Work as you know it will be a relic of the past

Work as you know it will be a relic of the past.

You wake up, pour a coffee and join a call. The first ten minutes are spent talking about why a colleague typed 1524 units in cell B4 of your shared spreadsheet. It should be 1759 units, obviously.

In the afternoon, you complete your third rewrite of an offer for a new prospect. You get a message just as you finish it. The scope has changed. Again.

This is a normal day for most knowledge workers. In five years, this type of day will seem archaic, and as strange as it sounds, we’ll probably miss it.

We’ve been here before after farming gave way to factory work, and then that turned into the knowledge work most of us do today. New tools automated the old jobs, and the nature of work changed with them. AI agents are doing the same to desk jobs.

Brilliant and useless at the same time

We recently worked out how to turn AI from a just question-and-answer machine into something that does actual work. Coding is a clean example. Two years ago, AI coding was mostly defined by autocomplete and small scripts. Today, engineers at Anthropic report AI now writes up to 90% of their code, with some no longer coding by hand at all.

So why do hallucinations and basic errors still happen? Every few weeks a new example makes the rounds: a model can’t count how many Rs are in strawberry; another insists you walk to the car wash because it is only 50 meters away (stepping through the suds, sprayers and rollers doesn’t seem like a great idea). The labs patch each one and a fresh embarrassment turns up the next day.

AI researcher Andrej Karpathy calls this jagged intelligence. Models crack extremely complex problems and then trip over something a child would get right. The lesson is that you can’t tell in advance which you’ll get.

In other words, you can’t extract a human from the process, drop in an agent and assume the output is fine.

What to hand over

What an agent can be trusted with hinges on two questions. What does a mistake cost, and what does checking it cost?

Error cost is the damage when the model gets it wrong. A hallucinated citation in a court filing is expensive, i.e. fines and reputational damage. A rough draft of a meeting summary not so much.

On the other hand, verification cost is how easy it is to check if what the agent produced is right. Mathematics sits at the inexpensive end of the spectrum, since proofs can be checked programmatically whereas a business strategy sits at the highest end, as you need deep expertise to properly evaluate it.

Gauge your tasks against those two questions. The higher a task scores on either, the more human verification, oversight and expertise is needed.

One point to keep in mind here is that what we see as one task can often involve several. Customer service, for example, might look like one task, but there is a big difference between routine first-level interactions, which have lower verification costs and clear escalation paths, and more complex second and third-level interactions, where the cost of errors and verification is much higher.

Klarna found this out very publicly, going hard with automation then hiring people back once quality dropped. CEO Sebastian Siemiatkowski's conclusion was that customers need to know a human is always there if they want one.

Why partial automation makes people more valuable

ATMs spread through banking in the 1970s, and there are now more than 400,000 of them in the US alone. The obvious prediction was fewer bank tellers. Instead, the number of tellers went up, and so did their wages.

Radiology is the modern version. AI tools now read some scans better than people do. Radiology departments aren’t sitting idle, however. Open positions can’t be filled and demand 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 remaining humans more valuable rather than less; they're now gatekeeping a far larger volume of higher-quality work. Demand only falls when the whole chain automates, and as long as jaggedness and hallucination are with us, that's tough to picture.

The focus effect

Automation frees up time, and where that time goes decides whether any of this pays off. If it’s spent well, it goes to the bottleneck tasks, like building relationships with a potential client, understanding what a client really needs and making judgement calls, it can improve the quality of the finished work and raise the bar for what gets automated next.

But to make this happen, leaders have to identify what agents can do, actively hand them over and then restructure processes so people can move to this higher value work rather than babysitting the machine.

Five years from now

Picture this: a dozen agents running in parallel, drafting offers, qualifying leads, clearing support tickets. The tedious and repetitive work is gone. Your cognitive load is higher because you’re having to mentally juggle all these tasks, giving feedback, while making sure nothing slips through the cracks.

Protect your mental bandwidth, keep the agents working for you rather than the other way round and spend what you get back on the things only you can do.

Otherwise, we’ll look back in a few years and wish we could argue about cell B4 again.

We've featured the best AI website builder.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit