From Reading Documents to Asking Questions

Research is moving from reading whole documents front to back to asking documents specific questions and getting sourced answers back. The unit of work is shifting from the page you skim to the question you pose, and that changes where an analyst’s time and attention actually go.

That is the shift in one line. The rest of this piece is what it looks like in practice, what gets faster, what stays stubbornly human, and why the sourced answer, not the fast answer, is the part that matters.

How research used to run: read everything, hope to notice

For decades the default way to work through a filing was linear. Open the annual report, start near the front, and read toward the back. Do the same with the quarterly results, the notes to the accounts, the concall transcript, and the previous few years for context. You read all of it because you did not know in advance which sentence would matter, so the only safe move was to pass your eyes over every one.

This is thorough and it is honest work. It is also enormously expensive in the one resource a good analyst never has enough of, which is attention. A single annual report runs to a few hundred pages. A serious desk follows dozens of names. Reading everything front to back does not scale, so in practice people cut corners: they skim, they read the sections they always read, and they quietly hope the important line is somewhere they happened to look.

The failure mode is not laziness. It is arithmetic. When the job is to notice one changed clause buried in three hundred pages, across many companies, every quarter, human attention runs out long before the pages do. We wrote about the day-to-day version of this contrast in investing before and after AI; here the focus is narrower, on the act of reading itself.

The shift: pose a question, get a sourced passage

The new default inverts the order. Instead of reading the document to eventually arrive at a question, you start with the question and go straight to the answer.

You want to know whether the segment mix moved this quarter. You ask that. You want the exact words management used about margins, and how they differ from last quarter’s words. You ask that. You want every place a related-party transaction is mentioned in a report, or the sentence where the auditor qualified something, or the line in the notes that explains why cash and profit diverged. You ask, and you get the passage, with a link back to the page it came from.

The reading does not disappear. It concentrates. You still read the two or three passages that carry the judgement, and you read them slowly, because those are where the thinking lives. What goes away is the hour of turning pages to locate them. The tireless first pass, finding where the answer sits, becomes cheap. The expensive part, deciding what the answer means, stays with the person.

The old skill was reading everything and remembering where things were. The new skill is asking the right question and checking the answer.

This is a genuine change in craft, not just a faster tool. It rewards people who know which questions to ask, which is a different and more valuable talent than the patience to read linearly.

Why this is not just keyword search

It is tempting to file this under search. It is not, and the difference is the whole point. Keyword search, the Ctrl+F reflex, finds strings of characters. It cannot find meaning, it misses the same idea when a different filer words it differently, and it cannot actually answer anything. We take that apart in detail in the death of Ctrl+F in annual reports; the short version is that matching words is not the same as understanding a question.

Asking a question is different in kind. “Did working capital tie up more cash this year, and where does the report explain it?” is not a word you can search for. It is a question about the relationship between three numbers and the language around them. The same problem shows up sharply in earnings calls, where the words that matter are rarely the obvious ones, which is why earnings transcripts break ordinary search in the first place. A question-answer workflow is built to handle meaning, not matches, and that is the line between a search box and a research tool.

What changes about time and attention

The honest way to describe the gain is that it removes grunt work, not judgement. Here is roughly where an analyst’s hours go on a document, and what moves.

Task	Old workflow	Question-answer workflow
Locate the relevant passages	Read linearly, skim, hope	Ask, get the passage with its source
Read the passages that matter	Squeezed in after all the skimming	The main event, done slowly
Cross-check across quarters and years	Manual, tab by tab	Ask the same question of each period
Form a view	Fully human	Fully human

The first row is where the time is saved, and it is real time. The last row does not move at all, and it should not. Forming a view about a business is the job. Nothing about asking better questions relieves you of the responsibility for the answer. If anything the responsibility grows, because you can now cover more ground, which means more claims you are accountable for.

There is a subtler gain in attention. When locating information is cheap, you are more willing to check a hunch. The old cost of “go find where the report discusses this” was high enough that people often did not bother and went with memory. When the cost drops, you actually verify, and verification is where errors get caught.

The sourced answer is the whole point

A fast answer that you cannot trace is worse than useful, it is dangerous. In investing, a number or a claim without its source is a rumour you have not yet checked. The value of a question-answer workflow is not speed on its own. It is speed plus a link back to the exact filing, page, and date, so you can read the original and satisfy yourself it says what the answer claims.

This is Altys’s stance and it is not negotiable: an answer about a company must carry its source. Not because sourcing is a nicety, but because the entire reason to move to this workflow is to check more, not to trust more blindly. An answer you cannot audit gives you the confidence of having checked without the substance of it, which is the most expensive kind of mistake on a research desk.

A caution the new workflow does not remove

One old problem survives the shift, and it is worth naming. When a company reports a quarter, it often restates the prior-year comparable, for an accounting change, a demerger, a discontinued operation, or a redefined segment. So the history you can ask questions of today is not always the history that was knowable on the date a decision was made. A question answered against “the numbers as they read now” can quietly flatter a past call, because the past has been tidied up since. This is the shape of lookahead bias, and asking sharper questions does not fix it. It only matters more, because you can now ask a decade of history in seconds and must remember that the decade has been edited.

The good version of this workflow respects when each fact was known, not just what it says today. Asking a document a question is a large step forward. Asking it the right question, and holding the answer to the standard of a source and a date, is the part that turns a faster tool into better research.

The quiet reframing

Step back and the change is about what an analyst is for. When reading everything was the job, diligence looked like patience. When asking the right questions is the job, diligence looks like curiosity with a paper trail. The documents have not changed. What has changed is that you no longer have to walk every inch of them to find the few sentences that decide your view. You get to spend your attention on the sentences, which is where it belonged all along.