AI Agents Will Reshape Knowledge Work, but Humans Remain the Fail-Safe, TechRadar Argues

The article opens with a familiar office scene: a knowledge worker starts the day with coffee and a call, spends the first ten minutes discussing why a colleague entered 1,524 units instead of 1,759 in a shared spreadsheet, and later rewrites a prospect offer only to learn the scope has changed again. Such days will look strange within five years, the author writes, and workers may miss them.

The author compares the shift to earlier transitions from farming to factory work and then to knowledge work. New tools automated old jobs and changed the nature of work, and AI agents are now doing the same to desk jobs, according to the article.

Coding is presented as an early example. Two years ago, AI coding mostly meant autocomplete and small scripts. Today, engineers at Anthropic report that AI writes up to 90% of their code, with some no longer coding by hand at all, the article says.

Yet the article stresses that AI systems remain unreliable in basic ways. It cites recurring examples: a model that cannot count the number of Rs in 'strawberry,' and another that tells a user to walk to a car wash because it is only 50 meters away, even though stepping through the suds, sprayers and rollers does not seem like a great idea. Labs fix one error and new ones appear, the author writes. AI researcher Andrej Karpathy calls this 'jagged intelligence': models can solve extremely complex problems and then fail at something a child would get right. The lesson, according to the article, is that a human cannot be removed from the process and replaced by an agent without assuming the output will be acceptable.

The article says decisions about what to hand to an agent should depend on two questions: what a mistake costs and what checking it costs. A hallucinated citation in a court filing is expensive because of fines and reputational damage, while a rough draft of a meeting summary is not. Verification cost is how easy it is to confirm the agent's output. Mathematics is at the inexpensive end because proofs can be checked programmatically; business strategy is at the highest end because evaluating it requires deep expertise. Tasks that score high on either measure need more human verification, oversight and expertise.

The author warns that a single task may contain several components. Customer service may look like one job, but routine first-level interactions have lower verification costs and clear escalation paths, while complex second- and third-level interactions carry much higher error and verification costs. Klarna learned this publicly, the article says, by pushing hard on automation and then hiring people back after quality dropped. CEO Sebastian Siemiatkowski concluded that customers need to know a human is always there if they want one.

The article also points to ATMs and radiology as evidence that automation does not necessarily reduce human roles. ATMs spread through banking in the 1970s, and there are now more than 400,000 in the United States alone. The obvious prediction was fewer bank tellers, but the number of tellers rose, as did their wages. In radiology, AI tools now read some scans better than people do, yet departments are not sitting idle: open positions cannot be filled and demand has never been higher.

Citing Nobel-winning economist Michael Kremer's O-Ring theory, the article says modern knowledge work is multiplicative rather than additive. One faulty step can drop the value of the whole output to zero. In automation, a single weak link can sink the whole output unless a human catches it. That makes the remaining humans more valuable rather than less, because they gatekeep a much larger volume of higher-quality work. Demand would fall only if the entire chain were automated, which the article says is hard to picture as long as jaggedness and hallucination remain.

The author argues that automation frees time, and where that time goes determines whether the shift pays off. If it is spent on bottleneck tasks, such as building relationships with potential clients, understanding what a client really needs and making judgement calls, it can improve the quality of finished work and raise the bar for what is automated next. To achieve that, leaders must identify what agents can do, actively hand over tasks and restructure processes.

Editor's Summary

TechRadar argues that AI agents will transform knowledge work but cannot yet replace human judgement. Because model errors can be costly or hard to verify, the article says people should oversee higher-risk work while automation takes on more routine tasks. It cites coding, customer service, ATMs and radiology to argue that partial automation can make remaining human roles more valuable.