Investing Before AI and After AI: How the Research Day Actually Changes

Before AI, an analyst’s working day was mostly manual reading and data entry: scrolling through hundred-page reports, using Ctrl+F to find one number, typing financial statements into a spreadsheet by hand, and chasing a single figure across several filings before any real thinking began. After AI, that grunt work gets delegated to an AI associate that reads the documents and does the first-pass extraction, so the human reaches the judgement sooner and spends the day on what only a human can do: deciding what the numbers mean and standing behind the call.

This is not a piece about whether AI replaces analysts. It does not, and we argue that case separately in will AI replace financial analysts. This is about something narrower and more useful: what the actual working day looks like, task by task, on either side of the shift. If you want the tool-by-tool playbook, AI for equity research covers it. Here we walk through the day itself.

The old day: reading everything to find the one line

Start with the classic task. A company reports, and you need to understand what changed. Before AI, that meant opening the annual report or the quarterly filing and reading it, often all of it, because you could not know in advance which paragraph held the line that mattered. A hundred pages of accounting policy and boilerplate to find the one sentence about a segment reclassification or a change in revenue recognition.

The tool for this was Ctrl+F, and Ctrl+F is dumber than it looks. It matches strings, not meaning. Search “margin” in a long document and you get forty hits, most of them irrelevant, and you still miss the paragraph that describes the same idea in different words. Search a transcript for “guidance” and the useful comment sits under “we expect”, “we are targeting”, or “directionally”. Plain text search breaks exactly where the language gets interesting, which is why earnings transcripts break search the moment you need nuance rather than a keyword.

Then came the spreadsheet. You would spread the financials by hand, typing revenue, costs, and balance sheet lines from a PDF into your model, quarter by quarter, year by year. Retyping numbers from a document is slow, and it is where errors creep in: a transposed digit, a wrong sign, a line read off the prior-year column. Half a day could disappear into data entry before you formed a single view.

And the numbers rarely sat in one place. To check whether a figure was consistent, you would open three filings and reconcile them: the press release, the detailed results, and the notes. One number, chased across documents, because a company can present the same item differently in different places, and you had to be sure which version to trust. If you have ever built a three-statement model from raw filings, you know the tax this takes before any analysis begins.

Here is the honest summary of the old day: most of it was mechanical. The reading, the typing, the cross-checking. The judgement, the part you are actually paid for, often started in the late afternoon, when you were already tired.

The new day: asking the documents questions

After AI, the shape of the day inverts. The mechanical work moves to an AI associate, and the human starts on the questions sooner. Go task by task and the contrast is concrete.

Reading becomes asking. Instead of scrolling a hundred pages to find one line, you ask: what changed in the segment disclosure this quarter, and where does the filing say it? A capable system reads the document and points you to the passage. You still read that passage yourself, but you read the right one, and you read it in the first ten minutes rather than the third hour. The reading does not vanish; the searching for what to read does.

Extraction becomes a first draft. Rather than retyping statements into a spreadsheet, you get a first-pass table pulled from the document, which you then check against the source. This is the single biggest time saver, and also the single biggest trap, because a model can produce a wrong number with complete confidence. The rule does not change from the old world: verify every figure against the filing before it enters a model or a note. What changes is that you spend your time verifying rather than transcribing, and verifying is faster than typing from scratch. Whether a machine can be trusted to read a balance sheet at all is a fair question, and the answer is: as a first pass you check, not as a final word you trust.

Chasing one number becomes cross-referencing on demand. The tedious reconciliation across filings, the same figure in three documents, is exactly the kind of high-volume, low-judgement work that delegates well. You ask for the number and its sources, then you adjudicate the differences. The judgement of which version is right stays with you. The labour of gathering the versions does not.

A quick side-by-side of the same day:

Task	Before AI	After AI
Find what changed	Read the whole filing, Ctrl+F	Ask, and get pointed to the passage
Spread the statements	Type from the PDF by hand	Check a first-pass extraction
Reconcile one figure	Open three documents, compare	Request the number with its sources
First view formed	Late afternoon	Late morning

Notice what the right column is not. It is not “the AI decides”. Every row still ends with a human doing the deciding. The tool moved the start line closer to the finish; it did not run the race.

What does not change, and must not

The parts of the day that AI leaves alone are the parts that were always the point.

Judgement stays human. Consider Reliance for the quarter ended September 2025. Its Oil to Chemicals segment did roughly ₹1,60,600 crore of revenue at about a 9% segment margin, while Jio did about ₹42,700 crore at roughly a 52% margin, and the upstream Oil and Gas segment did about ₹6,100 crore at around an 83% margin. An AI can pull those numbers in seconds and lay them in a table. What it cannot do for you is the reading of them: that the largest revenue segment is one of the thinnest by margin, that profit and revenue live in different places inside one company, and that a single consolidated topline hides all of it. That interpretation, and what you do with it, is the human’s job. The extraction is the easy part. Comparing businesses well is still a skill you own, not a query you run.

Conviction stays human. A model can summarise Asian Paints telling the market it expects “high single-digit volume growth in the band of about 8-10%” and an EBITDA margin band of 18-20% for the year ahead. It can even flag when next quarter’s language drifts from confident to cautious. But deciding how much to believe the guidance, and how that shifts your view, is conviction, and conviction is not a summary. Reading management guidance as a promise you grade over time is a discipline, not an output.

Accountability stays human. When a call is wrong, a person is answerable, not a tool. This is why the analyst does not disappear even when every mechanical task is delegated. Someone has to own the view, and ownership cannot be handed to a model.

There is a quieter trap worth naming. AI makes it trivially easy to backtest an idea across years of “as reported today” data. But a company that files a new quarter usually rewrites last year’s comparable for a demerger, a discontinued operation, or a re-cut segment, so today’s tidy series is not the one that existed when an old decision was made. Run the test on those cleaned-up numbers and a mediocre idea can look convincing. Speed makes the slip easier to commit, which is why lookahead bias matters more after AI, not less.

The real gain is where the hours go

The point of delegating the reading is not the summary. It is the extra hours you spend thinking instead of typing.

Add it up and the shift is not “AI does the analysis”. It is that the low-judgement volume work, the reading, the extraction, the cross-referencing, moves off the human’s desk, and the human arrives at the interesting questions with more of the day left to spend on them. The best desks are already rebuilding their workflow around this split, and over time it starts to look less like a faster tool and more like an operating system for the whole team. What that reorganised day looks like end to end is its own subject, covered in the institutional research workflow.

The mechanics change a great deal. Judgement, conviction, and accountability do not change at all. An analyst who understood the craft before AI is more valuable after it, because the machine handles the volume and hands the person more time for the one thing it cannot do: decide what it all means, and stand behind the answer.