Work as it is known today is quickly becoming a thing of the past. Many knowledge workers start their day with coffee and a call where the first ten minutes are spent debating why a colleague entered 1524 units in a spreadsheet cell when it should be 1759. Later, they might rewrite a sales offer only to find the project scope has changed again. This is a typical day for many office workers, and in just five years, it may seem outdated — even if it's missed. AI agents are reshaping desk jobs, much like tools have done in the past. For example, AI coding has evolved from basic autocomplete features to the point where engineers at companies like Anthropic say AI now writes up to 90% of their code. Some engineers now code less by hand. However, these systems are not perfect. They may fail to count letters in a word or suggest walking to a nearby car wash. While labs work to fix these issues, new ones often arise. AI researcher Andrej Karpathy calls this "jagged intelligence," where models can solve complex problems but struggle with simple tasks. This means you can't just replace a human with an AI agent and expect perfect results. The reliability of AI depends on two things: the cost of making an error and the cost of checking that work. For instance, a wrong citation in a legal document has serious consequences, while a rough draft of a meeting summary is less critical. Checking the accuracy of the work also varies — math is easier to verify than business strategy, which requires deep expertise. Customer service is a good example of a multifaceted task. Routine interactions are easier to automate because they have clear rules and escalation paths. But more complex interactions require human judgment, as mistakes can be costly and difficult to verify. Klarna, a financial services company, found this out when they relied too much on automation and had to rehire people after quality declined. Their CEO stressed the importance of ensuring customers know a human is available if needed. Historically, automation hasn’t reduced the number of workers but has instead increased their value. For example, ATMs led to more bank tellers and higher wages. In radiology, AI tools help read scans, but the demand for radiologists remains strong due to the complexity of the work. Economist Michael Kremer’s O-Ring theory suggests that modern knowledge work is multiplicative — meaning one error can ruin the entire project. This makes human oversight more valuable, as people can catch mistakes and ensure quality. Automation frees up time, and how that time is used determines its impact. If it’s used for high-value tasks like building client relationships or making strategic decisions, it can improve the quality of the work and raise the bar for future automation. However, this requires leaders to identify which tasks can be automated, assign them to AI agents, and restructure processes so people can focus on meaningful work instead of monitoring machines. In five years, a dozen AI agents might be working simultaneously, drafting offers, qualifying leads, and handling support tickets. The repetitive tasks would be gone, but individuals would need to mentally manage multiple tasks, providing feedback to ensure nothing slips through. It’s important to protect mental focus, making sure AI agents support humans rather than the other way around. The time saved should be invested in tasks only humans can do. Otherwise, in a few years, people might look back and wish they could argue over that spreadsheet cell B4 again.