AI-Assisted Screening — Casting a Wider Net, Not Picking the Winner

A screener is a database query. You specify conditions — market cap above a threshold, promoter holding above a level, five-year sales growth above a rate — and the tool walks the whole listed universe and returns exactly the rows that satisfy them. The output is reproducible. Run it again on the same data and you get the same set. If a company is missing, either it failed a condition or the data is wrong, and both are checkable.

A language model does none of that. It has no table to walk, no sort, no comparison operator. When you ask it for companies meeting numeric conditions, it produces text that statistically resembles an answer to that question — which means well-known names in the right sector with figures in the plausible range. It cannot tell you it did not compute anything, because from its own point of view nothing unusual happened.

The failure is dangerous specifically because the output format is identical. A real screener output and an invented one both arrive as a tidy list with a company name, a ratio and a two-line reason. There is no visual tell. The only defence is a rule you apply before you look at the output: if the question required sorting or filtering the universe by a number, the answer did not come from a model.

This site ships a real scanner for exactly that reason — numeric conditions against live NSE and BSE data, run deterministically. The division of labour is clean. The scanner narrows three thousand companies to a workable set using conditions you can defend. AI then reads the language around that set, or around a sector, and surfaces things a ratio cannot express.

A screener sees fields. Ratios, growth rates, holding patterns, price and volume. It is very good at that and it is the right place to start. But an enormous amount of what actually moves a business lives in prose that never becomes a field.

Consider a management team that spent four consecutive quarters saying they were "confident of maintaining margins" and this quarter says they are "watching the input cost situation closely". No ratio has moved yet. The reported numbers are identical to what a screener would have shown last quarter. Something material has nevertheless happened, and it happened entirely in language.

Or consider a cost complaint appearing in eleven of thirty concalls across an industry within the same fortnight — freight, a particular raw material, a wage revision. Individually each mention is unremarkable. As a cluster it is a sector-level observation that no single company’s financial fields would ever surface, because the pattern exists across documents rather than within any one of them.

Or the shape of a filing calendar. A company that files three related announcements in eight days — a board meeting notice, a fund-raising enabling resolution and a plant-related intimation — is telling you something about its own tempo that no ratio encodes. AI can read all three and describe what they collectively refer to. Reading them is the value; concluding anything from them is still your job.

The right mental model is a funnel with four stages, and AI occupies exactly one of them. Getting this arrangement right is most of the article.

Stage one is the universe: every listed company on NSE and BSE. Stage two is the numeric filter, run in a real scanner, which cuts that to a set small enough to read — perhaps thirty to eighty names, depending on how strict you were. Stage three is the qualitative pass, and this is the AI stage: you feed it the documents belonging to those names and ask it to read across them. Stage four is verification, which is you, opening the source and confirming every claim before it goes anywhere near a decision.

Notice what the AI stage does to the funnel shape. It does not narrow further in the way a filter does. It re-sorts — it tells you which of the thirty are worth reading first and why, and it occasionally widens by surfacing a name adjacent to your set that shares the language pattern. That is the "wider net" in the title. The net gets wider at stage three and then the verification stage does the actual narrowing.

The most common way this goes wrong is stage-skipping: running the AI pass on the whole universe rather than on a filtered set, which is the exact moment you stop supplying documents and start asking the model to recall. If you cannot paste the sources for the names you are asking about, the stage is being run on memory, and memory is where invented figures live.

Here is the honest version of what AI does at stage three. You have thirty concall transcripts. Reading all of them carefully is perhaps two full days of work, which is why most people read four and guess about the rest. A model can pass over all thirty in an afternoon and tell you, with quotes, which four contain the language you specified. You then read those four properly, and you read them knowing what you are looking for.

That is a real gain and it is worth being precise about what kind of gain it is. It is triage. It changes the order in which you read, and it means you are unlikely to leave a document unopened purely because it was twenty-eighth in the pile. It has not evaluated any company and it has not compared any two businesses on merit.

The prompt for this pass has the same architecture as every other extraction prompt in this module: pasted source, fixed output structure, mandatory quotes, and explicit permission to come back with nothing. That last part matters more here than anywhere else. Ask a model to find guidance softening in thirty calls and it will find guidance softening in thirty calls, because producing an empty result reads to it as failure. You have to say, in writing, that "no such language in this transcript" is the expected answer for most of them.

Run it one document at a time, in a fresh thread, and collect the outputs yourself. Batching all thirty into one enormous prompt sounds efficient and produces the worst of both worlds: the earlier transcripts get less attention as the context fills, and the framing from company one leaks into the reading of company nine.

Once you have thirty identically-structured outputs, a second kind of question opens up: what is common across them? This is where the qualitative layer stops being a faster way to read and starts producing something a numeric screen genuinely could not.

Paste your own collected outputs — not the raw transcripts, your extracted and structured blocks — and ask for recurrence. Which cost line was mentioned by the most companies. Which phrase appears in several unrelated managements’ guidance. Where two companies in the same sector describe the same demand environment in opposite terms, which is often the most interesting single observation available, because at most one of them can be right about a shared market.

The output of this pass is a set of observations with counts and citations back to your own blocks. It is descriptive. "Eleven of thirty calls mentioned freight cost; here are the eleven quotes" is a fact about a set of documents. It is not a forecast, it does not imply that any of those eleven companies is attractive, and translating it into a view is entirely your job — done with the documents open.

It is also worth running deliberately in the opposite direction. Ask what appears in only one of the thirty. Sector-wide language tells you about the environment; a single company saying something nobody else is saying tells you either that they see something others do not, or that they are describing their own situation differently from their peers. Both are worth a careful read of that one transcript.

Everything above produces one thing: a small set of names you are going to look at more carefully, each attached to a specific reason and a specific quote. That is the deliverable. It is deliberately unexciting.

Write the shortlist down in a form that keeps the reason attached to the name, because the reason is the part that decays. "Company X — guidance wording moved from confident to conditional between the Q1 and Q2 calls; quotes attached; source Q2 FY26 transcript, management opening remarks" is a research task. "Company X — looks interesting" is a memory you will misremember as a conclusion within a fortnight.

Then do the actual work. Read the full transcript rather than the extract. Check the corresponding numbers in the filed results, from the filing itself. Look at what the company said in the previous two calls, not just the one you screened. The AI pass got you to the right four documents faster; it did not read them for you, and nothing in it constitutes a view on whether any business should be owned.

And keep a note of what your numeric filter excluded before stage three ever ran. A screening process is defined as much by what it rejects as by what it surfaces, and the exclusions are where a filter’s hidden assumptions live. A debt condition that removes every company mid-way through a capex cycle is a decision about what kinds of businesses you will never see, made silently by a threshold you set once.

The failures here are not dramatic. Nothing errors out. The process keeps producing tidy shortlists while the ground under it has gone.

The first is the numeric slide. You start with the discipline intact, then one day you ask the model a small quantitative question — "roughly what is this company’s debt level?" — because opening the filing is inconvenient. It answers. The answer is plausible. From then on the line between computed and generated figures inside your own notes is gone, and you will not be able to reconstruct which is which.

The second is confirmation. If you tell the model which four names you already like before asking it to read thirty transcripts, you will get a reading that supports those four. Screening prompts must be blind to your existing preferences — no mention of what you own, what you are watching, or what you hope to find.

The third is the illusion of coverage. Thirty transcripts processed feels like thirty companies understood. It is not. It is thirty documents skimmed by a process that reports what you asked it to look for and is silent about everything else. The set of things your prompt does not ask about is enormous, and it is invisible in the output by construction.

The fourth is drift in the prompt. Edit the screening prompt halfway through a sector and the last fifteen outputs are no longer comparable with the first fifteen — but they still look identical, and the cross-document pass will happily find patterns in a set that was produced two different ways. Version and date the prompt, freeze it for the duration of a sector run, and note the version at the top of the output file.