The AI Research Analyst Is Here: What Actually Works and What Is Just Marketing

The AI research analyst that actually exists today is a fast, tireless reader, not an oracle. It can summarise a 300-page annual report, extract a decade of line items into a clean table, and flag a management change the morning it is disclosed. What it cannot do, despite what the louder demos imply, is tell you which stock to buy or guarantee that its guesses will beat the market.

The gap between those two descriptions is where most money gets wasted. Below is a plain buyer’s guide: what genuinely works in AI equity research right now, what is hype dressed up as capability, and the short list of things to demand before you sign anything.

What AI Actually Does Well Right Now

Strip away the pitch decks and the real progress is concrete and, frankly, unglamorous. These are the tasks where AI has moved from novelty to genuinely useful:

Summarising filings and transcripts. Annual reports, quarterly results, and earnings calls are long, repetitive, and mostly boilerplate. Modern language models compress them into a readable brief in seconds, and they are good at surfacing the one paragraph in the notes that actually matters. This alone can save an analyst hours a day.

Extracting structured data. Turning a scanned PDF or a messy XBRL filing into a clean, comparable table of revenue, margins, and cash flow is exactly the kind of pattern work machines are built for. Done properly, it removes a huge amount of manual keying and the transcription errors that come with it.

Monitoring for changes and events. Software does not get bored. It can watch a universe of companies continuously and flag a new filing, a rating change, an auditor resignation, or a shift in disclosed guidance the moment it appears, instead of you finding out three weeks later.

Drafting first-pass notes. A model can assemble a structured first draft: the numbers, the context, the obvious questions, so the analyst starts from a scaffold instead of a blank page. The draft is a starting point to be challenged, not a finished view.

Answering questions grounded in documents. Ask “what did management say about capex plans last quarter” and a well-built tool returns the passage, with the source attached. This is retrieval done well, and it is genuinely powerful when the answer is anchored to a real document.

Notice the common thread. Everything above is grounded in something that already exists: a filing, a transcript, a disclosed fact. AI is excellent at reading, organising, and retrieving. That is a real productivity gain, and it is available today.

The Hype to Be Sceptical Of

Now the marketing. A worrying share of “AI for investing” pitches quietly cross the line from reading documents to predicting the future, and that is where scepticism should kick in hard.

Magic stock picks. If a tool claims it will tell you what to buy, treat that as a red flag, not a feature. Reliable, on-demand stock selection is not a solved problem, and no model has solved it. In India this is also a regulatory line: firms that are not registered advisers should not be handing out recommendations, and you should not be relying on software that pretends to.

Guaranteed or “market-beating” outperformance. Any claim of guaranteed returns is, at best, a misunderstanding of how markets work and, at worst, a deliberate lie. Backtests can be curve-fitted. Cherry-picked wins prove nothing. Past performance is not a promise.

Black-box predictions with no sources. A number with no lineage is a rumour with better formatting. If a tool produces a forecast, a score, or a “signal” and cannot show you what it was built from, you have no way to check it, audit it, or trust it. Confidence is not accuracy.

A number you cannot trace back to a filing is not research. It is a rumour with better formatting.

The tell is almost always the same: the more a product leans on the promise of prediction and the less it shows you its sources, the more you are buying marketing. The tools worth paying for are confident about what they can show you and honest about what they cannot.

Works Today vs Just Marketing

Genuinely works today	Be sceptical of
Summarising filings and earnings calls	”Tells you what to buy”
Extracting clean, structured financial data	Guaranteed or market-beating returns
Continuous event and change monitoring	Black-box scores with no sources
Drafting first-pass, human-reviewed notes	”Set it and forget it” auto-investing
Q&A grounded in real documents	Predictions presented as certainty

The left column is about compressing work you already do. The right column is about outsourcing judgement you should never fully hand over. A good tool makes you faster at the left and does not pretend to do the right.

The Point in Time Problem Nobody Advertises

Here is a subtle failure that the glossy demos never mention, and it matters enormously for anyone testing historical performance.

When you ask an AI what a company’s numbers looked like in, say, mid-2021, you need the data as it was known then, not the version that was later restated. Financial statements get revised. Prior periods get reclassified. Companies split their stock. A tool that quietly serves today’s restated figures while pretending to show you the past is not lying on purpose, but it will make any historical analysis look far cleverer than it really was.

This is called point in time correctness, and it is one of the hardest things to get right in financial data. It is also one of the easiest to fake in a sales demo, because the flaw only shows up when you check the tool against what was actually knowable on a given date. Ask about it. The answer tells you a lot about how seriously a vendor takes the difference between a demo and a discipline.

Auditability Is the Whole Game

If there is one principle that separates a serious research tool from a toy, it is this: you should be able to trace any number back to the document it came from.

That means clicking on a revenue figure and landing on the exact line in the exact filing. It means a summary that footnotes its claims. It means a monitoring alert that links to the disclosure that triggered it. When you can audit the machine, you can trust it selectively, correct it when it is wrong, and defend your work to a client, a committee, or a regulator.

When you cannot audit it, you are simply taking a black box on faith, and faith is not a research process. The best AI tools are built so that a sceptical human can always check the work. The weakest ones are built so that you never have to look, which is a polite way of saying you are never allowed to.

Coverage matters here too. A tool trained and tuned on one market may be superb there and quietly useless in another. Indian filings, disclosure formats, accounting conventions, and corporate actions are not the same as US ones. A model that has not genuinely reckoned with local structure will produce fluent, confident, and wrong output. Fluency is not coverage.

These are the principles Altys is built around for Indian equities: everything grounded in the source document, correct as of the date it mattered, and traceable back to the filing. Not because it is a marketing line, but because research that cannot be checked is not research.

What to Demand From an AI Research Tool

Before you buy anything, make the vendor answer for each of these. If they cannot, walk.

• Cited sources. Every number and claim links to a specific filing, transcript, or document you can open. No citation, no trust.
• Point in time correctness. Historical data reflects what was knowable then, not what was restated later. Ask them to prove it.
• Auditability. You can trace any figure back to the exact line in the exact source. The box is glass, not black.
• Transparency about limits. The tool tells you what it does not know and where it is uncertain, instead of bluffing.
• Coverage that fits your market. For Indian equities, it must understand Indian filings, accounting, and corporate actions, not just approximate them.
• No promises of prediction or profit. A serious tool sells you speed and rigour, never guaranteed returns or “what to buy.”

The AI research analyst is real, and it is genuinely useful. It reads faster than you, forgets nothing, and never skips a footnote. Just remember what it is: a very good assistant that shows its work, not a machine that knows the future. Buy the first one. Ignore anyone selling the second.

Altys Labs is not a SEBI registered investment adviser. Nothing here is a recommendation to buy or sell any security, and no tool can guarantee investment returns.