A note on where this comes from: I have no relationship with Citigroup — financial, employment, consulting, or otherwise — and no non-public knowledge of its AI organization. Everything below is read from five of Citi's own public job postings and one April 2026 corporate blog post, all listed in full at the foot of this piece — not Citi's internal org chart, which nobody outside the firm has seen. That's not a limitation to apologize for: it's the same vantage point anyone sizing up a competitor's, a partner's, or a prospective employer's org design ever has, and it's usually enough to see the shape of a structure without the blueprint. What follows reads that shape deliberately, stated once, here, rather than re-qualified in every paragraph below.
In July 2026, Citi posted a job opening for Head of Applied AI & Agent Factory — Managing Director, C-16, $250,000–$500,000, New York. The posting describes a role that operates "an industrialized capability for the design, build, evaluation, and certification of enterprise AI agents," owns "the official enterprise agent catalog," and is measured, in year one, on "a working delivery-to-platform feedback loop demonstrably reducing duplicate build across the firm." It names two peer functions explicitly: a Head of Core AI Platform, who provides "the agentic runtime, guardrails, and evals," and a Head of Responsible AI, who "owns approval and controls." All three sit under a newly created Group Head of AI, alongside a fourth peer — a Head of AI Strategy, whose own posting describes ownership of the firm's investment thesis, board narrative, and build-vs-buy calls, but not delivery itself. Four functions, cleanly divided on paper: set the strategy, build the runtime, deliver into the business, govern the risk.
Citi posted several adjacent roles in the same window, and together they make the division more specific, not messier — which is the more interesting outcome. The role built to serve the platform pillar for one business line — corporate and investment banking (CIB) — is titled "Applied AI & Agentic Platform Engineering," with technical requirements spanning knowledge graphs, RAG pipelines, and multi-agent orchestration, built specifically for that business line, on its own GCP stack. The Responsible AI Lead's posting is explicit about which of banking's three lines of defense it occupies: First Line, operationalizing controls that Second Line (Independent Risk) and Third Line (Internal Audit) hold it accountable to, with named ownership of a firmwide "AI Command Center" for real-time monitoring of every agent's performance, fairness, and compliance. That's a level of structural specificity worth sitting with — Citi isn't inventing new governance vocabulary for agents, it's routing agent oversight through three-lines-of-defense scaffolding that already exists for every other risk domain in the bank, which is a genuinely different design choice than treating AI governance as its own novel discipline.
What "AI Factory" actually means right now
"Agent Factory" isn't a coinage unique to this posting — it's Citi's specific entry into a term that's being defined in at least three incompatible ways across the industry in 2026, and the posting is worth reading against all three.
NVIDIA's usage is the most literal: an AI Factory is infrastructure, full stop — a validated hardware-and-software reference design that turns data and electricity into tokens at scale, the same category of thing a physical factory is to raw materials. McKinsey's usage is organizational: a standardized, repeatable pipeline — ingest, develop, validate, deploy, monitor — built to cure what the industry now calls "pilot purgatory," the state where a majority of agentic pilots never reach production. The scale of that problem is why the term has momentum: multiple 2026 analyst estimates put pilot-to-production failure rates somewhere between 70 and 90 percent, and Gartner has projected more than 40 percent of agentic AI projects will be cancelled outright by the end of 2027, citing cost, unclear value, and inadequate risk controls. Microsoft has its own "Agent Factory" branding, closer to a tooling-and-certification program than either of the above.
Citi's Head of Applied AI & Agent Factory role reads closest to McKinsey's sense — the catalog, the certification process, the lifecycle management. But the CIB-specific platform engineering role sitting one rung below it is squarely NVIDIA's sense — GPUs, orchestration frameworks, a from-scratch technical build. One title, two senses of a term the industry hasn't finished arguing about, deployed inside a single org chart without apparent friction. That's not a contradiction so much as a sign of how early this vocabulary still is: the same word is doing infrastructure work in one job posting and organizational-process work in the next, and nobody involved seems to have noticed they're not quite the same claim.
The Applied AI cadre is Forward Deployed Engineering, under its own name
The other half of the Agent Factory role — "build and lead a cadre of Applied AI engineers embedded directly into business lines" — is a specific, well-documented industry pattern wearing a Citi-specific label. Forward Deployed Engineering, the practice of embedding engineers directly inside a customer's or business unit's workflow rather than shipping software over a wall, has gone from a niche Palantir-associated term to a genuine hiring boom in the space of about eighteen months. Forward-deployed engineering postings on Indeed grew roughly 729 percent year over year between April 2025 and April 2026; LinkedIn data shows demand for the role increasing roughly 42-fold since 2023. Amazon Web Services committed $1 billion to stand up its own Forward Deployed Engineering organization in 2026. OpenAI spun out a separate Deployment Company backed by more than $4 billion, absorbing roughly 150 experienced forward-deployed engineers through its acquisition of Tomoro.
What's specific to Citi's version, and worth noticing precisely because most FDE coverage skips it, is the return path. AWS's announcement, OpenAI's, and most of the trade coverage all describe embedding engineers with the customer; almost none of it describes what's supposed to flow back the other way. Citi's posting makes the return path an explicit, named, measured function — "a deliberate 'delivery-to-platform' feedback loop" graded on "reducing duplicate build across the firm" as a first-year success metric. Embedding engineers with the business is the easy half of this model to describe in a press release. Building a real mechanism for what comes back from that embedding, and grading someone on it, is the harder half, and it's the half Citi's posting is unusually specific about wanting.
The fourth function nobody's named
All three of these mechanisms — the certified catalog, the three-lines-of-defense governance, the delivery-to-platform loop — share a dependency none of the postings assigns to anyone: some model of what's true across the firm that every certified agent, in every business line, is checked against. Without it, "certified" and "cataloged" and "fed back into the platform" can all be true of ten agents that quietly disagree with each other about what a customer is, and nothing in what's publicly described would catch that until it surfaced in production.
That's a real question — what kind of shared memory an enterprise owes the agents built on top of it, who owns it, and how it's governed. It's also too big a question to answer honestly in a closing section of a piece about org design, so this piece won't try. It gets a full piece of its own: The Backplane Isn't One Memory. It's Four., which argues the shared-memory layer these three functions need isn't one thing to build, but at least four.
The honest caveat
Two limits worth stating plainly rather than qualifying away. First, the methodological one already named: five job postings and one corporate blog post are not an org chart, and the gaps identified above are gaps in what's publicly describable, not confirmed gaps inside the firm. Second, "AI Factory" itself is a term still being fought over by NVIDIA, McKinsey, and Microsoft simultaneously, and mapping Citi cleanly onto any single one of those definitions would flatten a genuine, unresolved ambiguity in the industry's own vocabulary. The more accurate reading is that Citi's structure borrows pieces of all three uses of the term at once — less tidy than a single-definition story, and closer to what's actually happening across every large enterprise racing to build the same thing this year.
What is clear, and doesn't need much hedging, is the shape of the bet: a $200 billion bank has decided that AI-led transformation is worth a C-16 seat built specifically to industrialize agents the way software has been industrialized for a decade, staffed by people whose job titles didn't widely exist two years ago. Whether that bet pays off will show up first in whether the catalog actually compounds or just accumulates — and that, in turn, comes down to the fourth function nobody's named yet.
Sources. Citi job postings and materials: "Head of Applied AI & Agent Factory – Managing Director," Citi Careers, jobs.citi.com; "Director, Applied AI & Agentic Platform Engineering," Citi Careers, jobs.citi.com; "Managing Director – Responsible AI Lead – C16 – NY," Citi Careers, jobs.citi.com; "Senior Vice President, AI Product Management," Citi Careers, jobs.citi.com; "Head of AI Strategy – Firmwide AI," Citi Careers, jobs.citi.com; "Introducing AI Agents: The Next Phase in Our AI Journey," Citigroup, April 2026, citigroup.com. On AI Factory definitions: NVIDIA, "AI Factories: The New Infrastructure of Intelligence," blogs.nvidia.com; coverage of McKinsey's enterprise AI framing, AI Conference London, 2026; Microsoft Agent Factory, microsoft.com/en-us/ai/agent-factory. On pilot purgatory and adoption-gap statistics: AgentMarketCap, "The Enterprise Agent Deployment Maturity Model 2026," agentmarketcap.ai. On Forward Deployed Engineering: Forbes Technology Council, "Beyond The Proof Of Concept: How Forward Deployed Engineering Accelerates Enterprise AI Adoption," February 2026; Techstrong.ai, "AWS Launches Forward Deployed Engineering Team to Speed AI Adoption"; AIwire/HPCwire, "OpenAI Launches Deployment Company to Scale Enterprise AI Adoption," May 2026.