Cited AI Research Changes the Analyst’s First Draft, Not the Final Call

Source-linked AI agents can compress filing review, but the durable workflow is a citation check before an analyst trusts the answer.

Layered investment documents flowing into a cited evidence map and a human verification checkpoint

The useful change in AI investment research is not that a model can summarize a filing. It is that a connected agent can produce a first-pass answer with links back to the filing, the official macro series, and the underlying transaction record. That changes the analyst’s first draft from a blank page into a set of claims that can be checked.

It does not change who owns the conclusion. A November 9 Valyu demonstration describes six jobs run against SEC filings, earnings data, Form 4 transactions, FRED, and prediction-market data. Its examples are revealing: the system found several figures, but the analyst still had to distinguish GAAP EPS from adjusted EPS, separate gifts from insider sales, and inspect Form 4 footnotes. The practical lesson is a citation gate, not an autonomous research desk.

The workflow has moved one step earlier

The old workflow was: open a 10-Q, search for a number, copy it into notes, then read surrounding context. The new version can start with a structured question: “What changed in this company’s disclosure, what is the exact source passage, and what still needs human review?” The agent searches a bounded source set, extracts candidate claims, attaches URLs and dates, and labels ambiguity.

That is a meaningful change for earnings preparation and event monitoring. A research team can ask for new Item 5.02 filings across a watchlist, route possible leadership changes to an analyst, and retain the source passage in the handoff. It can also request the official FRED series ID and observation date rather than relying on a rounded number in a news story. The result is less time spent hunting and more time spent deciding whether the evidence changes the thesis.

But retrieval is not verification. In the Valyu example, a company’s GAAP diluted EPS was $2.46 while the feed reported an adjusted $2.22 and estimates differed by provider. A fluent answer that says “EPS beat estimates” without naming the basis is not a small formatting error; it can change the meaning of the research note. The model can find the conflict. An accountable analyst must resolve it or preserve the uncertainty.

A reproducible cited-answer protocol

For a small coverage list, the workflow can be implemented without granting an agent trading or portfolio permissions:

  1. Define the source boundary: SEC EDGAR, the company release, one market-data provider, and official macro series. Record the retrieval timestamp.
  2. Ask for a table with claim, value, source URL, filing or observation date, definition, and confidence. Require “not found” rather than an inferred value.
  3. Sample every material number against the primary document. Check units, GAAP versus non-GAAP labels, periods, restatements, and footnotes.
  4. Ask the agent to list disagreements between sources and to explain what evidence would resolve each one. Do not let it silently select the most convenient figure.
  5. Write the investment implication separately from the extraction. The analyst should state whether the evidence changes a monitored hypothesis, leaves it unchanged, or is insufficient.

This makes the AI output an auditable research artifact. It also makes failure visible: missing filing coverage, stale data, a broken link, an unsupported claim, or a definition mismatch can be measured instead of hidden in polished prose. The index-maintenance workflow and near-miss log are useful companion ideas because both treat the handoff—not the model’s confidence—as the control point.

A safe test: 60 minutes, no trades

Choose three companies or three public filings that you already follow. Establish a baseline by having a human analyst extract ten material facts from each document in the normal way, recording elapsed time, factual errors, missing citations, and unresolved ambiguities. Then give the same documents to a cited-answer workflow for 30 minutes. Measure the same four metrics, plus the percentage of claims whose links land on the exact supporting passage.

The output is a comparison sheet, not a signal or recommendation. Stop if the agent invents a source, cannot reproduce a cited number, or produces a material claim without a definition. Do not extend the test into paper trading: the question is whether evidence handling becomes more accurate and reproducible, not whether a fast summary predicts returns.

The comparison will likely be mixed. The agent may reduce search time while increasing review time for edge cases. A small sample can flatter it; easy filings, survivorship in the chosen companies, and a patient analyst can all bias the result. A vendor’s own demonstration is also an incentive-laden source, not independent evidence of investment performance. There is no causal basis for turning a lower research time into an alpha claim.

For discretionary investors, the skill to practice is asking for definitions and source passages. Systematic teams should log retrieval dates, missingness, and schema changes before using extracted fields. Builders should optimize for provenance, abstention, and correction loops rather than a more persuasive answer. The frontier is therefore narrower—and more useful—than “AI does research”: AI can change the first draft when every consequential claim remains easy for a human to challenge.

Sources: Valyu, AI Agents for Financial Research (November 9, 2026); SEC EDGAR; FRED.


阅读中文版本 →