The Self-Checkout Supervisor Thesis

Or: A Question About AI and Work That Might Be Worth Asking*

Let us begin with what is actually happening to people.

In 2025, nearly 55,000 US job cuts were directly attributed to AI. Amazon eliminated 14,000 corporate roles. Microsoft cut 15,000. Salesforce reduced its customer support workforce by 4,000. Between the final quarter of 2025 and January 2026, approximately 245,000 further technology sector employees were laid off globally. These are not abstractions. They are people who had mortgages, routines, professional identities, and Tuesday evening plans.

Whether some of these cuts constitute what Forrester calls "AI-washing" — attributing financially motivated layoffs to AI that doesn't yet exist — is a question Harvard Business Review explored recently. The impact on the individuals affected, of course, is identical either way.

Against this backdrop, the prevailing conversation about AI and employment has settled into two well-worn grooves. In one, AI eliminates everyone's job and we all learn to forage. In the other, new and wonderful roles emerge that we cannot yet imagine. Both framings share an assumption so fundamental that it has become invisible: that you will continue working for one employer at a time, and that whatever productivity AI unlocks will belong to the firm.

A third possibility occurred to me recently, and I suspect it has occurred to others too, though I haven't found it articulated quite this way. Rather than claim some grand insight, I'd like to describe what I noticed and see whether it holds up.

The Supermarket

There is a moment in every supermarket visit — usually around the third time the machine insists there is an unexpected item in the bagging area — when you become aware of the supervisor. Where twenty checkout assistants once operated twenty tills, four or five people now oversee forty self-checkout machines. The workers have not vanished. Their role has shifted from scanning tins of beans to managing the machines that scan tins of beans.

Here is the bit that caught my attention: each supervisor is effectively using their time multiple times over. They are not working faster. They are not working harder. They are simultaneously present for forty transactions that used to require twenty individual humans giving their undivided attention. One person's time is being applied concurrently across many workflows.

Now consider what happens when this pattern reaches knowledge work.

The Multiplication

A marketing manager currently earns £80,000 from a single employer. They spend their time across strategy development (which they're brilliant at), internal meetings (variable value), creating presentations (disproportionate time consumed), and handling admin (actively annoying). Their time, like everyone's, is finite. They sell it to one employer because that has always been the deal.

But what if they deployed AI agents to handle the execution layer of marketing — the content drafting, the campaign assembly, the reporting — and shifted their own contribution to creative curation, quality judgement, and strategic oversight? Oversight that, crucially, does not require continuous attention. It requires availability, expertise, and the authority to intervene when the machines lose the plot (which they do, frequently).

If they can do this for one employer, why not for six? Or eight?

Each employer pays £15,000–20,000 for the oversight. The total: £90,000–160,000. Each employer pays less than a full-time salary. The worker earns more. And — this is the part that matters — the worker is not dividing their time between clients the way a fractional executive does, working Mondays and Tuesdays for Company A and Wednesdays for Company B. They are applying their judgement concurrently across all of them, intervening where needed, while their AI agents handle the continuous execution.

The person doing ten jobs simultaneously might, if they were so inclined, also find time for foraging. The point being that their time is no longer the constraint. Their judgement is.

Why This Wave Is Different

For the entire history of employment, the fundamental dynamic has been this: the worker has finite time and must sell it. The employer, who owns the expensive capital equipment — the factory, the assembly line, the bank of self-checkout machines — sets the terms. When supermarkets automated checkouts, the investment ran to hundreds of thousands of pounds. Individual workers could not install their own. The returns flowed to the firm. The worker's role simply changed shape within the employer's orbit.

AI breaks this pattern. The same subscription available to an enterprise is available to any individual for £20–200 per month. For the first time in an automation wave, the individual can own the means of amplification. Your supervision capability becomes portable. You are no longer dependent on a single employer's technology stack.

Starting from no AI development experience in September 2025, using only consumer-priced tools costing under £100 per month, I developed structured methodologies for AI-assisted product development, measured the economics, and shipped working software — documented in two prior pre-prints. This is not intended as a boast; it is intended as evidence that the barrier to entry is genuinely low and the software itself can train the user to use it.

The inversion of the traditional power relationship — where it was always the employer who owned the capital and therefore captured the productivity gains — is, if it holds, genuinely novel. For those who can participate.

The Very Large Caveats

That last phrase — for those who can participate — deserves its own section, because it is doing a lot of work.

This model applies to knowledge work. A paramedic cannot agentify themselves. A plumber cannot supervise AI agents that fix your boiler from across town. A care worker's value is irreducibly physical and present. The thesis describes a possible future for a specific category of work, not a universal transformation.

Even within knowledge work, the model presupposes expertise. The supervisor needs genuine domain knowledge that AI currently lacks — the taste, the contextual judgement, the ability to know when something is technically correct but somehow completely wrong. Not everyone has this, and crucially, the entry-level positions where people traditionally developed it are being eliminated at pace. If the model requires ten years of domain experience to work, and the first five years are being automated away, the pipeline of future supervisors may dry up.

The paper is honest about the risk of a two-tier workforce: those with the expertise and digital literacy to curate AI agents across organisations, and those whose value was primarily in task execution who find themselves displaced. Fewer people doing the work is not, in itself, a cheerful outcome. It becomes a potentially positive one only if those fewer people genuinely capture the value — and only if society figures out how the rest participate.

There are also real questions about liability across multiple clients, whether employers will accept shared supervision, and whether AI platforms will keep tools accessible or lock them behind enterprise pricing (the history of technology platforms is not encouraging on this point).

The Paper

"The Self-Checkout Supervisor Thesis: AI Agents, Portfolio Employment, and the Future of Work" is the third Syntropic Works research pre-print, following Dr StrangeDev on constraint-based AI collaboration methodology and Economic DORA on measuring AI-assisted development economics.

It includes seven testable hypotheses, a critical evaluation that takes up roughly a quarter of the paper, and — because any thesis worth proposing is worth subjecting to scrutiny — an explicit acknowledgement that it may prove to be accurate, transitional, niche, or simply wrong.

Read it: DOI: 10.5281/zenodo.18474685

The question for the next generation may not be "what is my profession?" but "what is my architecture?" — what portfolio of AI agents, curated with what methodology, applied across what set of organisations, generates the most value?

But before that question can be asked, a prior one must be answered: how do people develop the expertise and taste that makes the architecture worth building?

Both questions, at least, seem worth asking.

Responses, disagreements, and testable counterarguments are welcome.