a working system. Any executive who still needs that lesson in 2026 is already late.
Goldman Sachs has been just as blunt about the broader numbers. Chief economist Jan Hatzius said in February 2026 that AI had contributed basically zero to US GDP growth in 2025, pushing back on claims that the boom had already transformed the economy. Goldman has also argued that official GDP data misses some AI infrastructure activity, especially when chips and hardware are made abroad, but that only sharpens the point. The spending is visible. The productivity surge is not.
The resistance inside companies is telling you something
A separate 2026 survey by Writer and Workplace Intelligence, reported this week by ITPro, found that 29% of employees in the US, UK, and Europe admitted they had sabotaged AI rollouts at work. Among Gen Z workers, the figure was 44%. The behavior included refusing to use AI tools, bypassing training, and misusing systems. The reasons were not mysterious: job displacement, loss of creativity, and heavier workloads.
You could dismiss that as insubordination. Don't. When nearly half of your youngest workers say they're working against your AI program, they're telling you the rollout has lost legitimacy. They may be wrong in some cases. They may also be seeing exactly what leadership presentations skip over: mandated tools that add review work, produce thin material, and make employees feel as if their judgment is being extracted before their jobs are reduced.
That is the link between workslop and employee sabotage. Both come from AI adoption pushed by pressure and narrative rather than by a serious view of the work. Leaders wanted the efficiency story. Workers got the cleanup job.
The companies that come through this with their knowledge bases intact won't be the ones that deployed fastest. They'll be the ones that set clear rules for what AI output needs before it enters shared systems, keep human review attached to the parts of the work where judgment actually matters, and treat employee skepticism as evidence rather than attitude to be managed. Frankly, most organizations haven't had that conversation yet.
AI can still make work better in narrow, well-defined uses. Goldman has pointed to productivity gains in specific cases, and MIT's report found that a small minority of pilots did work when companies focused on real workflow problems instead of broad transformation theater. That is useful. It is also a warning. If you can't name the task, the owner, the review standard, and the cost of failure, you aren't adopting AI. You're just adding another way for bad work to travel faster.
Also read: Two ex-OpenAI founders built a tool to measure how well AI models actually know who you are • Microsoft is rewriting the economics of enterprise AI and the bill shock is just getting started • Stanford's 2026 AI Index confirms the enterprise window is closing faster than most founders think
