The alpha is in the process.

Two skills, one job.

Ask anyone who has watched a great stock picker leave a platform to run their own money. A striking number flame out, and not because the ideas got worse. Portfolio management is a different skill from stock picking, and it is the one nobody teaches. Raj Shah, formerly a partner at Light Street and now running a two-person fund with agents doing the work of a much larger one, puts it plainly: the job is at least two parts, and the second part is where the careers end.

Joe O'Donnell ran Tiger Global's short book for most of a decade before founding Canary. His account of why years go wrong is the other half of the same point: no investor ever looked back on a bad year and wished they had read more earnings calls. The year went wrong because a few big decisions were made badly. Which is also why he steers away from productivity tools: summaries are lossy, and lossy in the direction of conviction. The least lossy thing is still to read the transcript.

Put those together and the research-agent pitch is aimed at the wrong half of the job. Reading more, faster, does not fix the decisions that lose the year. Those decisions are made after the idea exists, and they are made, at most fundamental firms, on feel.

Five decisions after the idea.

An idea has to survive five decisions before it is a position, and each one makes or loses money independently of whether the idea was right. Click one.

None of those five questions requires the machine to know whether the company makes semiconductors or sandwiches. That is the test a former Balyasny quantamental PM applies: anything that can be determined from data without knowing which company it is belongs on the quant side. All five pass. They also run on first-party facts, your book, your borrow, your fills, your carry, so they do not decay when the fund next door buys the same agent. And they are mostly deterministic, which means they belong in the control plane, governed and versioned, not in the model's improvisation.

His other observation matters more: fundamental investors are, on average, not good at sizing, and sizing is a pretty well solved problem in the quant world. Conviction, crowding and upside can be quantified into a position size. Most fundamental firms still do it with a feeling and a round number.

Same idea, different size.

Here is the uncomfortable experiment. Give the machine no opinion on the stock. Tell it the analyst's conviction, upside and downside, then give it the facts the firm already owns: crowding, factor exposure, liquidity, borrow, margin and the existing book. Ask one question: how big?

The idea has not changed. The expected return has not changed. Yet the right position might be 80 basis points, 1.6%, 4% or zero. That difference is the process, and at many fundamental firms it is still a round number. Move the inputs.

Three engines, one size.

Sizing is not a formula. It is a small constrained optimisation: maximise the expected contribution of the position, subject to the risk the book can take, the liquidity you can actually trade, the financing you can actually get, and the mandate's hard limits. The objective comes from the analyst. The constraints come from three engines that most firms run separately and never at construction time. Click one.

Risk tells you how much the book can take. Liquidity tells you how much the market can take. Portfolio construction tells you what has to move to make room. Most firms run those calculations separately, after the fact, if at all. The opportunity is to run them simultaneously, before the trade exists.

Run together, the binding constraint becomes explicit: "2.1%, not 4%, because it is borrow-limited" is a different conversation from "risk said no." The trade-off is priced, because liquidity cost and carry enter the objective. And the rebalance arrives with the position: the hedge or the trim that funds it inside the risk budget, which is what a PM does by hand at 6 pm, badly. The output is a staged proposal that states the size, the binding constraint, the cost of implementation and the delta from the round number. Never executed.

One calibration caveat. Conviction tiers are only worth what the analyst's own hit-rate history says they are worth. Until that history exists, the sizer is a consistency tool, and should be sold as one. After a year of logged proposals and overrides, it becomes a return tool, and the override log is the most valuable dataset the firm will have built.

Flip the funnel.

The industry has the funnel backwards. The fashionable architecture puts the machine at the bottom generating ideas and the human at the top controlling risk. Flip it. Keep the human where the data is weakest: the idea. Put the machine where the data is strongest: the process. Toggle.

Let the human decide whether management has changed, whether the industry structure is breaking, whether the market has misunderstood something. Let the machine continuously ask whether the position is too big, too correlated, too crowded, too expensive to finance, too difficult to exit, or simply being traded badly. Raj Shah's version is to put the AI at the top doing portfolio and risk management and preserve the alpha at the bottom that is still human; every quant he has met says no human has sizing alpha. Ying Hua's is that taking sizing off the brain frees it for the part that is truly art.

The paradox of better research agents.

The better research agents become, and the more widely they are deployed, the faster their public-information advantage should decay. The information does not become less valuable; the time during which it is exclusively valuable collapses. And something else happens: when models, positioning and expectations converge, the surprise matters more. The same efficiency that compresses the pre-print opportunity amplifies the move when reality deviates from consensus, and practitioners who have watched a decade of prints describe exactly that. Process alpha rises as idea alpha falls. Drag the slider.

Jan Szilagyi of Reflexivity makes the case that alpha only compresses if everyone asks the model the same generic question; the relationships in markets are many, complicated and unstable, which is hard on a machine trying to find rules it can reuse.

The dataset nobody else has

Every fund is asking what external data it should feed its agent. Start with the dataset nobody else can buy: your own decisions. Every entry, every exit, every resize, every override, every locate, every fill. Every time a PM doubled down after a drawdown or sold a winner too early. Szilagyi's example is the one that lands: point the agent at your trade logs and it tells you that you are great at buying and terrible at selling, or that your Friday trades lose money, and nobody on the desk was ever going to say so.

A generic model has no edge reading the same earnings call as everyone else. It may have enormous value finding a repeated mistake in ten years of decisions that only your firm has made. Your mistakes are proprietary data. Fixing them is proprietary return.

The claims, tested.

Click a claim as it is typically pitched.

The memo is not the product.

A research agent that ends by producing a better memo is stopping before any of the five decisions. The workflow runs the length of the chain below. Click a step; the amber ones are where process alpha is made or lost.

If an agent identifies a material change at 9:02, produces a beautiful memo at 9:05, and the trade is discussed at 10:30, the firm has automated research and left the whole of the process untouched. The agent should not terminate at a document. It should terminate at a decision-ready proposal that is already sized to conviction and crowding, priced against real liquidity, financed at a known carry, checked against the book's factor exposure and limits, with the evidence for and against attached and the approver named.

That is not replacing an analyst's keyboard. It is compressing the distance between fact and action while making the five decisions on numbers instead of feel.

There is a consequence for the sell side too. If research is going to be consumed by this process, a PDF is the wrong object. The machine needs structured, permissioned facts with identifiers and timestamps: conviction, expected return, downside, catalysts, liquidity views, positioning and financing colour. The research provider that can show not merely that its note was read but that one of its facts changed the size of a position will finally have attribution worth paying for. The fuller note to the sell side is an addendum to Facts and hands.

Measure the process, not the machine.

Most AI metrics are activity metrics pretending to be investment metrics: memos generated, tokens consumed, prompts per analyst, hours saved by survey. None answers the allocator's question. The research metrics below are leading indicators. The process metrics are where the money is, and most firms have never computed them.

What to tell the allocator.

1. We use agents for coverage, synthesis and disconfirmation.Every approved source, every name, every day, with citations and a complete audit trail.
2. We do not claim the model generates ideas.Named people remain accountable for the investment thesis. The agent compresses the path between new information and a portfolio-aware proposal.
3. We claim it improves the process.Sizing is computed from conviction, crowding and upside. Liquidity, borrow and carry are checked before the proposal exists. Factor and concentration are measured before the trade, not discovered afterwards. Our own execution is audited by the machine.
4. We measure both.Citation accuracy, hypothesis survival and decision latency on the research side; realised size against conviction, slippage, crowding-adjusted return and behavioural drag on the process side; attribution over a long enough period, with the sample size.
5. The agent operates on our context, as us.It reasons against positions, exposures, financing and flows a generic agent does not have, and it does so under the permissions of the person who asked, not a service account.
6. Execution remains controlled.The model reasons and proposes. Deterministic systems enforce limits, permissions, approvals and execution policy, and record evidence, decision and outcome.

Where TSIQ sits.

TS Imagine already sits on the five decisions: positions, risk, liquidity, pricing and financing, across trading, risk, wealth, portfolio construction and financing, through one semantic layer that means the same thing by "position" in every one of them. TSIQ runs the process on top of that, and runs the three engines together: risk sets the budget and names what binds, the trading data sets the liquidity cap and prices the implementation, the portfolio analytics propose the offsets. When the research agent finds something, it does not write a memo. It asks: do we own it, what else in the book expresses the same risk, what happens to factor exposure if we add it, can we borrow it and at what, how many days to build, what limits apply, what contradicts the thesis, who approves. It runs as the person who asked, with that person's entitlements, and it stages; it never executes until the firm's own gate says otherwise.

The reasoning plane can be any model. Models will change. The control plane cannot: positions, risk, liquidity, pricing, financing, limits, permissions and approvals have to be deterministic, governed and owned by the institution.

The AI reasons. TS Imagine controls.

Keep the human on the thing the data cannot tell you: the idea. Put the machine on everything the data can tell you: the process. Because as everyone gets the same models, reads the same transcripts and generates the same summaries, the advantage moves somewhere else.

The idea gets you into the trade. The process determines what you make from it.