Interface Collapse: how task wins become workflow wins

Tasks are not jobs, but jobs are made of tasks

The first essay showed that AI plus verification has crossed cost and quality parity for a large class of professional cognitive tasks. The most serious continuity defence accepts that and answers with a distinction. AI can do tasks, it says, but tasks are not jobs. A job is a bundle of judgement, relationships, context and friction with the physical world. A model can draft a paragraph or fill a spreadsheet. It cannot run a practice, manage a team, close a deal or sit in a meeting. So jobs are safe even where their tasks are not.

This was the best version of the continuity story. It came from labour economics, productivity research and the business-school literature on augmentation, and it held for a while. It held the way a gate holds, until enough of it comes off the hinges.

The thesis never needed AI to replace whole occupations in one step. It needs only enough valuable task volume to cross the cost and quality threshold that firms begin rebuilding workflows around machines rather than people. A job is not a sacred object. It is a bundle of tasks, and once the bundle can be taken apart, so can the worker’s place in it.

What makes the bundle come apart is interface collapse.

Software fragmentation was a moat, and the wage paid for crossing it

A large share of white-collar labour was protected by the fact that business software does not fit together. The worker did not only think. The worker moved: from inbox to spreadsheet, from CRM to browser, from calendar to document, from dashboard to ticketing system, from codebase to wiki, from procurement portal to finance system. People were the integration layer between systems that could not talk to each other cleanly.

That was a real moat, and most of the people standing behind it could not see it, because the labour of stitching looked like the labour of thinking. A surprising fraction of what was called knowledge work was moving information between half-compatible systems while holding enough context not to break the process. The worker absorbed the friction the software produced. The wage paid for the absorption.

Models now operate the interface, not just the output

Once a model can see a screen, click, type, browse, call tools, handle files and persist across a long sequence of steps, the line between producing an output and doing the work starts to dissolve. The model stops being a text generator on the far side of a keyboard. It becomes an operator of the interface itself.

There are two routes in. The API automates wherever systems cooperate. The graphical interface automates wherever they do not. With both routes open, few ordinary enterprise environments can treat fragmentation as a lasting defence. Some systems are air-gapped, heavily permissioned, bespoke, legally restricted or tied to physical processes, and they will hold out longer. For ordinary white-collar workflow friction, fragmentation has stopped being a general protection.

The vendors are now building for persistence explicitly. In June 2026 OpenAI announced it would acquire Ona to give its Codex agents secure, persistent cloud environments, so that long-running agents can work across enterprise workflows rather than inside a single chat. That is infrastructure for work continuation, not for answering questions.

The benchmarks crossed the human baseline and kept climbing

Interface capability is measured in public. OSWorld-Verified is the standard test of desktop computer use. It asks whether a model can operate real software through screenshots, keyboard and mouse to finish multi-step tasks that span several applications.

The trajectory across recent frontier releases, as reported by OpenAI:

The reported human baseline on OSWorld-Verified is 72.4 percent. GPT-5.4 passed it. GPT-5.5 widened the gap, and also reached 98.0 percent on Tau2-bench Telecom, a test of long multi-turn customer-service workflows. In July 2026 OpenAI reported GPT-5.6 Sol at 92.2 percent on BrowseComp and 52.7 percent on the agentic benchmark Agents’ Last Exam. All of these are vendor-reported figures, and the evidence register files them as such.

The numbers describe a specific shift. The model is no longer producing an output that a person pastes into software. It is running the software layer, across applications, over many steps, with tools and memory. Work that used to need a person moving between systems now needs a person to authorise the start and check the end.

OpenAI’s own description of GPT-5.5 makes the point. It is presented as a model that researches, analyses, builds documents, operates software, moves between tools, checks its own work and pushes through messy multi-part tasks. That is not a text generator. It is a workflow operator. Within two model releases, the definition of AI capability moved from output generation to process execution.

People deploying the models describe the same shift. Cursor’s chief executive reported that GPT-5.5 “stays on task for significantly longer without stopping early, which matters most for the complex, long-running work our users delegate.” The pattern in these accounts is consistent: delegation, long-horizon execution, and work that continues rather than work that is assisted.

End-to-end scores already contain the compounding problem

The standard objection is that benchmark scores do not survive contact with production. Its most common form treats the score as a per-step success rate and compounds it. At 78 percent per step, a ten-step workflow would succeed end to end about 8 percent of the time.

That misreads the benchmark. OSWorld-Verified does not report per-step accuracy. It reports the rate at which the model completes the whole multi-step task, across real applications and file operations. The compounding is already inside the number. Exponentiating it again counts the same failures twice. The 78.7 percent figure is how often the entire task succeeds, and the human baseline it beats is measured the same way.

Production does add costs the benchmark leaves out: distribution shift, latency, permissions, audit trails, retries, escalation paths and the verifier’s time. Those costs are real. They belong in the unit cost calculation as deployment costs. They do not cancel the capability the benchmark measures.

So the comparison is not raw AI against an ideal human. It is AI with retries, guardrails, escalation and review, against people with training, supervision, review, error correction, management and quality control. The question is not whether the machine is perfect. It is whether the full machine stack reaches the same reliability at lower cost than the full human stack. For a growing share of professional digital workflows, it does.

Propagation runs in four layers, and the thesis lives in the fourth

The spread is best understood as a cascade rather than an event.

The first layer is task-level unit cost dominance. AI with thin human oversight produces professional outputs at equal or better quality, faster and more cheaply. This has crossed for a large and growing set of well-specified deliverables, and it is what the first essay established.

The second layer is interface and workflow dominance. AI works inside the same software where the work happens and handles the stitching between systems. This is crossing quickly, and the public benchmarks now track it directly. It is what this essay establishes.

The third layer is job-level dominance, where whole roles become unnecessary because enough of their task volume has been removed. This is partial and uneven, and the thesis does not depend on it. Many jobs will keep some human content for years. What matters is aggregate displacement, not the survival of any particular role.

The fourth layer is labour-market dominance, and this is where the thesis lives. Wage labour stops being the mass route to economic agency. It has not fully arrived, but the path from the first two layers to the fourth is now specified and visible.

Critics prefer to argue at the third layer, because residual human tasks are easy to find inside any existing job. The thesis was never defended there. It runs from the first two layers straight to the fourth, with the third treated as partial and contingent. Workflow recomposition can suppress hiring, break training ladders and stop mass absorption without a single occupation being formally abolished. That is why the displacement is quiet.

Displacement arrives as non-absorption, not unemployment

Mass displacement will not first look like mass unemployment. It will look like non-absorption: fewer entry-level roles, fewer graduate intakes, fewer promotions, more contractors, more review roles, productivity gains that do not reach wages, and incumbents holding their seats while new entrants fail to start. Headline unemployment can look healthy while the system underneath has stopped reproducing itself. This is the No-Scream Principle.

The signals to watch are specific. Entry-level hiring in AI-exposed fields. The ratio of junior to senior staff. Graduate absorption against stable aggregate employment. Wage compression in exposed cognitive work. Productivity gains that do not pass to labour. Contractorisation. Fewer new entrants alongside delayed retirement of incumbents. Review roles that do not grow into careers. Training ladders that break because AI now does the junior work from which expertise used to grow.

The first of these has moved, and it has kept moving. Stanford’s Digital Economy Lab first reported in August 2025 that early-career workers in the most AI-exposed occupations were losing ground relative to their peers. In an August 2026 revision built on a larger dataset, the authors report that employment of workers aged 22 to 25 in AI-exposed occupations now stands 19 percent below where it would be had it kept pace with less-exposed peers. More experienced workers in the same occupations have held up. That is the shape the thesis predicts. Not everyone fired at once, but the entry ladder narrowing first.

Job postings point the same way. PwC’s 2026 Global AI Jobs Barometer, built on more than a billion job advertisements, finds early-career postings flat in highly AI-exposed sectors, and AI-exposed junior roles seven times more likely than the least-exposed junior roles to ask for traditionally senior skills such as leadership and strategy. The junior job survives as a title while its content is promoted out of reach.

The strongest evidence against comes from the same kind of data. In May 2026, researchers at the Federal Reserve Bank of New York reported no divergence between junior and senior postings within highly exposed occupations, found that the relative fall in exposed vacancies began before ChatGPT’s release, and put fewer than 10 percent of workers and vacancies in highly exposed occupations. They conclude that the slowdown in entry-level hiring is hard to attribute to AI alone.

The two findings measure different things. Payroll records whom firms hired. Postings record whom firms say they want. If firms keep posting junior roles they no longer fill, or inflate the experience required for nominally entry-level roles, postings can show no split while payroll does. PwC’s seniorisation finding is consistent with that reconciliation, but it is a hypothesis, not a result. The conflict is open, and the evidence register tracks it as disconfirming until it is resolved.

The aggregate penetration figures remain low, which is itself what non-absorption predicts. The St. Louis Fed’s Real-Time Population Survey records generative AI assisting 1.7 percent of total US work hours in mid-2026, up from 1.4 percent at the end of 2024. That figure is self-reported, so it likely undercounts assistance built into ordinary tools. Calm aggregates are not a refutation while the entry-level signals move. By the time displacement shows in the headline statistics, the propagation has already done its work.

Augmentation is a stage, not a destination

The standard reply is that people will work with AI. The sentence is asked to carry more than it can, because there are three kinds of complementarity and they are not the same.

Genuine complementarity means AI raises the marginal value of human labour, and the worker captures part of the gain through pay, bargaining power or promotion. This is real. It is not guaranteed to last, because the model that makes a worker more valuable this year may absorb the role next year.

Transitional complementarity means people supervise, correct, validate, integrate and carry responsibility while the system improves. This is the phase most often mistaken for the future of work, and it is unstable by design. The human role exists to cover the system’s current limitations. As the limitations are fixed, the role thins towards nothing.

Theatrical complementarity means people remain for trust, liability, regulation, customer comfort or institutional appearance. The person is still in the room but no longer economically central. That is not augmentation. It is managed displacement dressed as augmentation.

The augmentation story points at the second stage and calls it the destination. The thesis calls it a corridor. Once the corridor has a name, “people will work with AI” stops being an answer and becomes a question. Which complementarity, at what margin, for how long, and under what competitive pressure? Genuine complementarity has to be shown, not assumed. The transitional kind is unstable by construction. The theatrical kind is displacement under another name.

No single decision looks like recomposition

Firms do not adopt interface collapse as a decision. It is the sum of thousands of small workflow changes, each defensible on its own terms.

No firm announces that it is rebuilding its operations around AI agents. It deploys a model in customer service. It widens the deployment. It connects the model to the ticketing system, then the ticketing system to the CRM, then automates the escalation path, then lets customer-service headcount fall as the system handles more cases. Each step is incremental and passes a quarterly cost review. The total is a recomposed workflow, but no individual decision looks like one. Each looks like ordinary process improvement.

That is why the change is so hard to govern from inside the firm. The decisions are too small to need board approval and too dispersed to be coordinated. They are made by line managers who own the budget and face direct pressure on cost and throughput. A chief executive who wants to slow the change has to override hundreds of locally rational decisions, and the firm that does so ends up with higher costs than rivals that do not.

What would refute this

Premise Two fails if interface-level capability stalls in a way that stops the propagation. That would mean frontier models that cannot reliably operate desktop software, navigate browsers, manipulate files or sustain multi-step work across applications, despite continued investment and progress at the task layer. OSWorld-Verified, BrowseComp and Tau2-bench Telecom track the relevant capability, and a sustained reversal across independent benchmarks would count. So would deployment evidence that integration across interfaces routinely consumes the task-level cost advantage.

On the labour-market side, the refutation is the opposite of non-absorption. If entry-level hiring in AI-exposed cognitive work recovers, junior-to-senior ratios stabilise or widen, and the Stanford early-career gap closes rather than widens while the New York Fed’s postings picture holds, then workflow recomposition is not suppressing absorption and the propagation the essay describes is not happening.

Showing that jobs still contain human tasks does not count. The premise never required whole jobs to disappear.