---
title: "Which Jobs Will Still Pay in 2028? AI Time Horizons Explained"
url: "https://goremy.ai/blog/articles/highest-paying-jobs-2028-ai-bottleneck-meter-time-horizon"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-17T09:20:10.704Z"
updated: "2026-11-17T09:20:17.317Z"
topic: Will AI Replace My Job
description: How METR time horizons, bottleneck economics and falling labor share point to which careers may keep paying as AI agents take on longer tasks.
---

# Which Jobs Will Still Pay in 2028? AI Time Horizons Explained

> How METR time horizons, bottleneck economics and falling labor share point to which careers may keep paying as AI agents take on longer tasks.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-17

## Which careers will keep paying as AI agents handle longer tasks?

The careers most likely to keep paying are the ones that sit at a bottleneck: work that stays scarce while AI makes everything around it cheap. Think roles that deploy and manage AI agents inside businesses, jobs gated by licensing and liability such as medicine and law, and positions where you own the assets that produce the work. Roles that mainly produce routine output that agents can now produce in volume are the most exposed. This is an argument built on trends, not a certainty, and it is worth understanding the logic behind it.

## TL;DR

- **AI agents are working longer without help.** METR, a research group, measures how long a task an AI can finish at human-level quality, and its early-2025 data showed that length doubling roughly every seven months.
- **Pay follows scarcity, not effort.** When something becomes easy to produce, its price falls, even if the work is still valuable. Programming is the example people point to.
- **The best-placed workers act like "centaurs."** They let AI do the volume and keep the judgment for themselves.
- **One person can now direct many agents.** A claim attributed to Anthropic says its agents log six to seven hours of work for every human hour inside the company.
- **Owning assets matters more as labor gets cheaper.** The US labor share of the economy was cited at 54.1% for the first quarter, reportedly the lowest since measurement began in 1947.
- **Agent operators are one clear bottleneck role.** The creator whose argument we draw on says median pay for forward deployed engineers is $200,000 to $250,000 a year, though you should check current listings yourself.

## What is the METR time horizon, and why does it matter?

METR is a lab that tests how long AI agents can work on their own while still producing human-level results. Its "time horizon" measure works like this. Give a model a set of tasks, each with a known human completion time. Find the task length at which the model succeeds about 50% of the time. That length is the model's time horizon.

Why care? Because it measures something closer to real work than a quiz score. Answering a question takes seconds. Fixing a software bug, drafting a report or running a research project takes hours or days. A model that can only handle a few minutes of work needs constant supervision. A model that can handle days of work can be handed a project.

In March 2025, METR reported that the time horizon was doubling about every seven months. At that point the horizon was measured in minutes up to roughly an hour. The creator we draw on argues that current models now run autonomously for days, and that researchers inside the major labs expect the doubling to continue or speed up. Treat that as a forecast, not a law. Trend lines can bend. But the direction is consistent enough that planning around it is reasonable.

## How does the bottleneck idea decide who gets paid?

The economic rule is simple. In any process with one constraint that limits everything else, the constraint captures the reward. Whatever is scarce gets paid.

Programming is the recent case. For years, companies and consumers wanted more software than there were people to write it, so programmers earned a lot. Once AI could write ordinary code quickly, the scarcity moved. The argument is that pay for routine programming has dropped, while pay holds up for people building very complex systems or working directly on AI models and agents.

Doctors and lawyers show the same rule from the other side. There are always more people in an emergency room than doctors to treat them, so doctors are the bottleneck and are paid accordingly. The legal field works similarly.

Notice what this does and does not say. It does not say your work is worthless. It says that when agents produce huge amounts of similar output, the price of that output falls because there is so much of it.

## What is a centaur, and why does it matter for your job?

A centaur is a human working with AI. The AI supplies speed and volume, like the horse body. The human supplies judgment, like the head.

The numbers behind this are striking, if you accept the claim. The creator reports that inside Anthropic, AI agents now spend six to seven hours on tasks for every one hour humans spend. That does not mean agents are slower. It means they are doing six to seven times the work. In AI research, which he calls one of the most complicated fields there is, one person running agents well might therefore equal six or seven people.

If that multiple keeps rising alongside the time horizon, one person directing agents could eventually do the work of many. For a worker, that cuts two ways:

- A person who learns to direct agents well becomes far more productive and more valuable.
- A team of ten doing routine work may need far fewer people if one person can supervise the agents doing it.

The practical question is which side of that line your tasks sit on.

## What does falling labor share mean for ordinary workers?

Labor share is the portion of the US economy paid out to workers as income. The source cites a first-quarter reading of 54.1%, described as the lowest since the government began tracking it in 1947. Verify the latest figure with official statistics before quoting it, but the long-running concern is real: more of the economy flows to capital than to wages.

The fourth premise follows from that. If labor becomes cheap and abundant, the value shifts to whoever owns the assets that produce the labor or the service. The creator cites the book *The Intelligence Explosion*, which compares the situation to "oil states," countries whose money comes from the ground rather than from citizens, and which have sometimes ignored citizens' needs as a result. The worry is that governments and large organizations lose their incentive to invest in people if AI can replace the labor.

You do not have to accept the political argument to take the practical point. Owning something, whether a business, a product, a client base or software, gives you income that does not depend on selling your hours.

## Which roles fit these four premises?

The source walks through the first of six careers in detail: the agent operator, also called a forward deployed engineer. This is a person who goes inside a business, finds work that agents can do, and builds systems to do it. The creator says the median pay in the US is roughly $200,000 to $250,000 a year, with the top 10% near $400,000. He also says Anthropic announced a plan to train 10,000 of these engineers by the end of 2027 and called them the hardest talent to find. Those figures come from one creator's account, so confirm them against current job postings.

The logic matches the bottleneck rule. The technology is becoming plentiful. The scarce part is the person who can get into a company, understand how it actually works, and make agents useful there.

From the source and the four premises, these categories look best placed:

| Category | Why it fits |
|---|---|
| Agent operators and forward deployed engineers | Scarce people who connect abundant AI to real business processes |
| Licensed, high-liability professions (medicine, law) | Demand outruns supply, and accountability stays with a person |
| Owners of businesses, products or client relationships | Value sits in the asset, not the hours |
| People who direct many agents in a complex field | The centaur multiple applies to their output |

The source did not finish its list in the material available, so these categories are my reading of its premises, not the creator's full six.

## What does an agent operator actually do all day?

The creator describes a straightforward process:

1. Meet with department heads or business owners and have them explain their processes.
2. Sort the work into tasks an agent can do entirely alone and tasks that need a human working alongside an agent.
3. Hand the first group to agents completely.
4. For the second group, train the people to use agents as thought partners.
5. Focus on growth work first. He uses the acronym RACE: reach, acquisition, conversion, expansion. His view is that small and mid-size companies, roughly under 100 staff, usually struggle more with the front end than the back end.
6. Map each step of the process. With permission, that can mean shadowing someone or capturing screenshots of their screen every few seconds, then feeding the log to an AI model to turn it into written procedures (SOPs).

The tools are modest. He mentions a coding agent such as Claude Code or Codex at about $20 a month, plus voice input and a meeting transcript. His point is that building is no longer the hard part. Getting into the business and getting people to explain how they work is.

## What this means for you

**If you do routine, high-volume knowledge work** (first-draft writing, basic reporting, standard code, data entry), assume agents will produce more of that output at lower cost. Move toward the part of your job that involves judgment, accountability and relationships.

**If you manage people or processes,** you are in a good position to become the centaur. Pick one recurring process this month. Write down every step. Try handing the first draft to an AI tool and review the result.

**If you are early in your career,** consider roles that sit between businesses and AI tools. You do not need an engineering degree to start. You need to understand how a department really works, and that knowledge is the scarce part.

**If you have savings or side income,** the labor-share argument says to think about ownership, even at small scale: a small business, a niche product, a service with repeat clients.

Whatever your situation, keep your estimate humble. Time-horizon trends are extrapolations. They can slow, and adoption inside real companies is usually slower than lab results suggest.

## Can I build my own process tracker instead of buying another subscription?

If you start mapping business processes the way an agent operator does, you will quickly need somewhere to keep them: a list of processes, owners, steps, which tasks are agent-ready and which need a person. A spreadsheet works for a week, then it sprawls. With [Remy](https://goremy.ai), you can describe that tracker in plain English (fields, status stages, who can see what) and get a deployed full-stack app with a database and sign-in, which you own as standard TypeScript in a git repo. Remy is $99 a month, or $79 billed annually, plus AI usage billed at cost.

## Frequently Asked Questions

### What is the METR time horizon?
It is the length of task, measured by how long a human would take, that an AI model completes successfully about 50% of the time. METR uses it to track how long agents can work independently. Its March 2025 findings showed the horizon doubling roughly every seven months.

### Will AI replace my job by 2028?
Nobody can say for certain. The argument in this article is that jobs built on abundant, repeatable output are most exposed, while jobs at a bottleneck are better protected. Your exposure depends on how much of your work is routine and how much rests on judgment or accountability.

### Why do doctors and lawyers keep getting paid?
By the bottleneck rule, demand for their services exceeds supply. An emergency room always has more patients than doctors. As long as that holds, and licensing and liability keep a human responsible, pay tends to hold.

### What is a forward deployed engineer?
It is someone who works inside a client business to find tasks AI agents can take over and then builds the systems to do it. The creator cited here says the median US pay is about $200,000 to $250,000 a year. Check current job listings for real ranges.

### What should I do first if I'm worried?
Pick one repeating task in your job and [try doing it with an AI agent](https://goremy.ai/blog/articles/hormozi-ai-jobs-bring-your-own-agent), then review the output critically. You will learn quickly which parts need your judgment. That is the part to build your career around.
