A percentage is not a quantity. It is a confession about a denominator.
I learned that at eighteen, hunched over fifteen ICO whitepapers in a university library in Tokyo. Every one of them advertised a number engineered to make me stop thinking — a 400% community growth rate, a 1,000% token appreciation projection. When I traced each figure back to its base, four of those projects collapsed under their own governance. The vesting schedules favored insiders. The "community" was three Telegram admins and a bot. So when a headline crossed my desk this month claiming Wall Street banks had boosted AI hiring by 1,721% for agent and governance roles, I did not feel excitement. I felt the familiar pull of an audit.
Context: job titles as fossils
Job titles are fossils. They preserve the shape of an organization's anxiety at a specific moment in time. When a bank posts a "Senior AI Governance Lead," it is not announcing a vision. It is confessing a pressure.
That pressure has a name and a date. The EU AI Act entered force in August 2024, classifying credit scoring, employment screening, and several financial decision systems as high-risk — which in regulatory language means mandatory human oversight, immutable documentation, and demonstrable transparency. In the United States, the Federal Reserve's SR 11-7 model risk management guidance has quietly governed quantitative models for a decade, and its logic now extends to machine learning systems that no one can fully explain.
Meanwhile, "agent" has crossed from research paper to procurement line item. The Agentic AI transition has its own timeline. Through 2023 and 2024, agents were demonstrations. By 2025, they were pilots inside sandboxes. The hiring language now suggests a third phase — deployment into workflows with real money and real customers on the other end. That is a meaningful threshold, and it is the strongest part of the story, even though the story never argues it.
Add the market context: we are in a bull phase, and bull phases reward narrative over architecture. Every funding round I have watched this year has a governance slide. Most of them are decoration.
Core: the arithmetic, and what the arithmetic hides
A 1,721% increase is a multiplier of 18.21. That is all it is. It tells you nothing until you know the base.
If the baseline year saw ten such postings, the surge produces 182. If it saw five, the surge produces 91. The categories in question — agent and governance — did not meaningfully exist as job families three years ago. Any growth rate computed from a near-zero base will produce an astronomical percentage while the absolute increment remains a rounding error inside a bank that employs 300,000 people.
The report never disclosed the absolute headcount. It never disclosed the time interval. It never disclosed whether the count came from a recruitment platform's listings, which sample only jobs advertised and only on that platform. That is not a data point. That is a vibe with a decimal point.
The second thing the number hides is what kind of work these roles contain. In a regulated institution, an AI agent is rarely an autonomous system that reasons and acts. It is far more likely a hybrid: robotic process automation holding the workflow, a language model holding the natural-language interface, and a human holding the liability. Banks do not deploy systems that can surprise them, because surprise is the one thing a regulator will not forgive.
The third thing it hides is the direction of the money. Hiring governance staff is a cost center, not a revenue center. Governance headcount grows when compliance obligations grow — not when profits grow. Reading a compliance-driven hiring spike as evidence of AI value creation is like reading a fire department's budget increase as evidence that fires are profitable.
And here is where I part company with the optimism. If Wall Street banks are hiring aggressively to build internal AI capability, the most likely loser is the pure API vendor. JPMorgan has built its own LLM suite. Goldman Sachs has internal tooling. The largest buyers of enterprise AI are, one integration cycle at a time, becoming their own suppliers. A bank that internalizes its intelligence layer stops paying rent on someone else's.
I have seen this movie before, and it was called PayPal and PYUSD. The company did not wait to be regulated into irrelevance; it moved first, became a licensed partner, and turned a compliance burden into a moat. Banks hiring governance leads are making the same move. They are not buying safety. They are buying the right to keep operating.
Now let me bring this home to the industry I actually live in. AI governance inside a bank is a committee, a policy document, and an audit trail stored in a system nobody outside the institution can inspect. It is trust requested, not trust proven. A committee cannot be queried. A hash can. When a bank's model risk team signs off on a credit model, the evidence lives in a folder a regulator may or may not open. When a protocol commits a model attestation on-chain, the evidence is available to anyone with a block explorer and the patience to look. That asymmetry is not a philosophical preference. It is an architectural one, and architecture is what survives contact with an adversarial market.
Truth is not consensus, it is verification. That distinction is the entire distance between a governance department and a governance protocol.
There is a genuine market forming here, and it is not the one the headline suggests. AI audit tooling, model risk management software, compliance-as-a-service for regulated models — these are the RegTech surfaces a 1,721% hiring spike actually reveals. When institutions hire that fast for a capability, they are also admitting they cannot hire their way to it. They will buy tooling. And the tooling they buy will increasingly need to produce proofs, not reports.
We build walls of code to protect hearts of flesh. The walls are going up on Wall Street right now. The question is whether they will be transparent walls or opaque ones.
What would change my read? Three numbers. The absolute headcount in both years. The share of those roles classified as engineering versus oversight. And the ratio of internal build to third-party procurement. Give me those three and I can tell you whether this is a demand signal for the AI industry or a supply signal of internalization. Without them, every confident conclusion is a guess wearing a suit.
Contrarian: the blind spot nobody is naming
Here is the angle the coverage missed entirely, and it cuts against both the bulls and the bears.
The hiring data measures intent, not capability. And in governance, intent without capability is a liability, not an asset. A bank that posts forty governance roles and fills them with people who cannot read a model card has not reduced its risk. It has added a layer of plausible deniability that will look, in a post-incident deposition, exactly like negligence.
I call this governance theater, and I have watched it in crypto for a decade. The ICO projects of 2017 had advisory boards. The NFT collections of 2021 had ethics councils. The structures existed to be photographed, not tested. During DeFi Summer 2020, I organized thirty volunteers to translate Aave and Compound documentation into plain Japanese, and when one protocol we had recommended took a flash loan hit, the panic was contained not by a committee but by a clear explanation published within hours. Governance that works is governance that gets exercised under stress. Everything else is furniture.
There is also a talent mismatch nobody is pricing. Governance roles require regulatory fluency, not machine learning fluency. The people who understand SR 11-7 are not the people who understand transformer architectures, and the people who understand both are rare enough to name individually. A 1,721% increase in postings against a fixed global supply of qualified people does not produce 1,721% more governance. It produces wage inflation and title inflation.
The blind spot, then, is not that the number is small. It is that the number is being asked to do work it cannot do — to stand in for capability, for budget, and for value.
Takeaway
The ledger remembers what the crowd forgets. In three years, the banks that hired governance staff and actually built verifiable systems will be distinguishable from the banks that hired governance staff and built slide decks. The difference will not appear in a hiring chart. It will appear in the incident reports they never had to file.
So here is the question I would put to anyone reading a percentage like 1,721% and feeling the pull of a trade: when the audit comes — and it always comes — will your governance be a signature, or a proof?
