AI and algorithms in fund management: hype and reality

Ask what AI is doing inside Indian mutual funds and you will get two answers, both wrong. The marketing answer is that machines now pick stocks better than humans. The cynical answer is that it is all a chatbot with a press release. The truth is more specific, and more useful.

Where it is actually used today

Overwhelmingly in operations, not in stock selection. In Indian asset management the deployed use cases are concentrated in:

• Customer service — chatbots, query routing, onboarding.
• Cybersecurity and surveillance — anomaly detection on transactions and access.
• Compliance and reporting — extracting data from filings, monitoring mandate breaches, generating regulatory disclosures.
• KYC and document processing — reading and verifying documents at scale.
• Customer segmentation and distribution analytics.

These are real, valuable and completely unglamorous. They lower the AMC's cost base, which over time can show up in your expense ratio — a more plausible route to benefiting you than any claim about better forecasts.

In investment decisions the use is narrower: screening large universes, processing unstructured text such as earnings-call transcripts and filings, and signal generation as an input to a human process. Quantitative and rules-based funds have existed in India for years and are a small but growing part of the market.

Why "AI picks stocks better" is harder than it sounds

Four reasons, and they are structural rather than temporary.

1. Markets are adaptive. A pattern that predicts returns attracts capital until it stops predicting returns. This is not true of the domains where machine learning has been transformative — a cat does not change its appearance because a model learned to recognise it. Financial data fights back.

2. The data is thin where it matters. There have been perhaps a dozen genuine market cycles in modern Indian history. For a model trying to learn how markets behave in regimes, that is a handful of examples — vastly less than the millions that other machine-learning successes were built on. Daily data does not solve this: ten thousand daily observations of one regime is still one regime.

3. Overfitting is nearly irresistible. With enough variables and enough compute, a model will find something that explains the past perfectly and predicts nothing. Backtests are cheap; out-of-sample performance is not. Any strategy presented with a spectacular backtest and a short live record deserves the sceptical reading.

4. Crowding. If many participants deploy similar models on similar data, they take similar positions — and unwind them simultaneously. Machine-driven strategies can amplify exactly the herd behaviour they are supposed to exploit.

What the regulator has done

SEBI has moved on this. The framework for responsible use of AI and machine learning by regulated entities is built around a familiar set of principles — ethics, accountability, transparency, auditability, data privacy and fairness — with concrete obligations attached:

• A board-approved AI governance framework at the intermediary, with internal controls, model documentation and algorithmic audits.
• Named senior-management responsibility for development, testing, deployment and monitoring.
• Sole liability rests with the regulated entity for any AI or ML tool it uses, whether built in-house or bought from a vendor. You cannot outsource the accountability along with the model.

That last principle is the one that matters to you. Whatever a fund house's model does, the AMC is answerable for it.

⚠️ This framework has developed through consultation papers and circulars and continues to evolve. Verify the current position before relying on specifics.

What it means for your portfolio

Be sceptical of "AI-powered" as a product claim. It describes a tool, not a strategy, and it is not a reason to buy anything. The questions that matter are unchanged: what does it hold, what does it cost, what did it do in a drawdown, and how does it compare within its peer group?

A quant fund is an active fund. Judge it exactly as you would any other — information ratio, rolling returns, drawdown, costs, and a track record spanning more than one regime. A systematic process has genuine advantages, chiefly consistency and freedom from the behavioural errors that this entire module is about. It has no advantage in forecasting.

Watch the live record, not the backtest. For a rules-based fund the only honest evidence is out-of-sample performance since launch. Three years of live data in one market regime proves nothing, however good the simulation was.

Note the cost argument, honestly. If a systematic process replaces expensive research staff, the fund should be cheaper. If it is priced like a conventional active fund, ask what you are paying for.

And notice where AI genuinely helps you rather than the fund. Better search, better explanation, better document processing, faster access to what a scheme actually holds. The investor-facing gains are in comprehension and administration — which is not nothing, and is more real than the alpha claims.

Pitfalls to avoid

• Buying "AI" as a strategy. It is infrastructure, not an edge.
• Trusting a backtest. They are optimised on the data they are shown.
• Assuming a systematic fund is lower risk. It removes emotional error and adds model risk.
• Ignoring costs because the process sounds sophisticated. Cost compounds against you regardless of how the decisions are made.
• Expecting AI to solve the behavioural problem. Nothing in a fund's process stops you redeeming at the bottom.
• Believing "the model has no emotions" means it cannot fail. It has the emotions of whoever chose its objective function and its training window.

Key takeaway

Inside Indian mutual funds, AI today does operations, compliance, surveillance and service — not stock picking, and that is where its real value is, because lower operating cost is a plausible route to lower fees. Genuine predictive advantage is structurally hard here: markets adapt, there are only a handful of real market cycles to learn from, and overfitting is nearly irresistible. SEBI now requires a board-approved governance framework and holds the regulated entity solely liable for any model it uses, in-house or bought. Treat "AI-powered" as a description of tooling rather than a reason to invest — judge a quant fund on its live record, its costs and its drawdown, exactly like any other active fund.