New York's AI Office Boom Exposes a First-Job Gap

AI-generated editorial illustration.

On October 2, the Partnership for New York City published two numbers that do not fit neatly in the same economic-development speech. AI companies had leased more than 2.2 million square feet of New York office space in the first half of 2026, more than twice the sector’s full-year 2025 total. Yet annual entry-level job postings in design, media and writing occupations had fallen 40.6% since 2022. Customer and client support postings were down 34.4%. These are different companies and different units of measurement. Their side-by-side display still creates a concrete problem for a city selling an AI boom: who is paying for the first years of work that turn a newcomer into an experienced employee?

The Partnership’s 12-page report draws its leasing figures from brokerage research and its job-ad figures from Lightcast. It does not follow the people hired by those office tenants or prove that AI eliminated a junior role. A lease reserves space. An advertisement solicits candidates. Neither records a young worker’s first payday. That missing link is measurable, if employers agree to report it.

New York already has a small, observable counterexample to a purely abstract workforce promise. In August, the city’s Economic Development Corporation reported more than 130 paid student placements at nearly 90 startups and venture firms during summer 2026. For the startup cohort, it also disclosed what happened next: 57% continued working after the internship and 27% received return offers or extensions from their hosts. Those are program results, not citywide employment rates. But they reveal the kind of denominator missing when a city counts leases, venture capital and AI pilots as if they were interchangeable with careers.

A lease buys time before it buys labor

The Partnership’s headline comparison uses 2.2 million square feet in the first half of 2026 against 936,000 for all of 2025. Its report says 13 existing AI firms signed leases that added a combined 1.2 million square feet to their footprints. The same document places 2025 AI startup venture fundraising at $16.7 billion, citing its analysis of PitchBook data. All three figures describe capital commitments or physical capacity. None specifies how many local residents started a paid junior job.

Even the floor-space numerator changes with the researcher’s definition. Colliers counted 1.50 million square feet of Manhattan AI leasing across 63 transactions in the first half; its broader technology category reached 4.15 million square feet. Savills and CompStak counted 1.7 million for AI occupiers in Manhattan over the same period. Their classification, geography and transaction data are not identical to the Partnership’s cited Newmark series. Choosing the largest estimate as a proxy for hires would hide the first measurement disagreement.

There is a company-level reason to keep the ledgers apart. An AI firm may lease a larger office for sales, security-sensitive client work or an anticipated research team, while placing its first round of hiring in experienced roles. Its customers may sit outside New York. A landlord can welcome that lease; an early-career worker sees no opening. Conversely, a company can hire locally while taking little additional space if it works remotely or shares a building. The city’s economic case should value both activities on their own terms. Requiring each lease to produce a fixed number of first jobs would be as crude as treating every new desk as a hire.

For a landlord, a lease is a contract with rent and duration. For an employer, it is an option to expand, consolidate or bring distributed staff together. Savills says the average Manhattan AI lease term in its sample lengthened from 49.0 to 90.9 months and the share completed through subleases fell from 42.3% to 17.5% when it compares 2020-2024 with 2025 through mid-2026. That signals a firmer real-estate commitment in the sample. It still does not tell a college graduate whether the tenant will post a junior position.

The city’s wider office market also has other tenants and other reasons to move. Colliers identifies Google, Ramp and Uber among the large technology commitments in its half-year data; its all-tech category should not be relabeled as AI startup hiring. Cushman & Wakefield put Manhattan’s overall vacancy rate at 19.3% in the second quarter, even after a quarterly decline. More leasing can repair a property market before it changes the employment market. It can also move an existing worker from one address to another.

There is a genuine upside to taking space ahead of staff. A growing firm may need secure rooms, labs, client space and room for the team it expects to build. A long-term lease can indicate that a founder is committed to the city rather than merely visiting. The claim needs a follow-up date. Six or twelve months later, how many seats are occupied, by whom, and how many of those workers entered through a local paid junior role? Without that follow-up, square footage remains a landlord’s outcome.

Junior openings fell while the city kept building

The Partnership’s posting table covers occupational groups, not named AI tenants. It reports changes in annual entry-level ads since 2022: design, media and writing -40.6%; customer and client support -34.4%; clerical and administrative work -30.5%; business management and operations -26.8%; and finance -23.4%. The report pairs these with an index of the share of occupational skills susceptible to AI augmentation, all above 50% for those groups. Exposure is an estimate of tasks. It is neither observed AI use nor a count of jobs displaced.

Its most revealing second figure is positive. Entry-level postings mentioning AI skills were 55% higher than in 2022. Employers may be asking a shrinking applicant pool to arrive with more applied knowledge. The series does not show how many such ads demanded experience, paid a wage above other junior roles or resulted in hires. A rising percentage of AI-related ads can coexist with fewer first jobs in the wider occupation.

The Center for an Urban Future’s September analysis finds a 49% fall in New York City entry-level tech postings since 2022. Its category is tech jobs, not the Partnership’s seven broader white-collar occupation groups. It reports a 29% decline in postings for the most AI-exposed occupations against a 21% increase for minimally exposed occupations. The center also finds a 20% increase in entry-level ads requiring AI skills. These figures sharpen the question, but the center and the Partnership both draw their posting trends from Lightcast. They are two analyses of related underlying ads, not independent replications of an employment effect.

A separate kind of evidence comes from the U.S. Census Bureau’s September working paper. Using administrative records on college graduates, Cody Orr, Lee Tucker and Lawrence Warren estimate that graduates in the most AI-exposed decile of majors saw initial employment probability fall by five percentage points and full-quarter initial earnings by 13% after late 2022, after the paper’s regression adjustments. It is a national study of graduates grouped by major. It does not locate a loss in New York, identify a particular employer’s automation decision or turn the Partnership’s job-ad changes into causal estimates. Its value here is that it observes employment and earnings rather than only vacancies.

That distinction matters for people who do find work. A graduate whose first job moves from a finance office to a retail shift still appears employed, but the pay and accumulated occupation-specific experience may change. The Census authors report that roughly half of their initial earnings decline came from lower pay within the industries that employed the graduates; the rest came from movement toward lower-wage industries. That is a national estimate for highly exposed majors, not a description of any particular New York graduate. It suggests that a city should count the quality and direction of first work as well as whether someone found a job at all.

The measures belong in separate columns. Posting counts can change when vacancies, duplicate ads or recruiting channels change. Administrative employment records observe paid work. A graduate can also move into another occupation or city, with different pay. To assess who benefits from AI growth, New York would need to trace those transitions across cohorts rather than multiply percentages from unrelated samples.

A posting is a request, not a hire

Late 2022 makes an attractive dividing line because ChatGPT arrived then. It was also the end of a pandemic-era hiring surge. Employers absorbed higher borrowing costs and reconsidered permanent versus contract staffing. The Center for an Urban Future names those pressures explicitly. Its authors spoke with training providers and reported that several had seen placement rates fall by at least a third. Those accounts indicate strain in the providers’ pipelines; they are not a citywide placement survey.

On October 5, the center’s Eli Dvorkin told the City Council that the tech economy remained comparatively resilient even as its first rung narrowed. He named higher interest rates, economic uncertainty and the retreat from pandemic hiring alongside AI. His testimony illustrates the policy stakes without resolving the statistical attribution. A council can respond to a shortage of paid entry work without first asserting that one tool caused every missing ad.

New York’s Comptroller described a low-hire, low-fire economy in May: weak new hiring has coexisted with relatively few layoffs. The office considered AI one possible reason, particularly in knowledge work, and also detailed the limits of current adoption. Its review of a survey of chief financial officers said the top stated reason for investment was production efficiency rather than labor-cost reduction; among adopters, 57% had put AI into no more than three business functions. That national firm survey is not a census of New York employers. It is a check against writing every empty junior requisition as a completed automation project.

The Partnership’s own sector figures are similarly narrow. Through August 2026, its analysis says New York financial services employment was up 0.6%, professional and business services up 0.04%, and information employment down nearly 2%. Those are broad industry totals, containing senior and junior staff at both AI users and nonusers. They can describe a city where the total workforce barely moves while fewer people enter it. They cannot identify which new lease, if any, replaced which first job.

Hiring managers face a real constraint even if software takes over some routine work. A first job traditionally lets someone learn a company’s data, customers, rules and unwritten exceptions under supervision. Removing the easy tasks can raise the bar for the first paid assignment. It may also release experienced employees to train newcomers more closely. Both outcomes are plausible. The difference depends on whether the employer budgets mentor time and defines work a beginner can safely own, rather than asking for the finished judgment of a senior employee on day one.

This is why a firm-level comparison would be more useful than a citywide correlation. Record the junior vacancies, applications, offers and accepted starts before and after a product deployment. Preserve the site’s demand and business-cycle context. Track contractor conversions, work quality and supervisor hours. Then compare the cohort to a similar team that did not deploy the tool. That design still has limitations, but it gives a buyer a narrower claim than ‘AI changed hiring’ and gives a worker a clearer view of where the entry point went.

New York has programs with different denominators

The city is not starting from zero. Its Economic Development Corporation’s AI Nexus, operated by C10 Labs and Plug and Play, was announced with targets of up to 165 AI startups and 96 pilots through 2029. It connects young companies with sector partners and adoption projects. A startup assisted and a pilot launched are outputs of an innovation program. The announcement does not promise 165 new employers hiring interns or 96 entry-level job offers.

The Startup and Venture Capital Internship Programs measure a different product: paid work. For summer 2026, NYCEDC reported more than 130 city-based students at nearly 90 startups and venture firms, with 82% of placements filled by City University of New York students. The startup portion alone placed more than 100 students at 68 tech-enabled startups. In that portion, 42% worked in software engineering or data science, 57% continued working after the internship, and 27% received a return offer or extension from the host. Continued work and offers overlap; the percentages cannot be added. NYCEDC does not give a common denominator for full-time employment twelve months later in this release.

One named account shows what the program can reveal without pretending it represents everyone. The NYCEDC release quotes Emotivo Health intern Dereck Taverne saying the placement became a continuing role, and Emotivo Health CEO Karen Lin describing a project that shipped features for clinician review. The release is the source of those accounts; Digidai did not interview either person. It does not disclose Taverne’s wage, contract length or later career outcome. Still, it describes a paid task, employer supervision and a continuation, a far more direct chain of events than an office lease.

At city scale, the program’s counts are modest. The Partnership’s report says New York’s AI startups raised $16.7 billion in 2025; the 2026 internship program reached a little over a hundred startup students. That juxtaposition is not a spending ratio: venture financing, city program funding and participant counts have different years and owners. It is a prompt for a budget question. If city officials want AI investment to open a first rung, which employers will fund paid seats, who will provide mentors, and what share of participants should receive a continuing offer?

The Center for an Urban Future proposes a six-month NYC AI Service Corps, initially funded with $10 million in city money alongside private and philanthropic support. It would place emerging technologists in agencies, nonprofits and small businesses to perform real projects. The center also asks for the next 100 large and midsize employers to build paid routes with CUNY and training organizations. Those are recommendations, not funded programs or observed outcomes. A city budget would need to specify seats, wages, mentor capacity and evaluation before the suggested dollars could be called a career-building result.

There are competing interpretations of what the city should buy next. NYCEDC Interim President and CEO Jeanny Pak described the internship program as a way to widen access to experience and networks. Tech:NYC’s Julie Samuels emphasized paid work with growing startups. Those views appear in NYCEDC’s own August release; they are program sponsors’ accounts, not independent evaluations. A public funder could instead prioritize broader employer apprenticeships that do not depend on startup survival. The New York Jobs CEO Council’s work has a much larger volume of low-income New Yorkers hired by member firms, but that measure spans multiple years, employers and job levels. It cannot be pasted into the startup internship denominator. The city needs to decide which population and kind of first opportunity it intends to serve.

AI-generated editorial illustration. The separate stages represent an investment, paid work and a job offer, which must each be measured.

Supervision is part of the bill

Employers cannot buy an experienced worker by purchasing a short course for a beginner. A useful first project has to fit an actual team backlog, and a manager has to make time to review it. That has a cost even when the student’s wage is subsidized. If an intern spends ten weeks testing an internal workflow, a senior employee may need to design access, review outputs, correct mistakes and judge whether a permanent role exists. If no one carries those duties, the placement can become a credential with little employer evidence behind it.

This is also where an AI-focused program can differ from a generic coding class. The city could ask participating employers to name a real business problem, a permissible data environment, the decisions the novice may make, and the work a mentor must sign off. A candidate might document errors in an AI-assisted customer workflow, build a retrieval test set, or measure where a model sends cases back to humans. Those are illustrative project types, not reported assignments in the NYCEDC cohort. Each would let a supervisor judge work quality without inventing a standalone ‘AI job’ that the employer has not approved.

The burden can fall unevenly. A small startup may have project work but no spare manager. A large bank may have mentors but a slow security review and a hiring freeze. A student who must earn rent cannot accept unpaid practice while waiting for access. The Partnership’s report asks New York to pair startup growth with workforce strategy; the internship release shows one way to do that at limited scale. Neither document provides a cost per sustained first job. That is the number a city finance office would need before expanding a pilot.

The employer’s alternative is a junior role redesigned around AI tools, with explicit training time in its cost center. That approach may be cheaper than a separate public program for some firms. It also means accepting lower initial productivity and measuring the later payback. A founder deciding whether to hire a graduate or another senior specialist cannot use the city’s aggregate posting chart as that return-on-investment calculation. The calculation depends on the team’s backlog, supervision hours, turnover and the length of time it takes the new employee to own work safely.

The worker’s cost is less visible in a capital or workforce plan. A paid placement can still be inaccessible if the wage does not cover commuting and living costs, the work hours conflict with classes, or an offer arrives months after the student needs income. CUNY students accounted for 82% of NYCEDC’s 2026 placements, according to the release, but it does not break out which participants accepted extensions, moved to another employer or left the field. A follow-up that asks only the host whether it hired someone will miss those paths. A worker-approved check after three and twelve months can show whether the initial opportunity led to a stable next step, with privacy and nonresponse rates reported alongside the result.

A conversion ledger for the first job

A city scoreboard should follow a cohort through successive decisions. The economic-development office, a participating employer and a training provider need to see different outcomes for the same people. The table below is an editorial measurement design; New York does not currently publish all these fields together.

Stage	Count and denominator	Proof and responsible party	Failure worth reporting
Investment	AI leases signed, occupied seats and new local hires, each with its own date	Landlord or broker for space; employer for payroll starts	Treating an expansion lease as a new headcount commitment
Demand	Unique junior requisitions, qualified applicants and approved paid seats	Employer HR and finance records, with a stable job definition	Counting reposted ads as distinct opportunities
Paid work	People who start and complete a paid project, wages and mentor hours	Program operator and host employer, with participant consent	Training enrollment that never reaches supervised work
Conversion	Offers, accepted starts and continuing work at three and twelve months	Host employer and, where lawful, verified follow-up	Adding overlapping offers and extensions or calling an internship a permanent job
Mobility	Pay, job level and occupation at entry and after a year	Employer and worker-approved follow-up, reported in cohorts	Losing workers who leave the host for another good job

This ledger has a practical denominator rule. For each cohort, publish the number eligible, selected, started, paid, completed, offered and still working at a defined date. Keep the same cohort through the sequence. A host that hires an intern after a project should count one continuing worker, even if it also issued an extension letter. A participant who moves to a different employer should not disappear from the outcome file if the program’s purpose was a durable career start rather than retention at the first host.

The system also needs a cost column that can survive a budget meeting: public subsidy, private wages, software, supervisor time and administration per paid placement, then per accepted job and per sustained job. Because the city has not published those costs alongside the 2026 internship results, no credible cost-per-job can be computed from the public release. Publishing it could expose programs that buy activity cheaply but do not reach a paid start. It could also vindicate a more expensive placement that produces durable work.

New York should resist collapsing the city’s labor market into one AI exposure index. The Partnership groups occupations by susceptible skills; the Center for an Urban Future focuses on tech entry posts; Census tracks graduates by major across the country. A conversion ledger can be segmented by CUNY status, prior work, occupation and employer size without claiming that any group is a perfect counterfactual. The first public version could compare cohort outcomes against a matched historical program cohort and disclose selection differences. A later study could test stronger comparisons. Precision improves when the city says which question each number answers.

Any comparison also needs a calendar. The student graduating into the 2022 labor market faced a different mix of vacancies and interest rates from the student graduating in 2026. A program that selects applicants with prior experience will look more successful than one deliberately opening the door to people with no professional network. Publish the selection rule and the share who declined an offer. Track employers that stopped hiring altogether alongside those that joined the program. Otherwise the most attractive number in a mayoral press release may simply be the one that lost its denominator on the way from the payroll file to the podium.

Paid work belongs in the renewal file

By the time an AI tenant renews a Manhattan office lease, its headcount plan may have changed several times. The landlord will know the square footage and rent. The startup will know its burn rate and whether the larger office helped it serve customers. The city should know something else: how many local entrants received paid supervised work, how many accepted a first offer and how many remained employed a year later.

The outcomes may diverge. A productive AI company could hire few beginners and still build valuable products. A training program could send graduates to other employers rather than to the startups that drew the original investment. Each public goal needs its own budget and measure. When NYCEDC next asks for money to expand a paid cohort, officials can request the number of participants who started, finished, accepted work and still held a job a year later. The square footage belongs in a different file.