The office job you know is on its way out

You grab a coffee before hopping on a Teams call to discuss why a colleague put 1524 units in cell B4 of your shared Excel spreadsheet when it should clearly have been 1759 (how could they be so careless?). Then in the afternoon, you complete your third rewrite of an offer for a new prospect only to get a message just before you finish saying that they’ve changed the whole scope of the project, again.    

This will sound all too familiar for lots of knowledge workers. But in five years they will probably look old-fashioned, and we may even miss them. Work has changed like this before. Most people used to work on farms, then in factories, and then in offices. Each time, new tools took over old tasks and the jobs changed with them. AI agents are no different and are now doing the same for office work. 

Capable and unreliable  

AI used to be a tool for answering questions, but coding has become one of the clearest examples of how it can now do real work. Just two years ago, AI coding tools were largely limited to autocomplete and small scripts, whereas Google now says AI writes up to 75% of its new code, which is up from 50% compared with late 2025.  

AI still makes surprisingly silly mistakes despite this progress. One model famously couldn’t say how many Rs are in “strawberry”, while another told someone to walk to a car wash because it was only 50 metres away, apparently overlooking the fact that they would have to walk through the brushes and sprayers. The labs usually fix each widely shared mistake, only for a new one to turn up shortly afterwards. 

AI researcher Andrej Karpathy calls this “jagged intelligence”, which means a model can solve a very hard problem and then get something wrong that a child would get right. That means you can’t simply take a person out of a process, put an AI in their place and assume the result is fine cecause there’s no way of knowing in advance which you will get. 

Deciding what to hand over  

Whether an agent can be trusted with a task comes down to two questions: what does a mistake cost, and what does checking it cost? 

The cost of a mistake is high for something like a hallucinated citation in a court filing, which can lead to fines and reputational damage, and low for a first draft of meeting notes with a few errors in it. The cost to verify an agent’s output also works differently, since math is inexpensive and can be checked programmatically, while a business strategy is expensive to check because you need real experience to judge whether it’s any good. The higher a task scores on either question, the more a person needs to review it and step in. 

It also helps to remember that what we see as one task can often involve several. Customer service is a good example, because routine first-level interactions are easy to check and have clear escalation paths, whereas harder second and third-level interactions cost much more when they go wrong and are much harder to check.  

Gartner predicts that by 2029 around a third of the employees laid off because AI replaced them will need to be rehired, and Ford, IBM and Klarna have already walked back AI-related layoffs and brought back engineers, HR staff and customer service workers, which suggests companies are finding out they still need people to catch AI’s mistakes.  

Why people have become more valuable 

Cash machines spread through American banks in the 1970s, and there are now more than 400,000 in the US. Although many people expected bank tellers to disappear, the number of tellers went up, and so did their salaries. Radiology is a more modern example, because AI can now read some scans better than doctors can, yet hospitals can’t fill their radiology roles and demand for radiologists has never been higher.  

Nobel-winning economist Michael Kremer’s O-Ring theory describes how modern knowledge work is multiplicative rather than additive. In other words, one faulty step drops the value of the whole output to zero.   

In automation, a single weak link sinks the whole output, unless a human catches it. That makes the people left in the process more valuable, since each of them needs to check a lot more work, and jobs only disappear when every step in the chain is automated, which is hard to picture while AI keeps making these mistakes. 

What you can do with the time you save 

Whether any of this pays off depends on what people do with the time they get back. If they spend it on the bottleneck tasks, like building relationships with a potential client, understanding what a client really needs and making judgement calls, the finished work gets better and more of the routine work can be automated next time.  

That won’t happen by itself, so leaders need to work out what AI agents can do, deliberately hand that work over, and restructure processes so people can move onto higher value work rather than just watching over the machines. 

Five years from now  

Imagine a dozen AI agents running at the same time, drafting offers, qualifying leads and clearing support tickets. The tedious and repetitive work disappears, but your cognitive load is higher as you try to keep track of what each agent is doing, reviewing their output and making sure nothing falls through the cracks.  

The sensible response is to look after your mental bandwidth, make sure the agents work for you rather than the other way round, and spend the time you win back on the things only you can do.  

If you don’t, we’ll be looking back in a few years and wish we could argue about cell B4 again.