INVESTSIGHT CAPITAL is a fintech and capital markets firm in Bengaluru, India. This article is part of our research library — for the official site see investsightcapital.com, or ask us a question.
For decades, institutional research has been bottlenecked by human bandwidth — how many filings an analyst can read, how many companies a team can cover, how quickly a thesis can be tested against new data. Artificial intelligence is removing that bottleneck, not by replacing the analyst, but by compressing the distance between question and evidence.
This piece describes what that looks like in practice — where the tools genuinely help, how a research desk reorganises around them, and the failure modes a careful firm designs against. Most of what is written about AI in finance is either a sales pitch or a warning. The useful ground sits between the two.
How is AI used in investment research?
In investment research, AI is used mostly for reading, sorting and flagging — not for deciding. Large language models ingest annual reports, earnings-call transcripts, exchange filings and news at a scale no analyst can match, and turn them into structured summaries, extracted figures and searchable notes. Machine-learning models screen thousands of listed companies against quantitative criteria, score the tone of management commentary, and flag statistical anomalies in reported numbers that merit a closer look. The output is a shorter, better-ordered list of questions for a human to pursue.
What AI does not do, in any serious research process, is originate the investment thesis or decide position size. It compresses the time between a question and the evidence needed to answer it. The analyst still has to ask the right question, judge whether the evidence is trustworthy, and decide what it means for the portfolio. That division — machine breadth, human judgement — is the working definition of AI-augmented research today.
Where AI is genuinely useful today
Document ingestion. An Indian listed company produces an annual report, quarterly results, investor presentations, concall transcripts, exchange disclosures and credit-rating rationales every year. Multiply that by a coverage universe of a few hundred names and the reading load is beyond any team. Language models can ingest the full set, extract segment data, capex guidance and related-party disclosures into a consistent structure, and let an analyst query across companies instead of reading them serially.
Transcript summarisation. Earnings calls are where management tone lives. A model can reduce a ninety-minute concall to the guidance given, the questions dodged, and the language that changed since last quarter. The summary is not the analysis. It is the map that tells the analyst where to spend the hour.
Screening at scale. Quantitative screens are not new. What changes is the ability to run them across the whole market continuously, combine fundamental, price and text-derived signals, and re-rank a universe daily rather than at quarter-end.
NLP sentiment and tone. Natural-language models score the sentiment of commentary, news flow and filings. The value is rarely the score itself but the change in it — a company whose disclosures have quietly become more hedged is worth a look before the numbers confirm why.
Anomaly detection in filings. Models trained on historical filings can flag when a ratio, a footnote or an auditor’s phrasing departs from a company’s own pattern or its peer group. This is coverage, not brilliance: the model finds the thing nobody had time to check.
How does the research process change?
The honest answer is that AI changes where analyst hours go, not whether analysts are needed. CFA Institute’s 2025 workflow survey found that, among analytical roles, generative AI was most used to help prepare research reports, cited by 27% of respondents, while an earlier 2024 survey found only 16% using it for industry and company analysis — the part of the job that depends on judgement (CFA Institute, January 2026). The same piece puts the shift plainly: with extraction and presentation handed to a model, “the analyst can focus on data interpretation rather than preparation.”
In practice a desk reorganises around three shifts. First, the top of the funnel widens: more companies get a first pass, because a first pass is now cheap. Second, the middle of the funnel gets more rigorous: time saved on reading is spent on the questions a model cannot answer — is this management team honest, is this moat real, what happens to the thesis if the cycle turns. Third, the review loop tightens: because every step is logged, it becomes possible to audit why a decision was made and whether the process, not just the outcome, was sound.
Our own research tooling follows this shape. The machine layer widens coverage and surfaces exceptions. The human layer owns the thesis, the sizing and the accountability.
What are the guardrails?
Every one of the tools above has a characteristic way of being wrong, and a research process is only as good as its defences against them.
Hallucination. A language model will produce a fluent, confident figure that does not exist in the source document. The defence is structural, not hopeful: every extracted number must link back to the page it came from, and nothing enters a model or a note without that link.
Look-ahead bias. A backtest that uses information which was not available at the time — restated financials, today’s index constituents applied to ten years ago — will look brilliant and be worthless. Point-in-time data is expensive and unglamorous, and it is non-negotiable.
Backtest overfitting. With enough parameters, any strategy can be tuned to fit the past perfectly. The more a result has been tweaked, the less it should be believed. Out-of-sample testing, walk-forward validation and a bias towards simple models are the standard defences.
Data-snooping. Run a thousand screens and a few will look extraordinary by chance. A signal that survives must have an economic reason to exist, not just a good p-value.
None of this is exotic. It is the same discipline that separated good quantitative research from bad before the current generation of tools arrived. What has changed is that the tools make it easier to skip.
Where it isn’t
AI does not understand conviction, does not carry accountability, and cannot substitute for the discipline required to sit through a drawdown without panic-selling. Markets are not purely statistical systems — they are shaped by policy, psychology, and events that have no historical precedent to train on.
A model trained on the last twenty years of Indian equity data has seen a handful of genuine regime changes. It has not seen the next one. When the environment shifts, a model’s confidence does not fall with its reliability, and that gap is widest precisely when the stakes are highest. We have written separately about what AI cannot do in investing; the short version is that everything downstream of the evidence — sizing, risk, whether to act at all — remains human work.
What does this mean for Indian investors?
Two things, one regulatory and one practical.
The regulatory point is that accountability is being written down. SEBI’s consultation paper on guidelines for responsible usage of AI/ML in Indian securities markets, published in June 2025, sets out a framework in which the regulated entity — not the vendor, and not the model — remains responsible for the tools it deploys, with expectations around governance, testing, disclosure and data protection (SEBI, June 2025). Whatever final form the rules take, the direction is clear: an adviser cannot outsource responsibility to software.
The practical point is that the useful question to ask any adviser or platform is not “do you use AI?” but “what is it allowed to decide, and who checks it?” A firm that can answer precisely — which steps are automated, which are reviewed, what happens when the model and the analyst disagree — has thought about the problem. A firm that answers with adjectives has not.
Our position
Technology should enhance human judgement, not replace it. That belief sits at the centre of everything we build in our AI Research Centre — from research automation to predictive analytics — because the investors who win the next decade will be the ones who combine machine-scale research with irreplaceably human discipline.
The practical expression of that belief is unglamorous: point-in-time data, sourced extraction, out-of-sample testing, and a written boundary between what the machine surfaces and what a person decides. The tools will keep improving. The boundary is the part that has to be designed.
