The not-so-secret agents: AI is not taking your job… yet

There are still accountants…

The advent of the spreadsheet in the late 1970s led many to predict the demise of the accounting profession. Since this amazing new technology could perform number-crunching with complex formulae on a previously unheard-of scale, why would anyone hire accountants? Fast forward a few decades and there are more accountants and many more spreadsheets than ever.

The arrival of widely available and affordable AI systems with impressively broad capabilities like ChatGPT, Claude and Le Chat have led many to make similar predictions about the fate of white-collar work in the modern world. The parable of the spreadsheet is a common retort and one that rings true with most of our experiences with AI. Rather than replacing our jobs, these technologies are enhancing them and in ways that are making us more productive in the process. Indeed, GitHub estimates that around 92% of US-based developers use AI coding tools and Microsoft CTO Kevin Scott predicts that within five years 95% of all code will be AI-generated. Yet big tech companies still employ armies of software engineers. And that's before we talk about how much we're not seeing lengthy queues of white-collar professionals at job centres. This is a classic case of Jevons’ paradox – increased efficiency in using a resource can lead to greater overall consumption of it. Exactly the same thing happened with coal in the nineteenth century and we’re now seeing it play out again in the age of artificial intelligence.… but you'll never meet a scrivener.

Yet in the longer run, maybe the Luddites had a point? Ensuring that all relevant parties have copies of any legal contract is an important matter, so important in fact that in the 19th century, the tedious job of accurately copying lengthy documents written in dense legalese was such a highly regarded task that an entire profession focused on it. This was the job of the scrivener. Today of course, you'll never meet anybody in this occupation --- the advent of the photocopier rendered it completely pointless.

Could AI push modern jobs the way of the scriveners? Up to now, it clearly hasn't and is showing little sign of doing so just yet. This might be about to change though with the advent of agentic AI.

The Rise of the agents

Roughly speaking, Agentic AI is when an AI system has the autonomy to act on the basis of its inputs. In particular, they operate independently, making decisions and taking actions without human oversight. The main reason ChatGPT has yet to take your job is that it's unable to do anything without a prompt from a human.

Consider, for example, the thermostat in your home. Typically, you program these in advance to turn on and off at fixed times and at a temperature of your choice. Some systems allow custom adjustments through an app, but the decision of what happens is still a headache for you and you alone. But imagine if your thermostat was controlled by an AI system in which your preferences are just one of the inputs. The system could also take into account local weather forecasts, data from sensors about ambient atmospheric conditions or perhaps even energy prices and the state of your finances. That would be agentic AI.

This is all well and good but is this approach capable of feats more interesting than simply tweaking your heating?

One perennial problem that AI suffers is its tendency to hallucinate, i.e. make things up. These can range from the comical, such as recommending glue as a pizza topping, to costly, such as when Air Canada had to spend millions of dollars in compensation to customers who took advantage of a hallucinated refund policy. One approach to the problem that's recently been investigated is to get agents involved. Instead of inputting your prompt and having a single AI system produce an answer, what if we instead give it to three agents? They can then independently come up with an answer and then compare with each other to decide if any of them have hallucinated. If there is a hallucination, they can disregard that answer use the others to produce the output. As simple as this sounds this approach does seem to be effective and the result is an AI system that's much less prone to hallucinating.

But having at least some human oversight remains important. One recent study ran a simulation of a software company run entirely by AI agents. The result was laughably chaotic. For example, when unable to identify the correct colleague to contact on some matter, an agent merely picked a random colleague and renamed them to pretend they were the correct employee to speak to.

Another difficult problem that many of our clients are concerned with is scheduling. Having previously helped several organizations with their timetabling, needs from applications in the military to arranging sporting fixtures, we have much experience in the area. Often in scheduling problems, a machine learning model will produce a schedule. Next, human oversight is then required to check the suitability of such a schedule. Agentic AI can take the human out of the loop making these decisions for us.

AI, Robot

If you've ever tried asking your favourite AI system to make you a cup of tea, you'll likely have been disappointed with the result. Without the ability to physically interact with and manipulate objects in the world around them AI systems can be extremely limited. But agentic AI could power robots that perform complex tasks autonomously, such as search and rescue missions in disaster-stricken areas or operate machinery on a factory floor.

At present the costs of advanced agentic AI systems as well as advanced robotic systems remain prohibitively high. But costs are falling all the time. For example, in January 2025 Chinese company DeepSeek released their R1 model. This stunned the world by exhibiting performance typically associated with the western models having been trained on much cheaper chips and for a mere $6m (£4.8m). This is a fraction of the hundreds of millions used to train other models with comparable performance at the time. It's only a matter of time before agentic AI-powered robots are much more commonplace.

And this in turn creates bigger ethical questions. Whilst researchers are already concerned with, for example, facial recognition technologies that struggle with different skin tones [1], imagine a system built with these biases – but unlike current technology also has the capability to physically act on them. In light of this, responsible AI – an approach to building AI systems with a view to what’s best for both people and planet – has become more important than ever. Just think of self-driving cars: agentic systems that frequently make every decision involved in driving, like what route to take, how rapidly to decelerate when approaching traffic lights or even how to avoid hitting pedestrians (just think of the famous trolley problem).

The future of agentic AI holds immense potential to be incredibly powerful, but also incredibly risky to workers’ prospects and possibly to the entire fabric of wider society. Yet its potential to improve our lives is equally immense and should not be underestimated.

[1] Kashmir Hill “Your Face Belongs to US" Penguin 2023

Posted on: 16/06/25