Ask an AI lab where its workforce is and you get a map, not an address. Researchers in London and Zurich. Engineers in Tel Aviv and Bangalore. Robotics technicians at sites that did not exist a year ago. And underneath all of it, the people who label the data, review the outputs and keep the models honest, a workforce that market researchers Dataintelo and Market.us estimate at roughly 3 million today and project toward 35 million as LLM training, reinforcement learning from human feedback and computer-vision crowd work scale. The supply chain behind it is already sprawling: a March 2026 report by SOMO counted almost 500 companies active in AI data collection and labelling, with Amazon, Google, Meta, Microsoft and Nvidia using at least 30 intermediaries between them.

Almost none of those people are employees in the traditional sense. They are contractors, agency staff, gig workers, quasi-workers with one client and no benefits. Staffing Industry Analysts‘ workforce benchmarks, cited by Conexis VMS, put contingent, freelance and temporary labour at 18 to 22 percent of enterprise workforces today and track a structural shift toward 40 to 48 percent among Fortune 500 and Global 2000 companies. Workday’s own VMS unit has told HR Executive that non-employees already average 36 percent of Fortune 500 workforces and are expected to reach 50 percent by 2027. Estimates of the global gig workforce vary widely by definition; Papaya’s own projection, presented to analysts, is growth from about 100 million to 435 million.

Whatever the exact numbers, the companies building on that workforce have a problem that has nothing to do with models. They need to recruit, classify, manage and pay people wherever the right talent exists, under whatever employment model fits, and do it faster than the regulations governing that work can change. Most of the tools available to them were built for a simpler question: how do I hire one engineer in Portugal?

An industry rebuilding itself

The global-employment market of the last five years was defined by employer-of-record platforms, a category that grew to include a dozen well-funded vendors, Papaya Global among them. The pitch was speed. Hire abroad without opening an entity, and let the vendor handle the employment through local partners and the payments through third-party rails.

That model worked when international hires were occasional and the risk was small. The question the whole category now faces is what happens at scale, when a company is standing up a data-labelling operation across eight countries or running a robotics workforce on three continents, and the rules are the product. By Papaya’s count, 48 countries require staffing licences to lease labour at all. Thirty-two cap how long an EOR arrangement can run, 18 months in Germany, 36 in France. Sixty-five apply strict contractor-classification tests and 28 actively enforce them. Misclassify a room of annotators in one of those and the bill is not a fine; it is back pay, benefits and, in some jurisdictions, a criminal referral.

The risk also enters earlier than payroll. TransUnion’s 2026 Gig Economy Worker Report, a survey of 1,012 US adults published in January, found that one in four gig workers have rented their verified accounts to others and one in five have sold them, rising to 31 percent of Gen Z and millennial earners for renting. Only 45 percent said the platforms they use have effective identity verification. Deepfaked identities defeat static onboarding checks. Project roles drift into indefinite employment without anyone deciding they should. A vendor breach exposes every client’s workers at once.

The vendors are responding along different lines. Some have expanded from EOR into contractor management, payroll for owned entities and broader HR suites, and now emphasise owned legal entities in the countries they serve. Others bundle global payroll into all-in-one HR, IT and finance platforms aimed at companies that want a single system of record. Several have pushed into contractor-of-record services to absorb classification risk. Papaya’s distinguishing move was to become a regulated payments company, and that is the version of “i” it presented to analysts on September 9.

Two cases the whole category is reading

Two court rulings, neither involving a payroll vendor, are being cited across the industry as a signal of where liability is heading. In 2024 a Canadian tribunal ordered Air Canada to pay $812 after its chatbot invented a policy, dismissing the argument that the bot was a separate entity (Moffatt v. Air Canada, 2024 BCCRT 149). In May 2025 a US federal court granted preliminary certification of a nationwide collective action in Mobley v. Workday on the theory that an AI hiring vendor can be liable as an agent of the employer; the case is in discovery through 2026.

Neither case is about paying contractors in Portugal. What they establish is a direction: when software makes a decision, the vendor behind it may own the outcome. Every EOR platform now uses AI somewhere in classification, onboarding or payroll validation, so the question of who is accountable when it is wrong applies across the category. Papaya has chosen to make it the centre of its pitch, which is a commercial decision as much as a legal one.

What Papaya says its infrastructure looks like

Papaya’s answer is to own the layers others partner for, and to connect them. The following is as presented at its Analyst Day; all figures are company-reported.

One system for every worker type. Sourcing and day-zero screening, classification, employment through owned entities or licensed partners, contractor and agency management, payroll for a company’s own entities, and payments, in one platform across 180 countries, 110 with native gross-to-net and EOR coverage, in 130 currencies. Employees, contractors, agencies and contingent workers sit in one record.

Money on rails it owns. Papaya Global acquired the licensed payments company Azimo in 2022 and holds six money licences across the UK, Europe, Hong Kong, Australia, Canada and the US. Client funds sit with J.P. Morgan and Citi. It says more than $50 billion a year moves through its rails and 95 percent of direct payments settle in real time. Most vendors in the category execute payouts through partner providers; whether owning the rails produces materially better outcomes for customers is the claim analysts will want evidence for.

AI that executes, experts who sign. Five agents, Connect, Comply, Hire, Validate and Pay, run on a single data layer. A compliance engine, Papaya ONE, grounds decisions in local regulation, the company’s own policies and cited sources, and flags what it will not decide alone. Above the agents sits a layer of in-country experts who handle ambiguous cases and approve anything consequential before it executes. Papaya reports support tickets down 24 percent and manual checks per payroll cycle down 55 percent on its own operation; these are internal measurements, not audited results.

Engineers before the contract. Forward-deployed engineers map a customer’s operation and build customised agents during migration. The stated ambition is 40-plus countries and 20,000-plus workers in under eight months.

A worker-side account. Banco gives each worker an account and card, with earnings landing in seconds, no local bank account required, and, by Papaya’s figure, up to 80 percent lower cost per payout. Other vendors also offer worker-side wallets or cards; the differentiator Papaya claims is that its own licences and rails sit underneath.

What is not yet proven

Three things the presentation asserts and the market has not yet tested.

The demographic projections are directional, not settled. The 35 million annotator figure and the 40 to 48 percent contingent share come from market-research firms and staffing-industry benchmarks with their own methodologies; the 435 million gig figure is Papaya’s own. The direction is widely agreed. The magnitudes are not.

The forward targets are targets. Papaya told analysts it expects threefold AI productivity in 2026 and more than 35 percent of revenue from autonomous agents and services by 2027, and labelled both as such.

And “owning the outcome” is a contractual posture until a court tests it. Papaya reports zero compliance claims to date. That is a strong record. It is also one that only moves in one direction, and no vendor in the category has yet been tested in court on it.

Building around opportunity, not geography

Strip away the vendor positioning and the underlying claim holds up: the AI economy is creating a workforce that is distributed, contingent and paid across borders, and the tools built for occasional international hires are being rebuilt for it. Most of the category is converging on broader platforms from the software side. Papaya is converging on the same destination from the payments and licensing side. Which approach wins will depend less on analyst-day slides than on which vendor is standing when a regulator, or a court, asks who was responsible.

Build anywhere. Employ everywhere. Pay in real time. Every vendor in the category now says some version of it. The one that can say “and we’ll answer for it” and mean it is the one the next decade of work will be built on.

Papaya Global figures are as presented at its Analyst Day on September 9, 2026 and are company-reported unless otherwise stated. Workforce projections are sourced to Dataintelo and Market.us (data annotation market research), Staffing Industry Analysts via Conexis VMS (contingent workforce benchmarks), HR Executive (Workday VNDLY estimate) and TransUnion’s 2026 Gig Economy Worker Report (credential sharing). Court case details are from Moffatt v. Air Canada, 2024 BCCRT 149, and Mobley v. Workday, N.D. Cal. No. 23-cv-00770-RFL. Category descriptions are from vendors’ public product pages.