How an AI Agent Reads Filings to Find Opportunities

The opportunity is public. Almost nobody reads it.

Every quarter, public companies file hundreds of pages of disclosure. A 10-K carries the risk factors. An 8-K flags a material event within days. Proxy statements reveal compensation and governance fights. Insider Form 4s show exactly what executives bought or sold, and when. This is the most reliable, most regulated, most timestamped data in markets, and it is free. The catch is volume. No human reads every filing for every name they hold, cross-references it against the day's news, and does it again tomorrow.

That gap is the whole thesis of an investing agent. Not a secret dataset, and not a faster trading algorithm. A reader. Something that ingests unstructured text at a scale a person cannot, and turns it into a short list of things worth your attention, each one tied back to the sentence that triggered it. The frontier in personal finance right now is not a better questionnaire. It is a system that can actually read.

What a robo-advisor does, and where it stops

To see why agents are different, look at what came before. A classic robo-advisor works from a questionnaire. You answer a handful of questions about age, income, and risk tolerance, and the system slots you into one of roughly twenty pre-built baskets of ETFs, then rebalances on static rules. Fortune, writing in March 2026, put it bluntly: robo-advisors are now regarded as a generic, incremental feature at best, systems that slot customers into one of twenty or so ETF baskets based on a questionnaire.

Nothing in that pipeline reads a filing. It does not know that a company you own just disclosed a going-concern warning, or that an executive dumped stock ahead of guidance. It reacts to price and calendar, not to text and events. That is fine for a set-and-forget index portfolio. It is useless for finding the specific, per-position opportunities and risks that live inside disclosure. The robo-advisor optimizes an allocation. It never investigates a company.

Point-in-time data: the unglamorous foundation

Before an agent reads anything, the data has to be honest about time. This is the part that separates a credible system from a demo. Filings get amended. Companies restate earnings. A ticker gets reused after a delisting. If an agent reasons on today's fully-corrected version of the past, it will look brilliant in backtests and fail live, because it quietly used information that did not exist yet. This is look-ahead bias, and it is the most common way analysis lies to you.

Point-in-time data fixes this. It stores each fact as it was known on a given date, so an agent reasoning about March only ever sees what was public in March. For an agent whose entire value is reacting to filings and news the moment they land, timestamp discipline is not a nice-to-have. It is the difference between a signal and an artifact. When you evaluate any tool that claims to read the tape, the first question is whether it respects when each fact became true.

From unstructured text to a cited signal, step by step

Here is the mechanism, in plain terms. First, ingestion. The agent pulls the raw documents, a new 8-K, an earnings transcript, a wire story, and splits them into passages. Second, retrieval. When it reasons about a position you hold, it fetches the passages most relevant to that name and that question, rather than trying to hold the entire corpus in its head at once. This retrieval step is what keeps the model anchored to real source text instead of its own memory.

Third, extraction and reasoning. The language model reads the retrieved passages and pulls out the fact that matters: a lowered guidance range, a customer concentration risk, a debt covenant, an insider purchase cluster. Fourth, and this is the part that makes it trustworthy, attribution. Every claim the agent surfaces carries a pointer back to the exact filing and passage it came from. The output is not the vibe of a document. It is a specific line you can open and read yourself. A good agent produces a claim and its receipt in the same breath.

This is also where per-position analysis replaces basket-picking. Instead of telling you to hold twenty percent bonds, the agent can tell you that one specific holding just filed a risk factor that contradicts the reason you bought it. That is a level of granularity a rules-based robo-advisor structurally cannot reach.

The real risk is not rogue trading. It is hallucination.

Language models generate fluent text, and fluent text can be confidently wrong. A model asked about a company can invent a revenue figure, misattribute a quote, or blend two filings into a fact that exists in neither. In casual chat this is annoying. In investing it is disqualifying. The honest literature on autonomous finance names this plainly, alongside related failures like over-trading, unclear accountability, and regulatory exposure. Academic work on AI-driven advice stresses that without clear safeguards, digital platforms risk reproducing the very problems they promise to solve, and it centers auditability as a first principle rather than an afterthought.

The fix is architectural, not aspirational. Retrieval grounds the model in real passages. Citations make every claim checkable in one click. And the workflow assumes the model can be wrong, so a human sees the evidence before anything happens. A citation is not decoration. It is the mechanism that converts a plausible sentence into a verifiable one, and it is the single most important feature to demand from any agent that reads and reasons on your money.

Find, propose, route: why consent-first design answers the fear

Reading is half the system. The other half is what happens next, and this is where design decides whether autonomy is safe. The pattern that holds up is consent-first. The agent finds an opportunity across every account you connect, banks, brokerages, and crypto, so it reasons about your whole net worth rather than one silo. It explains each finding in plain English, with the citation attached. It proposes the move. Then it routes that move to your broker the moment you approve, where supported, and not one second before.

You keep your own accounts and your own broker. The system is non-custodial, it never moves money on its own, and you hold a kill switch. If you want more autonomy, you can hire agents that act on their own strictly inside limits you set: a maximum per trade, a drawdown ceiling, a defined universe, with the same kill switch in your hand. Tengu is built this way on purpose, because the answer to the accountability question is not to trust the model more. It is to make every step visible, cited, and reversible. One capability makes the case concretely: cross-account tax-loss harvesting only works when a single agent can see every account at once, and it only feels safe because you approve the harvest before it routes.

Compliance matters here too. A serious agent proposes and you approve. It does not silently execute trades at brokers where that is not supported, and it does not promise returns. It hands you a checkable case and a button.

Where this is heading, and what to demand

The direction of travel is clear. Deloitte projects that generative AI could become retail investors' leading source of investment advice around 2027, with usage climbing toward seventy-eight percent by 2028. Fortune notes that more than a third of consumers across every age group already consult tools like Claude and ChatGPT about their investments, often before they talk to a human advisor. The demand is here. The open question is quality.

So judge these systems on three things. Does it respect point-in-time data, or does it quietly cheat with hindsight? Does every claim carry a citation you can open, or are you asked to trust the tone? And does it put you in the loop with real limits and a kill switch, or does it ask for the keys? An agent that reads filings and news can genuinely find things you would miss. An agent that reads them, cites them, and waits for your approval is one you can actually put to work.