Investing Then vs Now: What AI Actually Changed
AI did not remove the need to understand a business. It changed the unit of research from one company at a time to one sourced question across the market.
For most of the last two decades, investing research was a routine of gathering: filter on a screener, download the reports, read the transcripts, copy the numbers into a spreadsheet, and repeat it every quarter. The process had to change because the edge had quietly become surviving that routine rather than understanding the business, and the first thing AI changes in investing is not the answers, it is the workflow.
The routine that ate the day
Picture how it actually went. You opened a screener to filter for candidates. You downloaded the annual reports. You read the earnings-call transcripts. You clicked through the investor presentation, checked a couple of news sites, searched a name on Twitter, opened half a dozen browser tabs, copied figures into Excel, and eventually convinced yourself you had “done enough research.”
The process worked. It was also painfully inefficient, and most of the inefficiency was invisible because it felt like work.
The problem was never too little information
Investors today have more data than ever. Every listed company publishes annual reports, quarterly results, exchange filings, investor presentations, conference-call transcripts, credit-rating reports, shareholding patterns, and a stream of interviews. The scarce thing was never the information. It was making sense of it.
So research slowly became an exercise in information management rather than investment thinking. Hours disappeared into finding the right document, comparing this quarter to the last, verifying a number, and trying to remember what management had said two years ago. Even seasoned investors built private systems to cope: folders full of PDFs, notes apps, Excel trackers, bookmarked threads, handwritten margins. The edge was not only understanding businesses. It was surviving the process.
Then came the chatbot
When large language models arrived, it felt like the problem had finally been solved. Instead of reading hundreds of pages you could just ask, “summarise the last annual report,” or “what are the biggest risks here,” and get an answer back in seconds. Research became conversational, and that was a genuine leap.
Then the initial excitement wore off and a different set of problems showed up. A general-purpose model does not know which disclosure is trustworthy. It mixes old and new filings, struggles with financial context, cannot reliably separate management guidance from an analyst’s opinion, and does not grasp that a single sentence buried in a call can matter more than ten pages of the annual report. Most importantly, it rarely shows why it reached a conclusion. In investing, a confident answer you cannot trace is not a convenience. It is a liability.
The fix is not a smarter chatbot. It is retrieval grounded in the actual documents: every claim pinned to the filing, the page, and the date it came from; a summary you can audit line by line; the operating metrics management mentioned once on the call pulled out as structured, comparable numbers rather than left in prose.
Guidance has moved up through the year. The annual report raised full-year capex toward the top of the guided range, and the most recent call framed it as front-loaded spend in the first half:
The real shift is the workflow, not the summary
Here is the part most people miss. The biggest opportunity is not replacing PDFs with chat. It is replacing manual research workflows, the multi-step jobs that used to take a person days and that were almost impossible to automate before.
Take the deepest version first: you can hand over a whole company, not just a question. Ask for a primer and the work runs as a sequence: it reads the filings and the recent calls, pulls the reported numbers, maps how the revenue is built, checks the trend against the forecast, and flags anything that looks off, returning one note where every figure opens to its source. The catch, and it is the important one, is that a fluent multi-step answer can be multi-step wrong, so the sourcing matters more as the tool does more.
What drives it. Three segments; the energy business is still the largest share of revenue, the consumer arm the fastest-growing.
How the year went. Growth ran ahead of guidance for three quarters, then softened in the most recent one as retail margins compressed3.
It stops being sequential
There is a deeper change here, and it is the one that actually matters. Old research scaled linearly with analyst hours. You studied Company A, then Company B, then Company C, one at a time, and the only way to cover more of them was more people or more late nights. Coverage was a function of time.
AI breaks that relationship. The unit of work stops being a company and becomes a question, run across the whole market in a single pass. You do not walk down the list. You point one workflow at five hundred companies and read the comparison that comes back.
Think about the questions you always wanted answered but never had the time to:
- Which companies have quietly reduced their growth guidance over the last three quarters?
- Which management teams’ capital allocation has consistently matched what they promised?
- Compare every banking call this quarter for what was said about deposit growth.
None of these is a search problem, and none of them scales with your hours. Each is one workflow that reads the same passage in hundreds of companies, compares it against the quarter before, scores what it finds, and hands back a ranked comparison instead of a pile of notes.
One line, answered across the whole market instead of one PDF at a time. That is a research workflow, not a search.
What a market-wide question looks like
The change becomes clearer with a real example.
We scanned the latest consolidated FY26 filings for current Nifty 500 companies and asked one narrow question: how often did total-income growth and profit growth tell materially different stories?
| FY26 scan | Companies |
|---|---|
| Comparable total-income and PAT growth available | 401 |
| Gap of at least 20 percentage points between the two growth rates | 170 |
| Total income grew more than 10%, but PAT declined | 33 |
| Total income declined, but PAT grew more than 10% | 10 |
Source: Altys calculations from consolidated company filings for the year ended 31 March 2026. The universe is current Nifty 500 membership as of 9 August 2026. Growth rates compare FY26 with FY25 on the same basis. Figures are counts, not a score or recommendation.
The useful output is not “170 interesting stocks.” It is a research queue.
For one company, the gap may come from raw-material costs. For another, it may be interest expense, an exceptional item, operating leverage or a change in product mix. The cross-market scan finds where the two statements disagree. The analyst still has to explain why.
Doing that manually would mean opening hundreds of filings and maintaining a comparable spreadsheet. A workflow can complete the mechanical comparison first, leaving the human task where it belongs: investigating the exceptions.
Research should not stop when you buy
Traditional research has another hidden flaw: most of it stops the moment the investment is made. You spend weeks building conviction, then revisit the company only when results arrive, or worse, after the stock has already reacted.
In reality a thesis is a living document. Management commentary changes. Industry conditions evolve. Competitors ship products. Regulations shift. Capital-allocation decisions compound over years. Keeping track of all of that by hand is close to impossible once you own more than a handful of names, which is why most people simply do not.
This is the equation AI actually changes. Instead of restarting research from scratch each quarter, the system can watch continuously and surface the developments that genuinely matter, when your attention is needed, rather than leaving you to notice them late.
Management trimmed the full-year revenue outlook on this morning's call, the first cut in six quarters.1
The point is better information, not tips
Notice what none of this is. The most valuable AI in investing will not tell you which stock to buy. Professional investors do not want predictions. They want better information, and they want to spend less time collecting it and more time deciding what it means.
That is where the leverage is. The tool becomes an analyst that never tires of reading filings, comparing disclosures, tracking promises, monitoring a watchlist, or finding a pattern across a thousand companies. The investor still makes the call and still carries the risk. The tool simply makes sure nothing important is missed.
Research is becoming continuous
Research used to be episodic. You researched, you invested, you waited, and three months later you started the whole thing again. Tomorrow’s version looks different. Questions become conversational. Screening becomes a workflow. Monitoring becomes automatic. And you spend your time interpreting signals instead of hunting for them.
That is not just a nicer interface on the same job. It is a different way of investing, and the rest of this series is about doing it well.
Public sources
Related reading:
- If everyone has AI, who wins?
- When profit outruns sales: four checks
- How to use AI for stock research
This article is educational. Altys Labs is not a registered research analyst or investment adviser, and nothing here is investment advice or a recommendation to buy, sell or hold any security.
Frequently asked questions
How has AI actually changed the investment research process?
Less by answering questions and more by replacing the workflow. The old routine was gathering: filtering a screener, downloading reports, reading transcripts, copying numbers into a spreadsheet, and repeating it every quarter. The real shift is that whole research tasks, screening a question across the market, comparing every call, tracking a thesis over time, can now run continuously instead of by hand.
Can a general chatbot be trusted for stock research?
Not on its own. A general model does not know which disclosure is current, cannot reliably tell management guidance from analyst opinion, and rarely shows why it reached a conclusion. For investing, an answer you cannot trace is dangerous. The fix is retrieval grounded in the actual filings, with a citation on every claim, not a more confident chatbot.
Should research stop once you have bought the stock?
No, and that is one of the oldest flaws in the process. A thesis is a living document: guidance shifts, competitors move, regulation changes, and capital-allocation decisions compound over years. Tracking that by hand is impossible past a handful of names, which is exactly the work continuous monitoring is meant to carry.
Will AI tell you which stock to buy?
The useful kind does not. Professionals want better information, not predictions, so the value is in reading, comparing, and monitoring at a scale a person cannot, while the investor still makes the decision. This article is educational and is not investment advice or a recommendation.