A traditional quant fund's org chart looks like this: tens to hundreds of PhDs, divided along a pipeline — a data team cleans data, a signals team mines factors, a portfolio team optimizes, an execution team runs trading, a risk team watches exposure, and an operations layer handles reporting and compliance. The implicit assumption behind this shape is: research capacity = headcount of qualified researchers × output per head.
Over the past few months, we've refined this discipline against the full live pipeline and started testing a different assumption on small-money live capital: once the research process itself is frozen into an executable discipline, that equation stops holding. Enumeration, verification, monitoring, reporting — most steps on the pipeline can be run around the clock by AI agents, leaving only a small number of genuinely judgment-requiring decisions to a human.
This piece is about the design principles of that shape. It isn't a story about "using AI to be more efficient" — efficiency just means doing the same thing faster. What's happening here is structural substitution: certain roles aren't being accelerated — they stop needing to exist as roles at all.
Precondition: Discipline Before Automation
The most important point first, because it determines whether this shape holds together at all: the precondition for AI to take over a step is that the step's rules have already been made fully explicit.
Take signal research as an example. In the previous piece I described our admission discipline: every candidate signal must have factor exposure stripped out, clear a pre-registered out-of-sample test, be net-of-cost positive, and pass a tail-correlation gate — four gates, rules frozen before the data is seen, verdicts that depend on no one's discretion.
Notice a byproduct of this discipline: once a verdict requires no discretion, it no longer requires a human. Enumerating candidates, pulling data, running all four gates, writing the verdict, recording the conclusion in a ledger — the entire chain is pure execution. AI agents can screen overnight what a human researcher would take a month to get through, and they never get tired after the tenth candidate, never go easy on an idea they came up with themselves, and never forget to log a NO GO honestly.
The causal direction has to be stated precisely: it isn't "because we have AI, the process got automated" — it's "because the process was disciplined first, AI had something it could safely execute." Hand a process that permits discretion to AI, and what you get is discretion at scale — which is far more dangerous than human discretion, because it runs fast and looks objective. Discipline is the precondition for AI being usable, not something that comes after automation.
This principle generalizes into a test you can apply to any step, to decide whether it can go to a machine: can this step's correctness be verified without asking the executor's intent? If yes — hand it to agents, and build a pipeline for the verification itself. If no — keep it with a human, or redesign the step until the answer is yes.
This test is really the same distinction I used when writing about the Four Realms of Neural Networks: rules that can be explicitly verified live in the lowest realm — functions, inputs and outputs — no understanding required, only execution. But a question like "is the reason this signal works still going to hold tomorrow" lives in the curvature of a market manifold evolving through time — there's no shortcut, only a human weighing market structure, incentive shifts, and institutional constraints together. Discipline's job is to take the low-realm work off a human's hands, and purify their attention down to only the high-realm work.
What a Day Looks Like
With the abstract principle out of the way, here's the concrete shape. Over the course of a trading day, this "one person + AI" system runs roughly like this:
During market hours, no one watches the screen. The strategy computes a target portfolio at fixed times, the execution layer rebalances, and an independent risk engine checks every single order — this layer is purely machine territory, and a human was never supposed to be in it. Keeping a human out of this loop isn't just cutting a headcount — it eliminates an entire class of error: historically, human intervention during market hours has caused far more losses than it's prevented.
After close, agents take over operations. Data-quality checks, portfolio attribution, risk-metric sweeps, and a daily report sent to the fund's stakeholders — all run by agents off a checklist. The reporting step deserves its own mention: the narrative portion is written by a language model, but it's constrained by a whitelist mechanism — it can only cite real numbers produced by the deterministic computation layer; a fabricated figure fails validation and never reaches the final report. This is the correct posture for AI touching a formal external document: generation is responsible for expression, validation is responsible for fact, and the two layers never merge.
Overnight, the factory runs. Batch signal-screening jobs, scheduled model retraining and drift checks, next day's data prep — all completed while a human is asleep. Model training's role in this shape is deliberately unglamorous: it's one scheduled step on the pipeline, with its own acceptance criteria and rollback path, not a ritual that needs a dedicated attendant. The training decisions that genuinely need a human — changing the feature set, swapping a model family — happen far less often than outsiders assume, rarely enough to fit entirely inside a human's decision budget.
When something's wrong, the system wakes the human — not the other way around. All monitoring is designed as alert-on-anomaly, not status-on-display: on a normal day, the volume of information reaching the human approaches zero. This is deliberate — attention is the one genuinely scarce resource in this organizational shape, and any information flow that consumes attention without requiring a decision is a design mistake.
So what's left in a human's day? Read the verdicts, make the admission calls, set the direction, handle the real anomalies. All judgment, no execution.
What Stays With the Human: Three Non-Transferable Decisions
"Humans only do judgment" is too vague. Specifically, it's three things:
Admission. For every candidate signal the factory screens through, the last gate is always a human: does it get into the live portfolio or not. This isn't a ceremonial rubber stamp — the statistical verdict answers "did this hold up historically," and the human answers a question statistics can't: "is the reason it worked still going to be true going forward?" Crowding, the sustainability of the data source, the structural relationship to the existing book — these judgments draw on world knowledge outside the model. AI expands the funnel without limit; the human guards the last inch — neither side of this division of labor is replaceable.
Risk posture. Leverage level, drawdown tolerance, circuit-breaker thresholds — these parameters don't encode a statistic, they encode the capital owner's utility function. A machine can execute any risk posture, but choosing the posture itself is part of the fiduciary relationship — delegating it away is delegating away the fiduciary duty.
Direction. Which new data surfaces to expand into, which research line to shut down, how the system's own architecture should evolve — these decisions have feedback loops measured in months, no usable training signal, and are purely strategic judgment calls.
Flip it around: for anything that doesn't fall into these three categories, the default question is — why does this still need a person?
The Compounding Property of This Shape
Last point, and the most underrated one: this organizational shape's improvement compounds, where a headcount organization's improvement is linear.
For a headcount fund, one more unit of capacity means hiring one more person: hiring cycles, ramp-up time, and communication cost all grow super-linearly with team size — and people leave, taking process knowledge out the door with them. The organization's knowledge lives in people's heads; the organization itself doesn't actually learn.
In this shape, every process improvement settles into a permanent property of the system. A new validation rule becomes a gate every future screening round has to clear; the experience of handling one anomaly hardens into a new entry on the monitoring checklist; a falsified research direction gets recorded in an append-only ledger, never paid for twice. The system is a little smarter every day than the day before, and it never forgets.
Better still, the executor itself sits on the same compounding curve. The same frozen process that needed deep human involvement in a given step last year can, this year, be run independently by agents — the process hasn't changed, the executor got stronger. Betting on this shape means anchoring organizational capacity to the AI capability curve, rather than to the labor market. The gap between the slopes of those two curves over the past couple of years has been plain to see.
To be honest about the boundary: this shape currently holds for a specific scope — systematic, low-to-medium-frequency strategy families where rules can be made explicit. The trade-off is real: all tacit knowledge that's meant to enter the pipeline has to be forced into explicit form, and every "feel" that's meant to be executed by a machine has to be translated into a rule — an idea that can't be made explicit simply doesn't exist inside this system. We consider this constraint a feature, not a bug — an edge you can't write down is, in all likelihood, not actually an edge — but it does draw a real boundary around where the shape applies.
Closing
The endgame of competition in the quant industry was never "whose signal is better" — signals decay, signals get copied. What's actually being contested is the research cost curve: for the same research dollar, who can validate more candidates, retire the bad ones faster, and bank more negative knowledge.
A headcount organization's slope on that curve is fixed — set by the marginal cost of a human researcher. The "one person + AI" shape hands that slope over to the AI capability curve — and every discipline frozen into the system, every page added to the ledger, buys more with the next unit of research budget.
The organizational shape itself becomes a source of alpha.
This is the fifth piece in the Dnalyaw quant series. Earlier: The Geometry of Alpha, Two Languages of Risk Control, Dnalyaw: Engineering an AI Quant Trading System From Scratch.