The 1,721% jump in "agent orchestration" demand at big banks is real, but it starts from 108 mentions and ends at 1,967, about 1.4% of the AI job ads banks posted this year. The larger signal in the same data is that bank job ads cite governance skills roughly twice as often as model-building skills. If you are planning 2027 headcount, the scarce hire is the person who decides which agents run, what each one may touch and where a human signs off, and you will usually find that person inside your own engineering and operations teams faster than on the open market.
The numbers come from hiring-analytics firm Draup, which counted skills in public bank job listings for CNBC's October 2 analysis of JPMorgan Chase, Citigroup, Capital One and peers. Most coverage stopped at the headline percentage. The base rate and the pay table say more about what to do.
What Does the 1,721% Actually Measure?
It measures mentions of a skill in job postings, and the starting count was tiny. Agent orchestration went from 108 mentions in 2025 to 1,967 in 2026, against a total of 139,819 AI-related listings, which were up 49% on last year. Run the math: 108 out of roughly 93,800 listings in 2025 is about 0.1%; 1,967 out of 139,819 is about 1.4%.
Agent orchestration is the work of coordinating several AI agents toward one shared goal: deciding which agent handles which step, what data and systems each can reach, and how results pass between them and back to a person. It is a real discipline. It is also, by Draup's own count, a small one next to the keywords around it.
Look at the framework names in the same dataset. LangGraph mentions rose from 680 to 5,300, LangChain from 3,393 to 10,545, and prompt engineering from 3,595 to 11,368. Prompt engineering appears almost six times as often as agent orchestration. A job ad that lists LangGraph is asking for someone who can use a library; it says little about whether the bank has defined who owns an agent's decisions.
Two caveats apply to all of this. Draup's figures are skill mentions drawn from public listings and platforms including LinkedIn, so one posting can count toward several skills and reposted roles can count twice. And "mentions" is a recruiter's vocabulary. A rising count shows what hiring managers write into ads, and ads drift from the job.
Governance Skills Outnumber Model Skills Two to One
Bank job ads now cite governance skills about twice as often as the skills to build and run models. Draup counted roughly 16,000 governance-related skill references against about 8,400 for model training, deployment and operations. The governance figure is mostly three lines in the same table: responsible AI (1,026 to 7,764), AI governance (1,314 to 6,489) and AI risk management (335 to 1,538). Those three alone sum to 15,791.
That is about eight governance references for every agent orchestration reference. Business Model Analyst's breakdown of the same numbers lists what those governance roles involve in practice: permissions management, data access control, human-in-the-loop checkpoints, logging and third-party AI tool security.
Draup's CEO, Vijay Swaminathan, gave the reason in plain terms. "There is a lot of complexity in an enterprise. Sometimes these complexities are visible, but many times they are hidden. It takes a long time even to automate a simple process," he said, as quoted in the CNBC piece. He also pointed to "a renewed focus on soft skills like problem solving, creativity, ability to ask tough questions." Those are the skills of someone who knows the process well enough to say where an agent should stop.
The pressure is about to grow. JPMorgan's chief analytics officer, Derek Waldron, said in June that the bank's agents "don't just run for two or three minutes to carry out a goal or some instructions of a human, they can run for an hour or two," and described them working more like a "team manager than an individual worker". An agent that runs for two hours and delegates to other agents makes many more decisions between human checkpoints. Someone has to design where those checkpoints sit, and that work is what the governance line items describe.
Regulators Left Agent Governance to the Banks
The federal model risk guidance explicitly leaves agents out, so the agent side of this governance hiring is not answering a model risk template. Other obligations still apply, such as consumer protection, fair lending and the EU AI Act for banks operating in Europe, and they may be part of what is driving the job ads. SR 26-2, which the Fed, FDIC and OCC issued on April 17, 2026, replaced SR 11-7 and is aimed mainly at banks with over $30 billion in assets. Footnote 3 of the guidance itself reads: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." It adds that a bank's own risk management and governance practices "should guide the determination of appropriate governance and controls" for anything the document does not cover.
So there is no supervisory template for who signs off on an agent. Each bank is writing its own, which explains why the governance vocabulary in job ads is so varied: "responsible AI", "AI governance" and "AI risk management" are three names for overlapping work that no regulator has yet standardized. If you sit outside banking, this is still your situation. Nobody has told you what an agent control function looks like either, and the banks' hiring is the closest thing to a published reference design.
The strongest counterargument deserves a hearing. Governance keywords may simply be compliance boilerplate that legal adds to every AI posting, and the people actually hired may spend their days writing code. That is possible, and Draup's mention counts cannot rule it out. But the 2026 growth in the governance terms (657% for responsible AI, 394% for AI governance) far outpaced the 49% growth in AI postings overall, which means the phrases are spreading faster than the jobs carrying them. Boilerplate tends to grow at the same rate as the postings it is pasted into.
What Should You Pay an Agentic AI Engineer?
Budget for a modest premium over an ML engineer and a large one over a data scientist. Draup's median base salaries from bank postings in 2026 are $176,999 for an agentic AI engineer, $166,764 for a machine learning engineer and $134,986 for a data scientist. A generative AI engineer sits at $172,937, an AI manager at $182,957 and a generative AI manager at $190,836.
| Role (bank postings, 2026) | Median base salary |
|---|---|
| Chief data scientist | $235,158 |
| Generative AI manager | $190,836 |
| AI manager | $182,957 |
| Agentic AI engineer | $176,999 |
| Generative AI engineer | $172,937 |
| AI product manager | $167,748 |
| Machine learning engineer | $166,764 |
| AI engineer | $164,736 |
| Data scientist | $134,986 |
The gap between "agentic AI engineer" and data scientist is about $42,000, or roughly 31%. The gap between agentic AI engineer and ML engineer is about $10,000, or 6%. That smaller number is the useful one: the word "agentic" in a title buys a single-digit premium over an engineer who already ships production ML. For a non-bank reference point, the US Bureau of Labor Statistics puts the median data scientist wage at $120,230 as of May 2025, across all industries.
These are base pay only and exclude bonus and equity, which at the big banks can be a large share of total compensation. Treat them as a floor for a bank-competitive offer.
The Role to Write Is an Agent Operations Owner
The role banks struggle to fill combines process knowledge with the authority to constrain an agent, and that mix is easier to build than to buy. An agent operations owner is the person accountable for one agent workflow in production: which agents run, which systems and data each may touch, where a human approves, what gets logged and when the agent is switched off. It sits between the developer who builds the agent and the risk function that has to defend it.
Hiring this person from outside is slow for a structural reason. The hard half of the job is knowing your payment exceptions, your loan file or your claims queue well enough to see Swaminathan's "hidden" complexity. A candidate from another bank brings framework skills and none of your process knowledge. Your own senior developers, operations leads and second-line risk staff bring the process knowledge, and the framework skills are the faster half to learn.
Banks have already built the training machinery. Citi mandated AI prompt training for 175,000 employees across 80 locations in 2025, with training taking under 10 minutes for experienced users and about 30 for beginners. That is literacy, and it maps to the 11,368 prompt engineering mentions. The agent operations role needs a deeper track for a much smaller group: a few dozen people per business line who already own a process and can be taught to own the agent running it.
This is also where the pay data helps you. If you reskill a $134,986 data scientist or a process lead into the role and pay them partway toward the $176,999 agentic engineer median, you spend less than an external hire and keep the process knowledge you would otherwise have to rebuild.
What to Do Before the 2027 Headcount Plan Locks
This Week:
- Pull your own open AI requisitions and count how many name a framework (LangGraph, LangChain, CrewAI) versus how many name a decision the hire will own. If frameworks win, rewrite the ads before they go another cycle.
- List every agent you have in production or pilot and write next to each the name of the person who can switch it off. Blank lines are your hiring plan.
This Month:
- Write one agent operations owner job description with the five duties above: scope, data access, approval points, logging and shutdown. Have your risk lead and an engineering manager both sign it.
- Pick three to five internal candidates per business line from developers and process owners who already handle exceptions. Get your HR partner to price a reskill track against the external median.
Before Budget Sign-Off:
- Set pay bands using the Draup medians as a floor: around $167,000 to $177,000 base for engineers who build agents, and a separate band for the owner role tied to the process it governs.
- If you are a bank over $30 billion in assets, document in writing how your governance practices cover the agents SR 26-2 excludes. The guidance names no standard for them, so the one you apply has to be your own and on paper.
The Bottom Line
The 2026 bank hiring data reads like the early cloud years, when "AWS" in a job ad mattered less than whether anyone owned the account and the bill. Agent and framework terms grew fastest, but the volume sits with governance, and the regulator has left the definition of that governance to each bank. That leaves you free to define the role around your own processes and staff it from people who already know them.
Before you post a LangGraph requisition, find out who can switch off the agents you already run.
Continue Reading
- Human-in-the-Loop for AI Agents: Most Approval Gates Rubber-Stamp
- Best MLOps for Regulated AI: Domino, Then a Sign-Off Layer
- BMW Bets 100 Senior Jobs on AI Agents That Still Need a Signer
- At Sibos 2026, Bank AI Agents Fix Payments a Person Still Releases
- Claude Frontier Academy Wants Your Best Engineer and Names No Price
- Agent Orchestration Platforms: Score Exit, Not Features
- Agent Audit Logging: Platform Logs Miss the Decision You Must Prove