AI Builder Ads Grow While Early Careers Move to Contracts

AI-generated editorial illustration.

On September 30, workforce-data company Draup put two figures in its Fortune 500 hiring report. Its AI Builder family had reached 27% of tracked technology job-posting demand. Internships and contracts together also accounted for 27% of the early-career descriptions it classified, up from 13% in 2020. The matching percentages describe different groups of advertisements. They do not mean that more than a quarter of graduates became AI Builders, or that a quarter of AI Builders work under temporary contracts.

Five days later, Workday reported another split. Among roughly 550 employers using Workday Recruiting, mentions of basic AI skills such as simple prompting fell 25% after a January 2026 peak. Requirements for building AI tools, automating workflows and engineering AI systems rose 51% between September 2025 and July 2026. The two series use different time windows and categories. Together they describe employers asking for deeper technical work while an established route into that work grows less straightforward.

A job advertisement begins a hiring process. A graduate deciding whether to accept a long internship needs to know what comes after the advert: pay, supervised work and a date for a permanent-role decision. The manager needs to know who will teach that graduate. The finance leader needs to know where the labor bill lands. These reports can show where employers advertise demand; they cannot count the people who reached durable work.

Twenty-seven percent describes a posting share

The 27% figure comes from Draup’s Macro Labor Market Trends 2026. The company says the analysis draws on more than one billion job descriptions and one billion talent profiles from 2020 through August 2026. Its chart classifies technology demand into AI Builder, Forward Deployed Engineer, Guardrail, Infrastructure, Experience and Tech Support. The AI Builder share more than doubled since 2021 to 27%; Forward Deployed Engineer reached 18.5%. Those are shares inside Draup’s technology-posting taxonomy. They do not describe all Fortune 500 jobs or actual employees.

A career site can advertise a position several times, withdraw it, change its seniority or leave it open. Employers can fill work through internal moves and vendors without publishing a new vacancy. Draup’s classifications map stated demand. The published report does not trace each role-family label to a hire or six-month retention outcome. Its one-billion-description evidence base is much wider than the Fortune 500 technology-posting group behind the 27% figure.

The Federal Reserve gave the measurement problem a plain formulation in a September 30 research note. Economist Maria D. Tito used Lightcast’s online advertisements to study manufacturing skills. She distinguished that recruiting activity from the Bureau of Labor Statistics’ employer-reported vacant positions. The series have moved together recently, but an online posting remains an indicator of demand, not a direct count of openings. Completed hires sit another step away.

Draup’s role family also compresses several kinds of work. Its chart puts AI Builder beside Forward Deployed Engineer and a separate Guardrail family. A builder may create a retrieval layer, maintain an inference service, test an agent, or integrate a model into an existing product. Some work requires deep infrastructure skills; some requires sector knowledge and sustained review of failures. A common label helps compare advertisements, but it cannot price the training time or explain who handles the system after launch.

In Draup’s skills chart, demand for LLM fine-tuning rose 142%, agentic orchestration 128%, retrieval-augmented generation engineering 116%, and prompt and context engineering 97%. Manual data labeling fell 44% in its tracked taxonomy. These are changes in classified posting signals across the report’s observation period. They do not measure task automation inside every employer. A company can change the wording in an advertisement before it changes the work.

Vijay Swaminathan, Draup’s chief executive and a listed author of the report, argues for building and retaining talent internally where possible because senior AI workers carry a pay premium. The recommendation comes from a company that sells workforce intelligence. Its commercial interest does not make the posting data useless. It makes the denominator and conversion path especially important when a buyer uses the report to approve a training budget.

Internships are stretching beyond a summer

The early-career figures in the same Draup report refer to descriptions requiring zero to three years of experience. Internship mentions rose from 9% in 2020 to 19% in 2026. Contract-role mentions rose from 4% to 8%. Adding those two reported shares produces the 27% combined figure. The report does not identify a cohort of people who held both kinds of positions or establish how many were later hired full time.

Duration changed too. The reported average internship length increased from 3.6 to 9.7 months, and the average contract length from 7.4 to 12.9 months. A nine-month placement can contain serious paid learning and a route into a permanent role. It can also leave the entrant waiting most of a year for a conversion decision. The published chart gives duration, without wage terms, mentor hours, insurance coverage or conversion rates.

Those omissions are material to both sides of a contract. A student might accept a nine-month internship because it includes production code reviews, access to specialists and a written path to a salaried job. The same duration is a weaker proposition if the work is ordinary delivery under a temporary label, with no conversion target and no time for instruction. The employer’s bill changes too: a contractor’s invoice may sit in a project or services budget while a staff role sits in payroll. A headcount report can appear restrained even when the same work consumes more paid labor.

Draup also finds a lower Fortune 500 early-career posting share than at its April 2021 peak, even as overall tracked postings rose. That describes the mix of advertisements. Interest rates, business cycles, outsourcing, changed posting practices and a preference for experienced workers can all move the entry mix. A more technical AI role could increase graduate opportunities at a particular firm if it funds a credible learning route. The report cannot assign a share of the shift to AI.

An alternative reading deserves room. Longer internships may protect beginners from being asked to run a fragile model in a live business process after a short orientation. A contract can let a worker sample employers and gain portable experience. Those benefits need evidence: pay, supervision, completed projects, credentials that another employer recognizes, and a record of full-time offers. Duration alone cannot tell a graduate whether the arrangement is an apprenticeship or a delayed staffing decision.

One existing model sets a higher bar than the word “internship.” The U.S. Department of Labor’s Registered Apprenticeship description joins a paid job to structured learning under an experienced mentor, supplemental instruction, progressive wages and a portable credential. An ordinary internship or contractor agreement does not acquire those terms by lasting nine or 13 months. Draup’s category counts do not say which advertised placements meet them. An employer could publish the elements for its own cohort, and a candidate could compare the offer against them.

There is an employer cost behind each element. Protected mentor time has to be assigned to a person whose client or production work will pause. Instruction can come from a college, a vendor or the employer, but someone must pay for it. A progressive wage schedule requires a role budget beyond the first month. The expense can be justified if an apprentice learns to inspect a retrieval result, diagnose a failed automation and explain a decision to a customer. A long temporary placement without those responsibilities may be cheaper this quarter while leaving the firm dependent on costly senior hires next year. Draup’s charts do not settle that trade.

In August, the Bureau of Labor Statistics counted about 7.1 million U.S. job openings and 5.2 million hires across all industries. Openings describe positions available on the last business day of the month; hires count movements onto payroll during the month. Neither figure isolates AI roles. Even a national government count cannot supply the missing hires behind Draup’s specialized posting share.

Another dataset sees builders, not basic prompts

On October 5, Workday published a second employer sample. Its skills-demand series covers de-identified requisition information across about 550 enterprise employers using Workday Recruiting from September 2025 through July 2026. The categories distinguish basic AI skills, including simple prompting, from hands-on building, automation and AI engineering. Basic mentions peaked in January 2026 and then declined 25%. Build-oriented mentions climbed 51% across the stated September-to-July window.

The result supports a narrower conclusion than a claim that AI literacy has become worthless. Workday measured employer wording inside one software customer base. A recruiter might stop specifying “prompting” because it is assumed, because the organization revised its job template, or because it wants someone who can evaluate and build a system. The reported fall cannot identify which mechanism dominated. Nor can the 51% increase tell us how many applicants met the bar, accepted offers, or stayed.

Workday’s employee survey supplies a separate view of the training gap. In a September survey of about 6,000 full-time employees, 79% said they knew which skills they needed to succeed, while 66% said their employer helped them develop those skills. The 13-point difference is a difference between answers to two survey questions. It is not a measured shortfall in training seats, and it does not describe the 550 recruiting-system employers as a matched cohort.

The company also found that mentions of management and leadership skills fell 7% and training skills fell 13% in its requisition data from September 2025 to July 2026. A hiring manager might interpret those movements as priorities changing. A worker might hear a harder instruction: deliver more specialized AI work with less explicit commitment to teaching or supervision. The job posting cannot settle what happens at the desk. It can tell the employer where to ask for an explanation.

Phil Willburn, a Workday vice president quoted in the release, said employees may remain in the same job while the job changes around them. That observation helps connect the posting figures to internal work. If a company buys AI Builder talent from outside, existing staff still need to maintain business context, review outputs and know when to stop a system. A skill that disappears from a public advert may remain essential inside the company.

Workday sells the recruiting and workforce software that supplied these observations. Its sample is a large set of customers, not a random draw of all employers. Draup’s proprietary cross-company ads and Workday’s platform requisitions should therefore be read as separate directional evidence. Treating 27%, 25% and 51% as a single funnel would manufacture a conversion rate no source measured.

An October 12 account by Microsoft HR director Lesley Gong describes a different route to the skill Workday says employers want. Microsoft’s internal Career Hub and Learning agent point employees toward gaps, training, mentors and projects. A Workforce Insights agent lets managers examine staffing and hiring activity. Gong says role transitions and career sponsorship still require human judgment. This is Microsoft’s description of its own tools and practice; it does not publish completed moves, wage changes or the number of employees who became AI Builders.

For a manager deciding between an external requisition and developing a current employee, the missing Microsoft numbers are the useful ones. How many eligible employees applied for a project? How many were given protected time to learn? Did the worker move into a funded role, or only complete a learning path? Internal development may never show up in Draup’s public-posting dataset. A weak external hiring signal could therefore coexist with real internal opportunity at one company. It could also mask a frozen job ladder at another. An internal skills product is a way to organize the choice; it is not evidence that the choice was funded.

AI-generated editorial illustration. The open gap represents the unmeasured route from an advertised opportunity to lasting paid work.

Factories provide a different wage test

Maria Tito’s Federal Reserve note makes the hiring change visible in a sector often described as less exposed to office AI. In Lightcast online postings through July 2026, approximately 11% of manufacturing advertisements mentioned AI-related skills, compared with 8% of postings across the economy. General computer skills appeared in about 35% of manufacturing posts on average, against 24% across sectors. Manufacturing accounts for about 6.5% of new hiring demand in the posting and openings datasets the note compared.

The mix inside manufacturing matters more than the word “factory.” Broad AI and machine-learning skill requirements increased, but generative AI appeared in under 1% of manufacturing postings. Among production occupations, which include assemblers, machinists and welders, generative-AI requirements were essentially absent through the first half of 2026. Much of the AI-skill signal therefore sits in a different slice of the sector from the image of a production worker prompting a chatbot at every station.

The note also compares advertised pay. Across manufacturing occupations, postings that required AI skills carried an average advertised wage about 70% above all manufacturing postings over the sample period. The gap was volatile. For production occupations with AI-related requirements, the reported average differential widened after 2023 and was roughly 30%. Neither number is a raise delivered to an existing worker. The comparisons are between advertised positions that may differ in occupation, employer, location and experience; only around half of recent postings had nonmissing wage information.

The wage finding tests a different commercial question from the Draup role-family chart. If employers pay more for AI-related capability in production or engineering ads, a plant manager must decide whether to recruit scarce specialists, train adjacent staff, or buy the capability through a supplier. The published wage premium alone cannot price any one option. Training consumes experienced worker time and machine access; external hiring may bring higher salary and longer ramp; a vendor contract can move the cost outside payroll while leaving the factory responsible for safety and throughput.

The Federal Reserve analysis is valuable precisely because it resists a neat automation verdict. Postings measure what manufacturers ask for, and advertised wages reveal what some of them offer. They do not show plant-level deployment, output per hour, layoffs, or the earnings of workers who were actually hired. A higher advertised premium can also reflect a small number of senior engineering positions. An employer cannot use the 70% figure as a promised return on every operator’s AI training.

The distinction holds when a technology-company hiring report is brought onto a factory floor. Draup’s “AI Builder” taxonomy and Tito’s manufacturing skill classifier are not the same occupation. A factory engineer who builds a predictive-maintenance model might fit both definitions; a machinist whose job ad mentions machine learning might not. The only honest comparison is at the level each dataset actually observes: advertisements, classified skills and posted pay.

A company target is a different kind of number

Cognizant offers a named employer against which to test the aggregate figures. In a September 7 U.S. workforce announcement, the services company said it planned to hire 1,500 U.S. college graduates during 2026 and scale its Frontier Certified Engineer and Frontier Business Operator workforce to a combined 15,000 people. It presented the roles as funded positions and described university partnerships and registered apprenticeships.

The two Cognizant numbers have different meanings. The 1,500 is a planned graduate intake. The 15,000 is a workforce target that can include existing employees retrained or reassigned. It is not a verified count of net-new jobs. Cognizant’s earlier July role plan splits the target into 5,000 Engineers and 10,000 Operators, with a first assessed cohort expected in the fourth quarter.

The company gives a client example: a two-person Engineer and Operator pod built 17 production AI agents for a food-service account-management workflow and, it says, reclaimed about 11 hours per account manager per week. The customer is unnamed in the announcement, and the public account does not show the original work sample, time study, error rate or cost of ongoing review. It describes the kind of delivery Cognizant wants these roles to own. It is not a measured average for all 15,000 planned practitioners.

There is a financial constraint behind the role design. Cognizant’s second-quarter results reported 356,700 total employees at June 30, down 900 from March and up 12,900 year over year. Revenue was about $5.5 billion, up 4.5% year over year. Management also described Project Leap, an operating-model change intended to fund AI and workforce investments while generating in-year savings. Company-wide headcount and margin figures cannot be allocated to the Frontier cohort. They do show that the new job-family promise sits inside ordinary staffing and cost decisions.

For a prospective graduate, the question is less whether a company has invented a new title than whether it has funded a starting position, assigned a supervisor and built a route into client work. For a buyer of Cognizant’s services, the question is who remains accountable when an AI system needs repair after its launch. A certification is an input to those outcomes. It does not substitute for a staffed delivery team or a worker’s progression.

This site’s earlier Cognizant analysis traced the graduate pathway and its proposed evidence file in depth. The September announcement does not close the missing result fields. A new aggregate hiring report should not be used to declare them closed either.

A conversion file from advertisement to paid work

Employers and researchers can publish a more useful account without exposing individual applicants. The unit is a cohort tied to a defined role family and period. Each stage needs its own denominator. The table below is an editorial measurement proposal, not a claim that Draup, Workday, the Federal Reserve or Cognizant already keeps these fields together.

Stage	Count or definition to publish	Decision owner	What it can establish
Advertisement	Unique live requisitions, dates, geography, role level and duplicate-posting rule	Recruiting operations	Demand expressed in public or internal listings
Selection	Qualified applicants, interviews, offers and acceptance, with one cohort definition	Hiring manager	Whether the advertised requirement matched reachable candidates
Paid entry	Employee starts, internship starts and contract starts, each separated	Finance and talent leader	How many people entered paid work, rather than an ad database
Learning	Paid instruction hours, protected mentor hours and supervised production tasks	Delivery manager	Whether a junior route contains a teachable job
Conversion	Contract or internship completion and full-time offers accepted at a fixed date	Program owner	Whether the temporary route leads to durable work
Assignment	Live project scope, named reviewer, incident route and client approval	Operating leader	Whether a credential translates into responsible delivery
Progression	Six- and 12-month retention, pay and level change, with exits counted	People analytics	Whether the role becomes a career rather than a launch label
Business result	Customer or plant baseline, errors, output, labor hours and total cost	CFO and service owner	Whether a staffing model contributes to an outcome

Several choices in this file prevent easy inflation. Deduplicating requisitions stops repeated postings from looking like repeated jobs. Recording paid starts separately from offers exposes hiring plans that never turn into employment. A contract conversion denominator should include everyone who began the eligible cohort, with exits and unfinished placements disclosed. Keeping both median pay and the role/location mix stops a wage premium from being treated as a universal raise.

Timing belongs in the file too. A three-month-old placement cannot be marked a failed 12-month conversion just because a dashboard runs today. The employer can publish the original cohort size, the number due for a decision, the number still active, the number that left and the number that accepted full-time work. That lets a prospective entrant distinguish a program with a genuine conversion route from one that keeps resetting the clock. It also lets a CFO compare a proposed permanent salary with the full cost of a long paid placement rather than the first invoice alone.

The file also makes training legible as a budget. A manager who promises a graduate review time should identify who supplies it and what client work that mentor cannot do during those hours. If an employer books a contractor’s labor under services spend, the finance view should still count it. If an AI tool removes repetitive tasks, the entry plan should identify which remaining tasks can safely teach judgment and who signs off on progression.

One result can improve while another worsens. A firm may fill specialized builder jobs faster by paying above-market wages, yet lose the internal ladder that would have produced future builders. It may provide long internships with excellent mentorship but convert few people because the permanent budget is capped. It may train current workers and reduce outside ads while increasing productive capacity. A single “AI jobs created” number would hide all three patterns.

Q4 brings the first cohort test

The earliest check will be mundane. Cognizant expects its first assessed Frontier cohort in the fourth quarter. Draup can update its posting classifications. Workday can watch whether its basic and builder skill categories keep diverging. The Federal Reserve can observe whether AI requirements and advertised wages persist in manufacturing. None of those updates alone will reveal how many beginners secured a lasting paid role.

An employer that wants its AI hiring story believed can publish the stages it controls: planned places, offers, paid starts, completed training, client or plant assignments and later retention. A graduate can ask for the same sequence before accepting a long internship. A customer can ask who reviews an agent after the initial engineer leaves. These requests do not require a prediction about the future of all jobs. They require a roster, a budget and a date.

The September reports show demand changing on paper. For a person entering the next cohort, the decisive document will be smaller: a start date, a named mentor, paid time to learn, and a position that still exists when the contract ends.

This desk-research analysis was published October 13, 2026. The reported percentages refer to distinct samples, time windows and definitions. No original employer or worker interviews were conducted.