The empty-seat economy — what contracts first when AI takes over office work: online recruitment, per-seat software and the intranet

The first casualty of artificial intelligence replacing office workers is not the office worker — it is the industry that lived off hiring, equipping and housing them.

In September 2026, total white-collar employment in advanced economies has barely moved, yet its intermediaries are bleeding: Hays, one of the world’s largest recruitment firms, has reported its first annual loss in 23 years, Indeed and Glassdoor merged after layoffs, and software stocks went through a repricing that Jefferies analysts christened the “SaaSpocalypse”. All these businesses share one denominator: they bill their clients in proportion to headcount — per job ad, per placement, per licence “seat”, per square metre of office.

The verifiable thesis of this analysis: contraction follows the order of intermediation, not the order of employment. First to shrink are the markets that live off the flow of hiring — online recruitment, staffing, job ads —, because a hiring freeze cuts their volume before anyone gets fired. Then software decouples from seat counts, through the shift to consumption or outcome pricing. Professional networks lose the hiring product but keep attention and advertising. Office employment itself declines slowly and unevenly — through entry doors closed to the young, not through waves of layoffs.

An economy built on headcount

The dimensions are considerable. Worldwide enterprise software spending reaches about 1.4 trillion dollars in 2026, up 14.7%, according to Gartner. The online recruitment market is estimated at roughly 40 billion dollars — figures differ widely between research firms, so the order of magnitude is all that should be retained. LinkedIn alone took in 17.81 billion dollars in the fiscal year ended June 2025, according to Microsoft reporting. Intranet software — “the place where the company talks to its employees” — adds around another 22 billion.

The common denominator is the billing unit: the employee. A Salesforce licence, a Microsoft 365 seat, a LinkedIn Recruiter subscription, a staffing contract, a floor of offices — all are sold “per person”. The model was the commercial genius of the past two decades: the client’s growth automatically became the vendor’s growth. But the same arithmetic works in reverse: if a workflow goes from ten people to three people plus a software agent, the invoice shrinks mechanically, even as the client firm prospers. AI substitution does not need to produce unemployment to contract these markets — freezing hiring is enough.

The contraction has begun — among intermediaries, not on payrolls

That is exactly what the 2024–2026 data show. Recruit Holdings, the Japanese owner of Indeed and Glassdoor, laid off roughly 1,300 employees in July 2025 — about 6% of its HR-tech division — with the explicit justification of the shift to AI, according to TechCrunch; on 1 July 2026, Glassdoor was legally absorbed into Indeed. The volume of paid job ads on Indeed fell by about 7% in the last fiscal year, and the company is compensating for a shrinking market with higher prices — up 17% last year, according to Staffing Industry Analysts.

ZipRecruiter, the American barometer of job advertising, saw revenue fall from 905 million dollars in 2022 to 449 million in 2025 — a halving in three years, according to SEC filings. Honesty demands the reverse side too: in 2026 the company stabilised, with the second quarter at 118.1 million, up 5% year on year — proof that a halved market can find a new equilibrium rather than just the road to zero.

Staffing tells the same story with different numbers. Hays closed fiscal 2026 with a loss of 54.5 million pounds — its first in 23 years — and fees down 8%, citing AI as the leading factor of uncertainty. PageGroup lost 7.4% of revenue in 2025, Robert Half’s revenue fell by about 3% in the second quarter of 2026, and BNP Paribas Exane analysts downgraded the sector arguing that temporary workers are 40% more exposed to automation than the labour-market average. Challenger, the firm that tallies American layoffs, counted 87,714 layoffs explicitly attributed to AI in 2026 through May alone — 22% of the total, versus 5% across all of 2025.

And where hiring still happens, it bypasses beginners. The “Canaries in the Coal Mine” study by Stanford’s digital economy lab — on ADP payroll data covering more than 26 million workers — shows that hiring of 22–25-year-olds in the most AI-exposed occupations sits 19% below the trajectory it would have followed without this shock, while for their older colleagues in the same occupations hiring is rising, according to the August 2026 update. Unemployment among recent American graduates has climbed toward 5.7% — for the first time in decades above the general rate of about 4.3%, according to the Federal Reserve Bank of New York.

The stock market repriced software ahead of the results

For per-seat software, the shock came on 30 January 2026, when Anthropic launched industry plugins for its office agent. In the following week, software stocks lost around one trillion dollars — the sell-off that a Jefferies analyst quoted by Bloomberg called the “SaaSpocalypse”; estimates extending the count over the following two weeks rise toward two trillion, but they circulate through secondary sources and should be treated as an order of magnitude, not a measurement. Atlassian fell 36% in February and laid off about 1,600 people in March; Salesforce lost roughly a third of its value in 2026; HubSpot dropped 19% in a single day after announcing its new agent pricing.

The financial results, however, tell a more interesting story than the panic: not collapse, but a change of measuring unit. Agentforce, Salesforce’s agentic product, passed 1.2 billion dollars in annualised revenue, up 205%, according to CNBC — yet Marc Benioff said in May 2026 that the firm is hiring almost no one outside sales, after a fiscal year with no new engineers, while the support team shrank from about 9,000 to about 5,000 people, the work being taken over by its own agents.

Microsoft climbed from 15 to more than 30 million paid Copilot seats between January and July 2026 — and, in parallel, sells Copilot Studio on consumption credits, not on seats. Workday reports that more than a quarter of new contract value comes from AI. Software is not dying; the seat, as an economic unit, is.

One detail of informational hygiene: in the same wave of panic circulated the claim that Atlassian had reported “the first seat decline in its history”. The official shareholder letters show the opposite — seat expansion continued, and cloud revenue later accelerated to plus 29%. The narrative outran the data; it is a warning valid for both camps of the debate.

LinkedIn: the network wins, the hiring product loses

LinkedIn is the cleanest test of the thesis, because it holds all three businesses under one roof: hiring (Talent Solutions), attention (Marketing Solutions) and subscriptions. Total revenue has grown healthily — from 17.81 billion dollars in fiscal 2025, with growth accelerating toward 12% per quarter in 2026. But the engine has moved: for four consecutive quarters, growth was pulled by advertising, while Talent Solutions stagnated on a frozen hiring market and recovered only in the spring of 2026 — and the recovery is explicitly attributed to agentic recruitment products, which passed 450 million dollars in annualised revenue, a figure announced by Satya Nadella on 29 April 2026, according to Staffing Industry Analysts.

That same month, LinkedIn got a new CEO — Daniel Shapero, in a restructuring publicly tied to AI — and from fiscal 2027 Microsoft is reorganising its reporting so that Talent Solutions disappears as a separate line, absorbed into an “industry solutions cloud” metric. When a segment stops being reported, it usually stops being the growth story.

Meanwhile, the platform is fighting its own medium: a Pangram study of more than a million posts classified about 40% of long-form content on LinkedIn as entirely AI-written — the highest rate among the platforms tested —, and in July 2026 the company introduced a “seems like AI slop” button, pressed by more than a million users in the first two weeks; content so labelled loses about 40% of its views, according to LinkedIn.

On the candidate side, the platform receives about 14,200 applications per minute — 58% more than a year earlier, according to its own Workforce Confidence report of February 2026 —, and an average job ad gathers 242 applications. When both supply and demand are machine-generated, the signal disappears — economists call this “the market for lemons”. The pivot announced by Shapero is symptomatic: LinkedIn is investing in verification — that the person, the skills and their work are real. The network that got rich on CVs is repositioning itself as a notary of authenticity, because the CV, as a document, is dead.

The causal chain — and what could break it

The mechanism, made explicit. The cause: models are getting ever better at the standardised entry tasks of a profession — triage, synthesis, routine code, correspondence. The mechanism: firms do not fire en masse, they stop hiring at the base — “low-fire, low-hire” —, which first cuts the flow of hiring: ads, placements, recruiting licences. The first-order effect: hiring intermediaries contract years before any visible fall in employment. The second-order effect: software buyers renegotiate contracts onto consumption or outcomes, because seat counts are no longer growing — and vendors accept, so as not to lose the client. The third-order effect, the slowest: the composition of the office changes — fewer juniors, more agent oversight — and only that, eventually, moves payrolls.

The main counter-hypothesis deserves its full weight: part of the contraction is the economic cycle, not AI. LinkedIn’s own workforce report from January 2026 noted that, outside healthcare, hiring patterns look similar in highly and barely AI-exposed occupations — a sign that interest rates and post-pandemic normalisation explain a great deal. Yale’s Budget Lab and Brookings, who monitor monthly, find “no discernible disruption pattern” at the aggregate level since ChatGPT launched. The OECD estimates that only about 9% of jobs are realistically automatable in the short term. The honest verdict: the evidence supports the contraction of intermediaries and the closing of the entry door — not, for now, an employment apocalypse.

There are documented brakes too. MIT found that 95% of enterprise generative-AI pilots produce no measurable return; Gartner forecasts that more than 40% of agentic projects will be cancelled by the end of 2027; Klarna, the showcase of replacing support with AI, rehired people in 2025 after quality slipped; IBM, after its 2023 hiring pause, employs more people today, redeployed.

The Jevons paradox works against the story as well: if the hour of legal or accounting work gets cheaper, demand for it can rise — though Apollo’s chief economist, Torsten Slok, observes that the result would be “more lawyers, but fewer junior associates”. And physical offices, the surprise of the list: CBRE projects US office vacancy declining from 18.3% in 2026 toward 14.5% in 2031 and estimates that only 5% of office jobs are highly vulnerable — office demand is being pulled up precisely by AI firms.

The future in steps: how the film plays out, if the dominant scenario holds

What follows is projection, not prophecy — the film of the highest-probability scenario from the simulation below, with its branch points.

Step 1 (2026–2027): the silent freeze. Entry-level hiring stays shut; ads decline, prices per ad rise — the Indeed model. Recruitment platforms consolidate through mergers, not bankruptcies. Per-seat vendors monetise the installed base through price rises — LinkedIn Recruiter buyers report about 15% more at renewal, in the absence of a published list price — and bundle agents into ever more expensive subscriptions. The war on “AI slop” becomes a permanent operating cost of any professional network. Indicator to watch: the share of AI-attributed layoffs in Challenger’s reports.

Step 2 (2027–2028): the metric decouples. Financial reports begin to show stagnant seats next to growing consumption revenue; the first major vendor publicly abandons per-seat list pricing on a flagship product. Reporting itself is rearranged to hide the weak lines — Microsoft is already doing it with the LinkedIn segments from fiscal 2027. Generalist staffing enters a merger wave; offices bifurcate: premium buildings fill up, secondary ones empty out. Indicator: the first quarterly report in which a top-5 vendor publishes consumption revenue as its headline metric.

Step 3 (2028–2030): the agent becomes the customer. Our panel assigns this step the highest probability of all the questions put to it: platforms come to sell more to agents than to people — the machine interface becomes the product, the human interface becomes the oversight report. Classic intranets dissolve into the agent layer — the retirement of Viva Topics is the precursor —, and recruitment splits into two markets: verification of real humans (LinkedIn’s direction) and agent-to-agent matching, with no ad and no CV. White-collar employment drifts slowly — the simulation’s median: about minus 3% versus 2026 — but its composition is unrecognisable at the entry point.

Step 4 (2030–2035): who survived. The winners are identity and verification networks, owners of proprietary data, and platforms that managed the transition to outcome pricing. The losers are volume intermediaries: classic job boards, simple applicant-tracking systems, standalone intranets, undifferentiated staffing. The critical branch: if the rupture scenario — 15% probability — materialises, the whole film above compresses from nine years into three. The emergency signal would be entire cohorts of office workflows running with no human in the loop — Gartner’s threshold of 15% of daily decisions taken autonomously by agents, reached before 2028.

The scenarios of a simulation panel with nine virtual people

To put probabilities on this film, we ran the dossier through a Monte Carlo simulation with 20,000 runs over four scenarios, doubled by a simulated deliberation: an olivLaw panel of nine virtual people with deliberately divergent interests — an agency recruiter, an HR director, the CFO of a SaaS vendor, a commercial strategist at a large AI lab, a labour economist, an office investor, a 2026 graduate, a bank CIO and a sociologist of work. The result is a simulated editorial judgement on the facts above, not a measurement; the probabilities are estimates conditional on the simulation's assumptions, not observed frequencies.

How to read the numbers. The simulation and the panel explore scenarios: their results depend on the assumptions put in — the scenario weights and the indicator distributions, chosen editorially. The 20,000 runs do not prove accuracy; a large number of runs only stabilises the calculation on the same assumptions, it does not make them truer. The nine virtual people are AI-generated characters built to cover divergent interests; they are not a statistical sample of real people and do not represent the views of any population. The figures below remain useful as orders of magnitude and as forecasts with a deadline and a criterion, which can be checked.

The dominant scenario — “substitution in steps” (about 40%). The entry-level hiring freeze continues and widens; large layoffs remain the exception. Entry-level ads fall toward minus 30% versus 2026 by 2030, online recruitment and staffing lose about a fifth, per-seat software grows by only a few percent a year, and the hiring revenues of professional networks slip below their 2026 level. Visible trigger: another year of Stanford reports with the youth gap widening.

The edge scenarios. The “copilot plateau” (about 22%): AI remains a tool, the productivity gain is reinvested into more work — Jevons —, platforms grow more slowly, but they grow. “Friction and backlash” (about 23%): the wave of cancelled agentic projects, regulation and Klarna-style rehiring freeze substitution; today’s contraction proves cyclical. The “agentic rupture” (about 15%): agents take over entire workflows, seats decline nominally, the recruitment market collapses by more than 40%. The sum says something counterintuitive: the scenarios in which today’s platforms survive reasonably add up to about 45% — but none of the four brings back the world of 2021.

The point questions, with ICD-203 calibrated terms computed from the simulation: that LinkedIn’s hiring revenues will be below their 2026 level in 2030 is likely (about 65%); that platforms will sell more to agents than to people by 2030 is likely (about 79%); that online recruitment contracts by more than a quarter — roughly even chances (about 45%); that intranets disappear as a standalone product — roughly even chances (about 47%); that per-seat SaaS growth turns negative — unlikely (about 29%); that white-collar employment falls by more than 10% by 2030 — unlikely (about 28%).

The fault lines in the panel foreshadow the public debate. The lab strategist argues the panel underestimates speed — capabilities keep doubling, and 77% of enterprise API usage is already automation, not assistance, per Anthropic’s data. The labour economist answers with the Yale–Brookings aggregate null and the history of ATMs, which coexisted for three decades with more bank tellers. The SaaS CFO notes that AI-native margins — 50–60% gross — are structurally worse than subscription margins, so “the death of the seat” is also the death of software’s most profitable money machine.

The 2026 graduate cuts it short: for their cohort, the debate is over — the door is already closed. And the office investor reads CBRE and refuses to mourn: the physical office is the only “per-employee” asset the data show in recovery.

Verifiable predictions

Prediction	Horizon	Probability	How it is verified	Verification source
Another major online recruitment platform (from the list: ZipRecruiter, Monster, CareerBuilder, StepStone, Totaljobs) merges, is sold or is delisted	31.12.2027	55–70%	Official transaction announcement	Financial press, SEC
Microsoft reports at least 50 million paid Copilot seats	30.06.2027	55–70%	Figure in the earnings call	Microsoft IR, transcripts
AI-attributed layoffs in the US for all of 2026 exceed 150,000	31.01.2027	50–65%	Annual total from monthly reports	Challenger, Gray & Christmas
US office vacancy falls below 17.5% — the counter-narrative to “empty offices”	31.12.2027	55–70%	Quarterly market report	CBRE, JLL
A top-5 global SaaS vendor publicly abandons per-seat list pricing for a flagship product	31.12.2028	45–60%	Official pricing page / announcement	Vendor sites, press
The 2027 update of Stanford’s “Canaries” study shows the youth gap in exposed occupations above 19%	31.12.2027	55–70%	Publication of the update	Stanford Digital Economy Lab
Aggregate US white-collar employment does NOT fall by more than 5% versus 2026	31.12.2028	75–90%	Employment series by occupation	BLS (CES/CPS)

Method, definitions and figures

The simulation. Monte Carlo with 20,000 runs over four scenarios with editorial weights — substitution in steps 40%, friction and backlash 23%, copilot plateau 22%, agentic rupture 15% —, each with truncated normal distributions over nine indicators at the 2030 horizon: white-collar employment, entry-level ads, per-seat SaaS growth, the hiring revenues of professional networks, online recruitment, staffing, intranets, office vacancy and agent adoption. The panel deliberation runs three rounds over the same scenarios; for questions with a direct counterpart in the simulation, the answer is anchored in the simulated probability (65% weight), and the panel only tilts it — so that qualitative judgement cannot detach from its own numbers. The ICD-203 calibrated terms (“unlikely”, “likely”) are computed deterministically from the probabilities, not chosen by any model.

The limits. The scenario weights and distribution parameters are editorial judgements informed by the cited sources, not econometric estimates; the simulation propagates these assumptions coherently and exposes them to verification through the predictions table — it does not prove them. Market figures for online recruitment and intranets vary severalfold between research firms and are used only as orders of magnitude. The “two trillion dollars” estimate for the software repricing circulates through secondary sources; the one trillion in a week is attributed to Bloomberg via intermediaries. The Stanford/ADP data cover the US; extrapolation to other advanced economies is an assumption.

Definitions. “Per-seat” — a software licence billed per named user, monthly or annually. “White-collar” — office occupations: administrative, financial, legal, technical, managerial. “Entry-level” — roles with no experience requirement, typically under three years. “Staffing” — the placement of temporary or permanent personnel for a fee. ARR — annualised recurring revenue. Microsoft’s and Salesforce’s fiscal years end in June and January respectively; where the text says “fiscal year 2026”, the calendar period is specified at first mention. All amounts are in the currency of the original reporting: US dollars, pounds sterling or euros, as the case may be.