One's Agentic (LLM) Efficiency is Another's AI Slop
In the work context, many (if not most) CEOs are bullish on the implementation of GenAI (LLM) tooling across their organizations, with the hope of increasing profit through efficiency and productivity.
As many employees are pushed to adopt this agentic approach into their work lives--from software engineers, customer service reps, and even lawyers--we see many online posts and articles about how quickly these LLMs can ship work product and “increase” ones productivity to new heights.
Often less highlighted is the perspective of the recipients of such agentic work product. To one, quickly sending out an AI-generated presentation, report, memo, codebase, etc. seems brilliant (think of the time and money saved!). But what about the recipients or end users of such work product who are tasked with putting it to use?
Recently, BetterUp Labs partnered with Stanford Social Media Lab to conduct a study on AI-generated content in the workplace (https://www.betterup.com/workslop) (https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity). TLDR findings: not great. The researchers coined the new term “workslop” to describe AI-generated work product that looks good, but lacks meaningful substance--leaving the recipient colleagues to clean up and do the real thinking. Workslop eliminates the initial time/money saved, creating a lack of accountability and context, and negatively impacting the relationships and perceptions of colleagues.
A few key findings on the hidden cost of workslop (from Sep 2025 survey):
40% of US desk workers reported receiving workslop in the last month
2 hours was the average time to resolve each incident
$9M annual cost for a 10,000 person company
Teams also reported losing trust in their colleagues, becoming disengaged in the projects, and having an overall feeling of frustration and confusion.
The hidden impact of workslop seems to align with MIT’s recent findings that 95% of GenAI pilots at enterprise organizations fail (https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf). In other words, billions of dollars invested in enterprise GenAI tooling are yielding little to no results.
We also witnessed a fairly amusing (yet alarming) illustration of this problem when Deloitte (big four consulting firm) was caught using AI-generated content in a report provided to the Australian government to help crackdown on welfare--resulting in references to fictitious academic research papers and fabricated quotations from a federal court judgement, among other errors (https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund/). One can imagine that Deloitte’s leadership was initially pleased to see that GenAI was being used to expedite client work product--until realizing that Deliotte must now issue (at least) a partial refund of the $290,000 to accommodate for the workslop it provided.
With any emergent technology, it’s understandable to encounter some bumps along the way. However, with the billions (perhaps trillions) of dollars being poured into the development and deployment of LLMs, it appears that the inherent limitations of these systems are often being overlooked or blatantly ignored as organizations are pressuring employees to use these tools in all aspects of their daily work--as quickly as possible.
I’m reminded of Gary Marcus, who for many years has been (and continues to be) a voice of reason on the topic--highlighting the inherent flaws and limitations with LLMs specifically (Marcus on AI). We are watching many of Gary’s predictions play out in real time, with little to no acknowledgment by the leaders of these GenAI empires as to the prevalence of these issues and how these problems (hallucinations, etc.) will be solved by bigger data centers and more training data.
Personally, I am not an anti-AI luddite. However, as someone experiencing workslop at an increasing frequency, I am concerned that too much attention is being placed upon the productivity champions shipping out work product and not enough attention is given to the recipients of workslop that are left to clean up the AI-generated mess. Confronting this issue is inevitable (as shown in the Deloitte example), but the sooner we can acknowledge the problem and progress towards a more sustainable technological and organizational solution, the sooner we can materialize the true potential of AI-empowered humans.