The junior work we think we're saving

Published 30 September 2026

I lead payment teams and, among other things, act as hiring manager. Since many people in companies started using AI, I've heard several iterations of the same idea: AI takes over the junior work and we save money. We all instinctively have an intuition of its fallacy (or at least I hope we do :D), but it's tempting to ignore it given the quick returns.

Christian Catalini's essay The Economics of AI gave me the words for that intuition. It sums up "Some Simple Economics of AGI", a working paper he wrote with Xiang Hui and Jane Wu, and despite the title it's about how AI is already moving the value in everyday work, not about machines becoming generally intelligent one day.

Execution at machine speed, verification at human speed

The core idea is simple: producing work is becoming almost free, checking it isn't. An agent can write a specification, the code and a launch campaign overnight. Lovely, until someone has to find out whether it invented a fact, opened a security hole or took a clever shortcut around its goal. Production runs at machine speed. Checking runs at human speed, or at the speed of the real world when the answer only arrives years later. Catalini calls the widening distance between the two the Measurability Gap.

Why we take the risk anyway

So why not slow down wherever checking is hard? Because for the company that doesn't, the maths works. The savings arrive now and land on the company, while the cost of a hidden failure arrives later and often lands on someone else: customers, or the wider market. Skipping the checks is a rational choice for one company, even when it's a bad deal for everyone.

The cost that worries me most is the slowest one, and it brings me back to that idea about juniors. Experts learn their judgement through years of routine work, so if agents take over the junior work, in ten years there are fewer people able to supervise the agents. As someone who hires engineers, I keep asking myself: if juniors no longer do the work, how do they learn to check it? The essay's answer is to rebuild the apprenticeship on purpose, with simulations in which people decide before they see the answer.

The data worth keeping

Most companies treat their archive of finished work as the treasure, but finished work only teaches a model what to produce, and models see more of it every year. The data worth keeping is the record of what to reject and why: blocked deployments, near misses, postmortems, expert overrides. This one felt very familiar from payments, where moving money is the easy bit and the real knowledge sits in the exceptions, the disputes and the chargebacks. A competitor can copy a workflow, but not the years of cases that taught you when to trust it.

A second model is not a second opinion

AI will have to help check AI, as there is simply too much output for people to read. The catch is that a checker built on similar data and design shares the blind spots of the model it checks. People make independent mistakes, which is useful; agents built on the same foundations can make the same mistake a million times and politely approve each other's work. Real verification needs different methods, real-world outcomes and a person who stays responsible for the result.

Who stands behind it

I'll admit liability was the thread I was slowest to see (it turns out it runs through the whole essay). If anyone can produce a polished result, the customer stops caring which model made it and starts caring who guarantees it. In high-stakes markets the essay expects AI firms to be judged more like insurers than like software vendors, which favours banks and insurers, with their licences and balance sheets. The policy argument follows: as long as failures cost someone else, companies buy too little verification. Make the deployer pay, and verification tools get built without anyone prescribing how AI must be designed.

When more users make a network worse

The idea I liked most connects to what I'm studying. Card schemes are the textbook two-sided market: more cardholders attract more merchants, and the other way round. That network effect quietly assumes every participant is real. Agents can fill a marketplace with listings and reviews, and because they make the same mistakes, more activity can make a network less trustworthy rather than more useful. The network effect runs in reverse.

flowchart LR
    A[More agent activity] ==>|the same plausible mistakes| T[Less trust]
    T ==>|the best contributors leave| V[A less valuable network]

What counts then is how much participation people can trust, and a platform's history of resolved fraud cases, disputes and chargebacks becomes its real advantage.

What I take away

None of this is an argument against using AI: the essay's goal is "the maximum automation you can verify, insure, reverse, and still understand". So, back to the junior work and the money we save. For engineering teams, the saving is real only if we count the work we can stand behind rather than the work produced, keep the record of what we rejected and why, and give juniors a way to learn to check work they no longer write.