Botsitting: The Hidden Cost of AI Productivity
The industrial revolution reduced or eliminated tedious manual work. The intelligence revolution is well... different. Many employees have become unexpected managers of digital interns. Some feel that they are babysitting AI models, AKA "Botsitting".
You've used an AI model right? Have you ever just given it a task and not checked it's work? Just let it rip and hit "deploy", or turned in the report?
Today’s AI agents resemble a tireless, eager junior employee with poor memory. A junior employee produces useful work. They can accelerate progress. They can handle routine tasks, sometimes complex ones too, but they require supervision. The more capable they appear, the more dangerous unsupervised mistakes become.
The paradox is that organizations often measure the junior employee’s output while ignoring the supervisor's time spent reviewing and improving it. AI magnifies the same attribution problem at scale.
“Who is assisting whom?”
Simple productivity calculations assume that if AI completes a task in ten minutes instead of a human in one hour, fifty minutes have been saved. In practice, it's more complicated. AI has inverted the traditional relationship between humans and tools. Instead of AI assisting people, people assist AI by:
- Explaining the task
- Providing context
- Reviewing the output
- Identifying mistakes
- Revising prompts
- Re-running the process
- Integrating the results into existing workflows
Recent research suggests knowledge workers now spend more than six hours per week performing this kind of labor. An employee may generate a draft report in five minutes, then spend another thirty minutes validating facts, fixing formatting, and correcting subtle errors. Workers also move between multiple AI tools, comparing outputs and manually transferring information. The employee becomes the integration layer.
A frustrating observation from employees using AI is that current systems don't learn from their mistakes the way other employees do. They require repeated reminders, additional guardrails, and ever-growing documentation. Employees derisively call this work "botsitting" — assisting the model to do the work they used to do themselves.
“If you take a really smart person and throw them at a completely new job, they won’t be great at it at first but will likely learn the ropes quickly from instruction or trial and error.”
This same process is challenging for LLMs today. It will only get more complex as more models, APIs, tools, and techniques emerge, especially as teams tackle increasingly ambitious use cases. You won’t be able to judge business outcomes by individual inferences, which is how most people approach LLM engineering today.
The Emotional Cost
Many people derive satisfaction from creating something: writing code, solving customer problems, designing systems, producing analysis, or crafting a presentation. These activities require skill and judgment. They provide a sense of ownership, meaning, and engagement. What happens when AI automates the part of your job you actually enjoy?
Instead of doing the work, the human increasingly reviews the work. Instead of solving the problem, the human monitors a system attempting to solve it. Few employees wake up excited to spend their day correcting errors, rewriting prompts, or validating generated content.
Their complaint is not that AI is ineffective. The complaint is that AI is taking over the parts of the job they enjoy and replacing them with oversight and supervision. Organizations may discover an accompanying decline in motivation, craftsmanship, and job satisfaction.
When AI supervision is untracked, unbudgeted, and unrewarded, employees may reduce oversight and start delivering work they can't fully explain or defend. It is a slow surrender of agency.
"I don't like that I basically tell Claude what to do, and then either go do busy work or waste time on the internet."
– Comment posted on Hacker News: The rise of the 'botsitters'
Am I Serving AI or Supervising AI?
Disenfranchised employees view AI as a needy coworker, a source of verification work, and a system that automates the enjoyable parts of the job while leaving them responsible for quality control.
Employees who are heavy AI adopters see themselves as "AI Management". Same activities. Completely different interpretation:
- This isn’t babysitting, this is delegation
- AI is a junior employee and I am the manager
For them, AI takes care of implementation details and lets them focus on creativity, architecture, strategy, and judgement. They describe AI as a force multiplier, a junior engineer, a delegate, a productivity accelerator.
"Of course I need to supervise it. That's what managers do."
The deeper issue is about identity. For decades, knowledge workers have defined themselves by their ability to produce. Software engineers wrote software. Analysts produced analysis. Designers designed. AI is shifting many of them toward the role of manager, editor, reviewer, or quality assurance specialist. Some people love that transition. Others don't.
The dialog is about what kind of work remains for humans after the technology arrives.
Reducing the Need for Botsitters
A truly transformative approach will reduce the emotional costs and require less AI supervision, less correction, and less orchestration. Today, human oversight exists for three primary reasons:
- The AI lacks context → World Models
- The AI cannot reliably learn from mistakes → Closed Learning Loops
- We do not fully trust the output → Independent Validation, Critique Model
Organizations that achieve the greatest leverage from AI will be the ones that tackle these constraints. When AI learns from it's mistakes it reduces future oversight. When an AI system can independently verify its own output through testing, simulation, formal checks, or measurable outcomes, the need for human oversight declines dramatically.
World Models
A major reason managers exist is to provide context. They make decisions based on an understanding of:
- Customer needs
- Business objectives
- Organizational priorities
- Historical decisions
- Constraints and tradeoffs
- Culture
Jack Dorsey's concept of a "World Model" addresses this challenge. A world model is a structured, persistent representation of how an organization works: its customers, products, processes, goals, constraints, and the relationships between them. Rather than having context re-explained through prompts every session, the AI operates from a durable understanding of the business.
For example, a customer support AI with a customer "world model" would know which customers are most valuable, what products they own, their historical issues, company service policies, and current business priorities. It could make decisions like an experienced, seasoned employee rather than a stateless chatbot. As world models improve, much of the contextual guidance currently provided by managers can be embedded directly into the system.
Anything you can measure, you can train against - so the measurable slides to commodity. What increases in value are things a public model can't reach: correctness that exists only inside one firm's private data.
Closed Learning Loops
Many current AI deployments operate as one-way systems: Request → AI → Output
More advanced systems create learning loops: Request → AI → Action → Outcome → Feedback → AI
Examples include customer support systems learning from satisfaction scores, sales systems learning from conversion rates, recommendation engines learning from engagement, and software agents learning from test failures. Organizations that build effective feedback loops will need dramatically less supervision over time.
"This means the real opportunity is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound."
– Satya Nadella, CEO Microsoft
Independent Validation and Critique Model
The most powerful way to reduce oversight is to make outcomes objectively verifiable. When a result can be independently tested, measured, or validated, human review becomes unnecessary.
- Unit tests validate software changes
- Financial reconciliations validate accounting calculations
- Data quality checks validate transformations
- Monitoring systems validate infrastructure changes
- A/B tests validate business outcomes
In these environments, AI can operate autonomously because success is not determined by human opinion. The system itself can determine whether the work was correct.
The other option is to use a second, independent model to review and critique the output of the first model. This is called "LLM as Judge", "Verifier Model" or "Critique Model". When the second model agrees things can move forward without human oversight. When the second model surfaces concerns it can be raised to a human for review.
Summary
The defining challenge right now may not be building smarter AI. It may be reducing the need for humans to become full-time botsitters. Until then, organizations should be cautious when measuring AI productivity gains.