Citadel's head of Global Quantitative Strategies told Bloomberg on September 25 that his 180-person group was recruiting from AI labs including Google DeepMind — not to build models, but to direct automated systems and "manage machines." Goldman Sachs posted a VP-level AI Safety Platform Engineering role in New York at $130,000–$250,000 base, with a mandate to build a "unified company-wide agentic evaluation framework." While the AI trainer economy's canonical story runs through Mercor contractors in Cambridge and Outlier annotators in San Francisco, a parallel evaluation infrastructure has been assembling across Midtown Manhattan — and it pays differently, recruits differently, and solves a different problem.
The Infrastructure Driving the Demand
JPMorgan's LLM Suite has reached more than 200,000 daily users across more than 450 production AI use cases as of mid-2026, according to internal materials cited by multiple outlets. The bank targets 1,000 use cases by year-end, with a $18 billion annual technology budget funding the expansion. Goldman Sachs has deployed its GS AI Assistant firm-wide. BlackRock's Aladdin platform covers approximately $21 trillion in assets under supervision, with AI risk tooling embedded at every layer of portfolio management.
Each of those deployments requires human evaluation infrastructure that the traditional AI trainer economy isn't built to supply. Mercor's senior ML annotators — earning $94/hr at the Q3 2026 median — are calibrated to evaluate whether a model's reasoning about a constitutional AI question or a chain-of-thought problem is sound. They are not calibrated to determine whether a risk model's output violates a bank's SR 11-7 model validation obligations, whether a wealth advisor's AI assistant has breached a fiduciary line, or whether an agentic system should have halted before executing a trade.
The finance AI evaluator sits at the intersection of two expertise sets that rarely overlap: deep familiarity with AI system behavior, and internalized command of the regulatory and institutional context in which that behavior has consequences. That intersection is narrow. The demand for it is not.
The Comp Architecture
Goldman Sachs' VP — AI Safety Platform Engineering posting specifies building a "unified company-wide agentic evaluation framework," real-time LLM guardrail services, and automated governance controls. Base pay: $130,000–$250,000 annually, in New York. That is not a contractor rate. It is a W-2 base with Wall Street equity and bonus structure stacked on top.
JPMorgan's comparable role — AI/ML Technology and Cybersecurity Risk Management Vice President — posts at $147,250–$260,000 base in New York. BlackRock's Risk Strategy and AI Transformation Director runs $173,500–$270,000. Morgan Stanley's AI Oversight and Governance VP sits at $110,000–$190,000. Citi's Model Identification and Assurance Review, at the assistant vice president level, starts at $77,000–$116,000.
Across the sector, bank AI job postings totaled 139,819 in 2026, up 49% year-over-year, according to data from American Banker and PYMNTS. The fastest-growing skill category within those postings: agent orchestration — designing and directing multiple AI agents working in concert — which saw demand rise 1,721% across bank job ads in 2026. Wall Street now wants nearly two governance and evaluation skills for every model-building skill, according to analysis by BusinessModelAnalyst, reflecting the regulatory pressure arriving faster than internal hiring can absorb it.
The comp comparison to the Mercor ecosystem is instructive. A senior ML annotator on Mercor earns $94/hr at the Q3 2026 median — approximately $195,000 annualized at full-time hours. A Goldman Sachs AI Safety Platform Engineering VP earns a base of $130,000–$250,000, before Wall Street's bonus and deferred compensation structure. Both roles evaluate AI system behavior for a living. The inputs to the job — domain knowledge, understanding of model failure modes, the ability to write corrective guidance a technical team can act on — overlap substantially. The credential profile, the institutional context, and the output format do not.
Citadel and the Lab-to-Finance Corridor
Citadel's September push is the most visible confirmation that the finance AI evaluation market has reached escape velocity. The GQS group — Navneet Arora's 180-person research-and-data-engineering team — plans "healthy double-digit" headcount growth over the next year, per Arora's Bloomberg statement. A new team focused on systematic equities trading was launched under Alexey Poyarkov, who joined from TGS Management.
What distinguishes this expansion from prior quant hiring cycles is the explicit targeting of AI lab researchers. Citadel is approaching researchers at labs including Google DeepMind with a specific pitch: the ability to develop investment ideas and apply AI systems to trading at institutional scale. "People can now effectively manage two, four, and in some cases even more processes," Arora told Bloomberg. "Humans are now also becoming 'managers of machines.'"
That framing — evaluators who direct automated systems — is operationally identical to what Anthropic calls a "human oversight specialist" and what METR calls an "evaluation researcher." The difference is domain. Citadel's version evaluates AI trading systems against investment hypotheses and risk parameters. Anthropic's version evaluates frontier model outputs against alignment criteria. The underlying competence — understanding what the model is doing and why it's wrong — is the same. The career path and the paycheck are different.
Bridgewater has moved in a related direction from the opposite angle. Its AIA Labs team has built fine-tuned models that beat GPT-class performance on document-filtering tasks at 13.8x lower cost, according to the firm's own reporting. Building those fine-tuned models required the same evaluation-and-feedback loops that define the AI trainer economy. Bridgewater did it internally, with finance-domain experts, for finance-specific applications. The trainer economy framework applies. The Mercor marketplace does not.
Three Things to Watch in Q4 2026
Whether Goldman's evaluation framework becomes a shared infrastructure. The VP — AI Safety Platform Engineering role at Goldman specifies a "company-wide" agentic evaluation framework. If that framework stabilizes into a reusable architecture rather than a bespoke build, Goldman could become a reference implementation for how financial institutions structure AI evaluation governance — a role it has played before with trading infrastructure like SecDB and Marquee.
Whether the Citadel-to-lab talent pipeline develops bidirectionality. Current flow runs one direction: frontier lab researchers — particularly from DeepMind — moving toward finance AI roles for the comp and the applied challenge. If finance AI evaluator roles become a credentialing path in their own right, the pipeline may eventually run both ways. A researcher who spends three years evaluating AI trading systems at Citadel holds a specific, demonstrable AI safety record. Frontier labs evaluating that profile for red-team or safety evaluation roles would have a legitimate reason to hire it.
Whether regulatory acceleration steepens the demand curve. Bank regulators in the US are extending SR 11-7 model validation requirements to cover generative AI deployments. Each regulatory update creates a new evaluation requirement that needs a credentialed human to assess. The 49% job posting surge in 2026 is, in part, a regulatory response. If guidance from the Federal Reserve, OCC, and FDIC continues arriving faster than internal hiring can absorb it — a pattern that has held through 2026 — the evaluation function at major banks will become a structurally larger share of the total AI headcount.
The Mercor economy is about training frontier models from the outside in. The Manhattan evaluator economy is about deploying institutional models safely from the inside out. Both fall within the AI trainer economy's scope. They're just solving different parts of the problem — and paying differently to get there.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.