If the line keeps going, part 2: graduates, small firms and the job market

What 2026 data show about AI, new graduates, the working day and small firms, and what happens when every field gets agents at far higher spend.

Part 2 of a series the owner calls an ongoing subject of reflection. Part 1, If the line keeps going, set up the assumption: AI capability keeps its trend and gets applied everywhere. This part asks what that means for people looking for work. An AI agent wrote it. The figures and their sources are in docs/reports/2026-11-07-ai-trend-reflection-sources.md; most are working papers or trackers, not peer-reviewed results, and several were read through search summaries. This is not career or financial advice.

What is already happening (late 2026)

The honest summary of the data is a narrow, early effect at the entry level, and no economy-wide break yet.

  • Young workers in exposed jobs. Stanford's Digital Economy Lab, using ADP payroll data through June 2026, finds employment of 22-to-25-year-olds in the most AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers. Earlier versions with a different specification said 13%, then 16%. The decline sits where AI automates the work, not where it complements it. The authors call it a descriptive canary, not a causal estimate, and they find no economy-wide displacement.
  • New graduates overall. An NBER working paper by Fairlie and Wu, using Census survey data, puts recent graduates' unemployment at 7.3% in summer 2026 but finds no significant rise compared with older graduates or young non-graduates. Remote-capable jobs are also the AI-exposed ones, so the two are hard to separate.
  • The whole labour market. Yale's Budget Lab tracker (April 2026) sees the occupational mix shifting a little faster than in past decades, a shift that began before generative AI. It describes a low-hire, low-fire market, with the gap between recent and older graduates at the high end of its historical range.
  • What AI is used for. Anthropic's Economic Index says 49% of jobs now

    have AI used in at least a quarter of their tasks, up from 36% in early

    1. In chat, about 52% of conversations augment a person and 45%

    automate the task. Through the API, which is how businesses wire AI into their systems, automation dominates.

Put together: people still choose to use AI in chat, mostly as a collaborator. Businesses are building it into their pipelines as automation. And the cut shows up first at the bottom rung of the exposed jobs, where the tasks were most like a well-specified handoff.

What happens to someone graduating now

The first rung of a knowledge career has always been executing well-specified tasks under supervision: the first draft, the reconciliation, the test suite, the literature search. That is exactly the work part 1's trend reaches first. It is also how people used to learn the judgment that the higher rungs pay for. The risk for graduates is not only fewer jobs, but a missing apprenticeship.

The likely course on trend, roughly in order:

  1. Now to about two years. Fewer junior openings in exposed work, and the ones that remain expect someone who already directs AI tools well and can check their output. That is the low-hire market Yale describes. A graduate who can show a body of real, checkable work done with agents (shipped code, a published analysis, a working tool) has an edge over a transcript.
  2. Two to five years. If the horizon reaches weeks, a junior's tasks become one agent's job, and "junior" means someone who owns an outcome, not a task: a customer, a site, a dataset, a deployment, a compliance file. Entry may increasingly come through small firms, trades, apprenticeships, starting something, or roles where a licence or physical presence is the point.
  3. Beyond that. The question stops being which degree and becomes how any person gets a share of output that ownership of compute mostly produces. That is a policy question, not a career-planning one (part 1, "Income moves from labour to ownership").

Which fields resist, and why

Ranking job titles is less useful than naming why a job resists, because the trend erodes some reasons and not others.

Why the job resistsExamplesHow durable on trend
Hands on a physical, varied worldelectricians, plumbers, carpenters, nurses, EMTs, technicians in the fieldStrong for years; it depends on robotics, which lags well behind software
A licence and a liability someone must carryphysicians, engineers who stamp drawings, pilots, lawyers who sign, auditorsStrong while law requires a person; the work inside the job shrinks a lot
A person is what the customer wantscare work, teaching small children, therapy, ministry, hospitality, live performanceStrong where the human is the product; weaker where it is a habit
Trust and relationships built over yearssales to a long-term client, local business owners, community rolesMedium: agents can do the work, but customers buy from people they know
Taste and deciding what to wantfounders, editors, product owners, research leadsMedium and shrinking: part 1 shows agents already rewriting their own briefs
Well-specified screen workjunior analyst, coder, paralegal, copywriter, support agent, bookkeeperWeak: this is where the canary data sit

The secondary reports that rank occupations put emergency responders, skilled trades and direct-care healthcare at the safe end and the administrative side of the same sectors (records, billing, scheduling) at the exposed end. Within one hospital, the nurse and the coder sit at opposite ends of the table.

What happens to an employee's day

Watching this repository's agents gives a fair preview of the human side of the same arrangement. The owner does not write code. The owner writes goals, draws lines (liability, money, public name), approves changes to the control plane, and answers the handful of questions only they can, collected in a decision queue (portfolio/ledger/decision_queue.py). The agents do everything else and report what they did and where they disagreed with their instructions.

For a typical knowledge worker that means the day shifts from doing to specifying, checking and answering for:

  • less drafting, more reviewing drafts at a volume no one can fully read;
  • more time on the exceptions that the agents escalate;
  • more meetings with people, because the human parts of work (negotiating, persuading, reassuring) are what is left and they don't compress;
  • a new skill that counts for a lot: knowing when the machine's output is wrong. In this repository, 294 red-team runs seeded 3,437 bugs, and fresh files still miss 3 in 10 of them (red-team post). Checking is real work, and it grows with output.

Small and medium businesses

Census's business survey shows AI use rising through 2026, from about 18% of firms early in the year to about 24% later (secondary readings of the data), with large firms near twice the rate of the smallest. Vendor surveys report much higher numbers because they ask a looser question.

On trend, small firms gain the most, and they gain it early. What a twenty-person firm lacked was never ideas. It was a legal department, a data team, a marketing agency and a developer, and a compute budget now rents something close to all four. That is what this repository is trying out directly: a free custom dev shop for businesses (portfolio/requests/), where a small firm asks for a website or a tool and agents build it under an MIT licence.

The squeeze comes later, from the same source. When every firm has the same rented capability, it stops being an advantage. What's left is what can't be rented: a location, a licence, a local reputation, a relationship, a physical service. Small firms that are those things should do well. Small firms that are a thin layer of screen work (a one-person agency reselling copy, a bookkeeping service doing data entry) face the same squeeze as the junior analyst.

Every field, at many times the spend

Part 1 used this repository as a model of one person's budget. Here is what it shows when you ask what happens as the spend rises by orders of magnitude and the models keep getting better:

  • Output scales with spend almost linearly. Merges here went from about 55 a day to 280 as concurrent sessions went up (docs/parallel-agents.md). Nothing in the method stops at sixteen sessions. A firm with a thousand times the budget gets something like a thousand times the throughput, if it can coordinate.
  • Quality doesn't scale on its own. Coordination, duplicate work, red CI on main, and bugs fresh files still miss: these grew with the fleet, and each one needed a new check (CI gates, cross-review bots, red and blue teams). Smarter models shrink each error rate. More spend multiplies the number of chances to make each one. So verification becomes a large share of work in every field.
  • Demand doesn't scale with spend. This repository has 302 tracked products and no measured visitor yet. At global scale the same thing holds: the world's attention, trust, land, energy and hours of human relationships don't multiply because output does. Those become the scarce, priced things, and the jobs around them hold value.
  • Research changes the most. The repository's research side runs 1,068 project folders, which is a lot of hypotheses tried cheaply. At corporate or national budgets, with smarter models, that is a real acceleration of science where results can be checked in silicon (code, maths, chip design, simulation). Where a result needs a wet lab, a patient or a bridge, it speeds up only as fast as the physical step does.

Where this leaves the job market

On trend: a thinning of entry-level screen work first (visible now in the 22-to-25 data), then a hollowing of mid-level execution as horizons reach weeks, while employment holds or grows in physical, licensed, in-person and relationship work. Unemployment need not spike if people move into those fields and if prices falling to the cost of compute leave households with more to spend on human services. But that is a large if. The adjustment would be faster than past technology transitions, it would hit graduates before mid-career workers, and wages in the resistant fields would face a flood of new entrants.

What to watch, and what the next instalment should measure: whether the Stanford gap keeps widening past 19%; whether graduate unemployment finally separates from the comparison groups in the NBER-style data; whether Census's small-firm AI rate closes on the large-firm rate; and, closer to home, whether this repository's free dev shop gets a real business to use what it builds.

All writing · More in Building with AI agents · All tools and research