While artificial intelligence has led to workforce reductions at many companies, it is simultaneously creating new types of jobs. According to the latest data from a recent online recruitment campaign, over 5,000 internet companies released more than 200,000 job openings this summer. Leading firms like JD.com, Tencent, ByteDance, and Meituan collectively contributed over 46,000 positions, spanning cutting-edge fields such as AI algorithms, large model applications, and high-performance computing. The increase is not just in the number of roles but also in their variety, with new positions like algorithm engineers and prompt engineers emerging. A recently popularized role, "botsitting," has significant potential to become a major new employment pool.
So, what is botsitting? Just as babysitting involves caring for an infant, botsitting refers to supervising and managing AI systems. Departments like Microsoft's Copilot unit already have roles such as "AI Trainer," "Digital Adoption Specialist," and "AI Advocate." Their work essentially constitutes a form of botsitting. Specifically, it involves teaching colleagues how to use AI, checking the quality of AI outputs, and integrating AI into specific business processes. In practice, there's often a gap between the answers an AI provides and their practical application. AI can produce hallucinations or completely misinterpret instructions. In short, you need to spend time refining the answers AI gives you. How much time exactly? According to the "Work AI Index 2026" report, the average office worker spends nearly a full workday each week on botsitting tasks. Since botsitting isn't particularly difficult and requires only basic AI experience, it's well-suited for recent graduates looking to understand the industry. Much like data annotators in the early days of AI, this role started small but grew in importance and scale as AI developed.
What Does Botsitting Entail?
To understand what botsitting involves, consider an example. Suppose you ask AI to write a market analysis report. The AI initially lacks knowledge of your company's product lines, so you must first feed it the background information. After it generates a draft, you need to verify competitor data point by point. Finally, you must format the AI's output. This entire process can take one to two hours. The "Work AI Index 2026" notes that 87% of office workers already use AI at work, claiming it saves them an average of 13 hours weekly. However, they also spend an average of 6.4 hours per week on botsitting. This means half the time saved by AI is given back through these supervisory tasks. Furthermore, botsitting has a self-worsening tendency: 69% of people submit AI-generated content without any review. The report further indicates that of the time employees spend on AI, 37% is on botsitting, 36% on using AI to produce content, and the remaining 27% on learning tools and building agents. While these figures may seem surprising, the reality is even more striking. One reason botsitting consumes more time than actual AI use is that 36% of AI sessions "fail" completely and must be restarted. Another reason is the proliferation of tools; 77% of AI users switch between multiple AI tools weekly, with 33% using four or more simultaneously. Among Claude users, only 0.5% use Claude exclusively; the average user runs four other AI tools concurrently. Each switch to a new tool breaks the context of previously fed company background or project details, requiring re-input. The report terms this the "context tax." It shows that for every 10% increase in time spent feeding context to AI, the probability of employee burnout rises by 25%. Moreover, heavy AI users engage in botsitting more than twice as often as light users. Essentially, the more you rely on AI, the greater the proportion of your time spent on supervision. The more unstable the AI's output quality, the more human botsitting is required; more botsitting leads to greater fatigue; greater fatigue increases the tendency to skip review and submit directly; and as more people skip review, organizations fail to see AI's true return. The report reveals that while 75% of individual users believe AI boosts productivity, only 13% report significant organizational improvements from AI. The missing 62% is attributed to a lack of proper botsitting.
Why Botsitting Suits Recent Graduates
The National Association of Colleges and Employers "Job Outlook 2026" report indicates that 45% of employers rate the 2026 talent market as "fair," the worst rating since 2021. In the fall of 2025, corporate employers predicted only a 1.6% increase in hiring new graduates. Although this figure rebounded to 5.6% in a spring update, large tech companies reduced new graduate hiring by 25% in 2025. However, the "Work AI Index 2026" suggests that botsitting will create substantial employment opportunities for many graduates.
The first reason is that botsitting has an extremely low entry barrier but offers broad industry exposure. Many people associate AI-related jobs with coding, parameter tuning, or model training. Botsitting is entirely different. It doesn't require algorithmic knowledge, just basic human judgment. Can you tell if the text AI wrote is nonsense? Can you determine if a data analysis report's conclusions align with the preceding data? Can you identify seemingly professional but hollow terminology AI inserts to meet word counts? Any normal human with a basic university education possesses these capabilities.
The second reason is that this generation of graduates are AI natives. A reality for the class of 2026 is that their theses were likely completed with the help of ChatGPT, Claude, or DeepSeek. In contrast, many experienced senior employees still use AI merely as a faster search engine—posing a question and accepting an answer. This generation of graduates is different. Through repeated experiences of being misled by AI, they have developed an intuition: when to trust AI output, when to double-check, and when AI appears to be answering but is actually evading. More importantly, this generation inherently understands "prompt engineering," a lesson learned through academic trial and error. They also don't view repeatedly editing AI output as "extra work"; it's simply part of their normal workflow: have AI generate a draft, make substantial revisions, have AI polish it, then review it themselves. In other words, while botsitting feels like an "additional burden" to senior employees, for recent graduates, it's the "normal way of working" they've learned.
The third reason is the moderate labor intensity. As mentioned, it requires about 6.4 hours per week. For those graduates proficient with AI, this time can be even shorter. NACE's survey also shows that 70% of employers already use skills-first hiring, a 5% increase from last year. Skills-first hiring means evaluating candidates based on their ability to perform the job, not their school or major. Another often overlooked point is that botsitting isn't a dead-end job; it has a clear career progression path. Companies like Scale AI and Surge AI hire their best trainers directly into full-time roles as quality analysts and project managers. Starting wages range from $10 to $20 per hour, with full-time annual salaries reaching £40,000 to £60,000. For a recent graduate, promotion is based on quantifiable performance—catching more and more accurate errors than peers leads to advancement.
Will Botsitting Become a Permanent Fixture?
A more fundamental question remains: Is botsitting a temporary phenomenon of a transitional phase, or will it become a permanent role? To understand this, we can use a reference point: data annotation.
A decade ago, "data annotation" was almost unheard of. Take the task of teaching a model to recognize a cat: the solution was to hire people to label images one by one—this is a cat, this is not a cat, this is a cat even though half its face is obscured. Fortune Business Insights once published data showing that around 2015, China had only tens of thousands of data annotation practitioners, with a total market size of about 500 million yuan. By 2020, China's data annotation market reached 3.1 billion yuan. By 2025, it surpassed 10.5 billion yuan. Globally, the data annotation tools market was valued between $1.7 billion and $3.6 billion in 2025, projected to grow to $14 billion to $38 billion by 2034, with a compound annual growth rate over 26%. If models are getting stronger, why do they need more people for annotation? The answer is simple: stronger models handle more complex tasks; more complex tasks require more精细 training data; more精细 data demands more irreplaceable human judgment. Every step forward in AI raises the requirements for data annotation. Botsitting is following the same path as data annotation and will go even further.
First, the core bottleneck for botsitting is not technology but organization. In March 2026, Harvard Business Review published a significant article titled "The 'Last Mile' Problem Slowing AI Transformation." Authors including Harvard Business School's Karim Lakhani, Microsoft's AI at Work lead Jared Spataro, and Harvard D³ Institute's Jen Stave concluded that models can become infinitely powerful, but as long as they don't understand a company's internal logic, someone must be assigned to clean up after them. This isn't a problem of models being inadequate; it's that an information gap exists between models and our real world. Every enterprise has unique context, such as unwritten rules and corporate culture. These elements aren't in any public dataset, and AI can never learn them on its own. But for AI to function within an organization, someone must translate, supplement, and correct this information for it. This is precisely the core value of botsitting. It's not compensating for AI's inadequacies but bridging the information gap between AI and the organization. As long as companies are unique, this gap will always exist, and botsitting will always be needed.
Second, AI's working method inherently requires human oversight. A 2026 Forbes article, "Is AI Replacing Jobs? New Data Suggests It May Be Increasing Workload," posits that AI won't make jobs disappear; it will shift them. The article argues, "Once AI enters real workflows, someone must supervise, edit, verify, and provide a safety net." In 2025, a federal court in Mississippi encountered an issue where lawyers from Butler Snow law firm submitted court documents containing AI-fabricated legal precedents without review. Judge Britton Manasco was furious, stating in the ruling that "fabricating legal authority constitutes serious misconduct." A 2026 survey by the National Law Review of 85 legal professionals concluded that the future competitive differentiator for lawyers might not be which AI they use but their ability to validate outputs. "Human-AI collaborative workflows, quality control, and defensible review processes will become core competencies in the legal industry, not optional safeguards." Morgan Stanley launched an AI tool called Morgan Stanley Debrief in 2024; by the end of 2025, 98% of wealth management advisors were using it. However, Morgan Stanley has a rule: advisors must "review and adjust AI-generated outputs before finalizing" when using AI-generated meeting summaries and investment recommendations. Simultaneously, the Financial Industry Regulatory Authority's 2026 annual regulatory report, released in December 2025, specifically added a regulatory framework for "AI systems that autonomously execute tasks," explicitly requiring that once an AI system can take actions within a brokerage workflow rather than just generate content, the firm's supervision, record-keeping, and governance obligations must be substantially upgraded. In plain terms, AI can help write documents, conduct analyses, and even make decision recommendations, but the final signature must come from a living person. Moreover, the person signing must be able to explain and take responsibility for what they sign. Authority can be delegated, but responsibility cannot be outsourced, so botsitting is destined to persist.
Third, empirical evidence is already appearing. The World Economic Forum's 2025 "Future of Jobs Report" stated that AI and big data specialists are among the fastest-growing roles towards 2030, but it also noted that non-technical positions like AI governance and AI strategy are growing rapidly in parallel. A report from AI talent platform Mercor shows that global demand for human evaluators and trainers is growing at 25% to 35% annually. The report also mentions that most of these roles are fully remote, don't require a technical background, but highly value domain expertise and judgment. As AI transitions from a personal efficiency tool to an organizational infrastructure, botsitting evolves from "something anyone with free time handles" to a function requiring dedicated personnel.
Comments