Filings and Concalls with AI — Turning 300 Pages into Ten Sourced Claims

Compressing a document means throwing away most of it. That is the point. The question is what survives, and there are two answers with completely different consequences.

A summary preserves the gist. It tells you the business grew, margins were under some pressure, and management sounds cautious about the second half. It is pleasant to read and worthless to act on, because every sentence in it is unattributed. You cannot check it, you cannot quote it, and in six weeks you will not remember whether you read it in the filing or in the summary of the filing.

An extraction preserves the claims plus their coordinates. Revenue from operations for the year, with its unit, its comparative and the page it appears on. Total borrowings, with the note number. The auditor’s opinion type, and whether an emphasis-of-matter paragraph exists. It is drier, considerably shorter, and every line can be walked back to the document in under a minute.

Only the second one is research. The distinction is not about effort or intelligence — it is entirely about whether the route back to the source survived the compression. Everything else in this article is machinery for making sure it does.

A quarterly release is not one document. It is usually four: the financial results filed with the exchange, an investor presentation, a press release, and a concall transcript published a few days later. They overlap, they contradict each other in emphasis, and they are written for different audiences by different teams.

The temptation is to paste all four and ask for a combined view. It is exactly the wrong move. Every additional document crowds the same limited working memory, and accuracy on any specific question falls as that space fills. Worse, when a figure appears in two of the four with slightly different framing — one on a standalone basis, one consolidated — a blended answer silently picks one and does not tell you which.

Run them separately, in order, each in its own thread. The financial results first, because they are the audited-basis numbers and everything else is commentary on them. The presentation second, noting where its framing differs from the results. The transcript last, because the questions asked are only meaningful once you know what the numbers were.

You do the combining. That sounds like the machine failing to do its job, but the combining is the part that constitutes understanding the business. Handing it over is how you end up with a view you cannot explain and did not really form.

The instinct with a new filing is to ask what is interesting about it. That question guarantees you will find whatever you were already inclined to look for, and miss whatever you were not thinking about. A fixed checklist removes the choice, and removing the choice is the entire benefit.

A fixed list does two things a bespoke question cannot. It covers your blind spots on the days you are not curious about them — you will check related-party transactions in a quarter where nothing looked wrong, which is the only quarter in which finding something there would surprise you. And it produces identical structure across companies and quarters, so two extractions can be compared mechanically rather than re-read and interpreted.

The list below is deliberately short. Seven items, each of which can move your understanding of a business materially, and each of which is stated somewhere in a standard Indian filing rather than requiring inference. It is not a valuation framework and it is not a view on the company. It is a floor: the things you refuse to not know before forming any opinion at all.

Add to it if your process needs more, but add permanently rather than per-company. The moment the list varies by how interested you are in a name, it has stopped being a checklist and gone back to being curiosity with extra steps.

The first pass is mechanical and should feel that way. You paste one section of the filing, you run the same prompt you always run, and you get back a table with the same seven rows it always has. Nothing about it is clever, and that is why it is reliable.

Where the filing is long, run the pass section by section: the statements, then the notes, then the auditor’s report, then the management discussion. Each pass fills whichever rows that section can support and honestly reports the rest as absent. Rows fill in across passes, and by the end you have the full table with a reference on every line.

Expect gaps and be suspicious when there are none. A single section of a filing genuinely cannot answer all seven items. If a first pass over the profit-and-loss statement alone returns a confident promoter-holding figure, the model has gone outside the source and everything else in that table is now questionable too.

Keep the raw output. Do not tidy it into prose yet. The table with its gaps and its references is more useful than any polished version of it, and the polishing is what you will do at the very end, once.

A claim without a reference is a memory. You believe it, you cannot say why, and you have no way to re-examine it when something changes. A claim with a page reference is a position: you can go back, read the original sentence in its context, and decide whether it still says what you thought it said.

The gap between those two shows up months later, not on the day. A quarter after you formed a view, the company reports something that seems to contradict it. With references you spend two minutes confirming what the earlier filing actually stated. Without them you are arguing with your own half-remembered impression, and impressions defend themselves rather than update.

The reference is also the only practical check on the model. You cannot audit the reasoning inside it. You can absolutely search a PDF for a quoted string, and a quote that is not there is an invented claim caught in seconds. That is the entire verification loop, and it exists only because you demanded a reference and a quote in the first place.

There is one place this matters more than anywhere else, and it is arithmetic. A model can compute correctly from numbers in front of it. It can also produce a ratio built from one number you pasted and one it recalled, and the answer will look exactly like the correct one. Requiring both inputs to be shown, with both references, is what separates those two cases.

A concall transcript has two halves and they are not the same kind of text. The prepared remarks are written in advance, reviewed before delivery, and designed to be safe. The analyst question-and-answer section is unscripted. That is the entire reason it is worth more.

A model summarising the whole transcript at once will flatten this. It produces a blend in which management’s prepared framing and their answer under direct questioning carry the same weight, which is precisely the distinction you were reading for. Run the two halves separately and the difference stays visible.

What you want from the Q&A is not sentiment. It is a straightforward record: what was asked, what was answered with specifics, and what was answered without them. A question about a margin decline that receives a paragraph containing no figure is a data point, and it is one that only exists if you compare the question against the answer rather than reading the answer alone.

The most useful signal across quarters is the question that keeps returning. When analysts ask about the same subject three quarters running and the answer stays qualitative each time, that pattern says something a single transcript cannot. It is also the kind of comparison that only works if you extracted all three quarters in the same format.

One extraction tells you the state of things. Two extractions in an identical format tell you the direction, and direction is what a single snapshot structurally cannot give you. This is why the fixed checklist matters more than any individual prompt in this article.

The comparison is mechanical because both sides share a structure. Row three against row three, with both references carried across. Where a figure exists in both, the change is arithmetic. Where an item was disclosed last quarter and is absent this quarter, that absence is itself a finding — and a plain-language summary of either quarter alone would never have surfaced it.

Language deserves the same treatment as numbers. A risk described in one quarter as a possibility and in the next as something being actively managed has changed, even if no figure moved. Quoting both versions side by side makes the shift visible; paraphrasing either one erases it.

Resist the pull to have the model explain the delta. It does not know why anything moved and will produce a plausible reason, which then lodges in your notes as though it were sourced. Report the movement, keep the references, and form the explanation yourself — or leave it open until a document actually answers it.

The deliverable from a full pass over a quarterly bundle is small. Roughly ten claims, each one sentence, each carrying its document, page or note reference, and the date you checked it. It fits on one screen. That is not a disappointing yield from 300 pages — it is what 300 pages of a mostly-boilerplate document honestly contains for someone who already knows the business.

Every claim should be something you could read aloud to a sceptical person, who could then open the filing and confirm it. If a line cannot survive that, it is either an inference of yours, in which case label it as one, or an unsourced model output, in which case delete it.

Keep your open questions in the same file. The items where the answer was "not stated in the source", the analyst question that went unanswered, the item you could not compute because only one input was disclosed. These are the most valuable lines in the document, because they are the ones next quarter’s reading is for.

Then close the threads. They have served their purpose and their contents are no longer trustworthy after the fact — you cannot tell weeks later which line in a chat was verified and which was not. The notes file, where every surviving line has a reference and a date, is the whole point of the exercise.