During a collaboration between Harvard Business Review and the Stanford Social Media Lab, researchers uncovered an interesting finding. Employees reported spending an average of 1 hour and 56 minutes dealing with each instance of AI workslop. Based on the self-reported salaries of the survey participants and the estimated time to rectify each instance, that adds up to an invisible tax of roughly $186 per month per employee.
What is AI Workslop?
AI workslop is AI-produced output that seems credible on the surface but lacks any real, meaningful substance, or even contains factual inaccuracies. It may come in the form of an investment report that cites a bunch of statistics but is devoid of any wisdom for why the company is actually a good investment. It may come in the form of a model that incorrectly calculates several metrics. It may come in the form of a memo that was clearly not written in the person’s voice. Regardless of how AI workslop shows up, study participants reported a mixture of the same emotions: 53% felt annoyed, 38% felt confused, and 22% felt offended.
It is often the result of people blindly trusting AI outputs without quality checks, or any real thought at all. While AI models have undoubtedly improved since the study’s original publication in September 2025, AI workslop is still a problem that must be carefully considered when implementing AI.
Grounding
“AI, you are grounded.” No, not that kind of grounding. In the AI world, “grounding” refers to providing AI a carefully curated knowledge base to draw its answers from. It ultimately helps reduce the rate of “hallucinations,” the term for when generative AI says something that is verifiably incorrect.
There are very complex ways to ground AI, such as Retrieval-Augmented Generation (or “RAG”), but there are also simple ways to achieve relatively similar results. For example, when prompting a generative AI tool, you can upload the PDF you want it to draw from and say, “Only answer using the knowledge in this PDF.” There are also more purpose-built tools like NotebookLM, which fundamentally operates like a RAG system without all the complex setup. Regardless of the approach, grounding is a particularly powerful way to reduce hallucination-driven AI workslop.
Context
Leading generative AI models can hold nearly 1 million “tokens” (the unit of measure for AI, where one token equates to roughly 0.75 words) in their “context window,” or memory, at any given time. To put that into context … pun intended … it would take roughly 1.5 million tokens to capture the entire Harry Potter series. Yet people often use only a minimal number of tokens when prompting AI. In some ways, it is like owning a Ferrari but only driving it 5 miles per hour. (To be fair, there are dangers that come from driving a Ferrari too fast too, but we will focus on the pitfalls of driving it too slow for now.)
Generative AI models can produce brilliant results because they are trained on the knowledge of the world. But generic prompts (e.g., “Write a motivational email I can send to my team”) produce generic outputs. Context-rich prompts produce far better ones. Compare the prompt above to: “I am about to kick off my Q1 ‘Win New Client Business’ initiative with my team of 5 people. Each is expected to bring in $100,000 in new business, which is a big stretch but doable with some creativity and hard work. Write me a motivational email I can send to my team that is in my voice (example emails attached) but also blends Herb Brooks’ and Steve Jobs’ styles of motivation.” Context is a particularly powerful way to reduce generic AI workslop.
SKILLs
Whenever you hire a new employee, they need to be trained to perform certain tasks according to the firm’s standards. The same is true with AI. And the creation process is a lot simpler than most people would think. In Claude or ChatGPT, simply type “Help me build a SKILL that …” and fill in the blank. The tool will produce a first draft of that SKILL, which you can refine further. Once it is refined to perfection, these SKILLs are automatically leveraged whenever Claude or ChatGPT sense that the SKILL is relevant.
The SKILL tool is actually quite new (launched in late 2025) and is quickly expanding to other AI platforms like Gemini. As a result, its full potential is only partially appreciated. But it has already generated some early wins. Some people have created SKILLs that polish up PowerPoints to meet brand standards (e.g., every bullet must end with a period, titles must be set in 28pt font, always use the hexadecimal color #135AE1 for font). Some have created SKILLs that analyze a document according to a framework learned in a book (e.g., read this industry report and categorize the information according to the principles of Blue Ocean Strategy). Some have created SKILLs so that AI writes in a style more aligned to their voice (e.g., do not use em dashes, always lead with a shocking opener, sprinkle in an exclamation point every few sentences). SKILLs will keep being applied in a wide variety of ways, but at a minimum they are making real progress against irrelevance-based AI workslop.
Human-in-the-Loop
Will grounding, context, and SKILLs eliminate all AI workslop? Definitely not. Which leads to the final point. It is always important to keep the human in the loop. AI does not carry the same accountability and fiduciary duty that industry professionals do. Even more, AI does not have the same wisdom, judgment, and taste that humans do. While it can undoubtedly accelerate and enhance the creation of certain deliverables, the human should always review the output and have the final say.