One point runs through the whole report: using AI does not change the rules. An AI-driven strategy is an algo like any other under SEBI's retail framework, an AI tool that recommends trades is giving advice that needs registration, and a promise of assured returns from an AI product is a warning sign whatever technology sits behind it. The report describes public rules and public statements; it is not a QuantPower study, and it makes no forecast of what AI will earn for anyone.
AI in Indian markets at a glance
| Item | Position on 20 October 2026 |
|---|---|
| Reporting of AI and ML systems to SEBI | Required of intermediaries, exchanges and mutual funds since 2019 |
| SEBI's guidelines on responsible AI use | Proposed in June 2025; final guidelines announced as coming shortly |
| AI strategies under the retail algo framework | Treated as algos; hidden-logic strategies need a registered research analyst |
| AI that recommends trades to clients | Investment advice or research, which needs SEBI registration |
| Personal data in AI systems | Covered by the Digital Personal Data Protection Act, 2023 |
| Main fraud risk | Deepfake videos and voice clips of market officials and well-known investors |
How this report was put together
The report uses public sources only: SEBI's circulars and consultation papers, statements by SEBI's chairman as reported in the press, the exchanges' investor cautions, and the text of the data protection law. Each is listed in the references. Where a rule is proposed but not final, the report says so. Facts were checked on 20 October 2026; rules in this area are changing, so check SEBI's website for the current position before relying on any of them.
A short history of AI in markets
The idea of computers making trading decisions is older than most traders think. In the 1980s, banks built "expert systems" that wrote traders' rules of thumb into software. In the 1990s, researchers and some funds trained neural networks on price data, with mixed results that taught a lasting lesson: a model that fits the past well can still fail on new data. Through the 2000s and 2010s, machine learning spread through finance in quieter roles: credit scoring, fraud detection, execution algorithms that split large orders, and the surveillance systems exchanges use to spot manipulation.
Language models are the newest wave. In 2017 researchers at Google published the transformer, the model design behind today's large language models.1 In late 2022 conversational chatbots built on such models reached the public, and within a year brokers, fund houses and analysts were using them to summarise filings, answer client questions and draft research. Traders began asking chatbots for stock picks, and fraudsters began using generated video and voice to impersonate market officials.
In India the timing coincided with the fastest growth in retail options trading the market had seen, which meant that AI tools reached a large audience of new traders at the moment when SEBI was already worried about losses in that segment. Regulators followed. SEBI had already required market intermediaries, exchanges and mutual funds to report the AI and machine learning systems they use, in circulars of 2019.2 In June 2025 it published a consultation paper on guidelines for the responsible use of AI and ML in securities markets.3 SEBI's framework for retail algorithmic trading, in force for every broker since 1 April 2026, applies to AI-driven strategies as to any other.4
Where AI is used in Indian markets
| Area | Typical uses | Who uses it |
|---|---|---|
| Surveillance | Spotting unusual trading patterns, front-running, pump-and-dump schemes and fake social media tips | Exchanges, SEBI |
| Customer service | Chatbots, ticket sorting, call summaries | Brokers, mutual funds |
| Onboarding and compliance | Document checks, fraud detection, monitoring client communication | Brokers, banks |
| Research | Summarising company filings, earnings calls and news | Brokers, analysts, funds |
| Trading | Signals from prices, order flow and text; execution timing | Institutions, prop desks, some retail algos |
| Advice and distribution | Robo-advice, portfolio suggestions | Registered advisers, fund platforms |
The rules that apply
SEBI's proposed AI and ML guidelines
SEBI's consultation paper of 20 June 2025 proposed guidelines for the responsible use of AI and ML in securities markets.3 Its main points:
| Principle | What it asks of market participants |
|---|---|
| Governance | Senior management with technical knowledge is responsible for AI models through their life cycle |
| Testing and monitoring | Models are tested before use and monitored for drift and errors, with fallback plans |
| Third parties | The regulated firm stays accountable when it uses an outside AI provider |
| Disclosure | Clients are told when AI is used in services that affect them, with its risks and limits |
| Data | Data is good quality, free of bias and handled under data protection law |
| Proportion | A lighter regime for internal uses that do not directly affect clients |
Comments closed on 11 July 2025. Speaking at FICCI's Capital Markets Conference on 19 August 2026, SEBI chairman Tuhin Kanta Pandey said the final guidelines would follow a tiered approach and require human oversight, data controls and kill switches.5 On 10 October 2026 he said they would be issued shortly.6
Why the proposals matter to individual traders
The proposed guidelines apply to regulated firms, not to individuals trading their own accounts. They still matter to traders, because they shape the tools traders use. A broker's chatbot that answers questions about your account, an adviser's AI-assisted research and a platform's AI assistant all come from firms that will be expected to test their models, monitor them for errors, keep a person accountable and tell you when AI is involved. When choosing a tool, a trader can ask whether the firm behind it can explain how its AI is tested and who answers for its mistakes. A firm that cannot answer that question is unlikely to meet the standard SEBI has proposed.
The retail algo framework
SEBI's framework for retail algorithmic trading, in force for all brokers since 1 April 2026, applies whether or not a strategy uses AI:4
- every algo order is tagged with an identifier from the exchange, the Algo ID,
- algo providers are empanelled with exchanges and act as agents of brokers, and
- black-box strategies, whose logic the user cannot see or replicate, need a provider registered with SEBI as a research analyst, and must be registered again with the exchange whenever their logic changes.
Most AI trading models are black-box by nature, because their decisions come from learned weights rather than rules a user can read. Our guide to SEBI's algo trading rules explains the framework.
Advice and data protection
An AI tool that recommends specific trades to clients is giving investment advice or research, which needs SEBI registration as an investment adviser or research analyst. Personal data used to train or run AI systems falls under the Digital Personal Data Protection Act, 2023, whose main duties apply from May 2027 under the rules notified in 2025.7
Three cases under the rules
The rules are easier to follow with examples. These are illustrations of how the framework applies, not legal advice.
- You write your own code with a chatbot's help. A trader asks a chatbot to draft a Python strategy, tests it, and runs it on their own account through their broker's API, from a static IP registered with the broker, at fewer than 10 orders a second. The chatbot only helped write the code; the strategy is the trader's own. It falls under the rules for a client's own algo, and the broker's checks apply to every order.
- A vendor sells AI signals you cannot inspect. A company offers a "self-learning AI bot" whose logic is not disclosed and connects it to clients' broker accounts. That is a black-box strategy. Under the framework it needs a provider registered with SEBI as a research analyst, empanelment with the exchange through a broker, and a fresh registration whenever its logic changes.
- You use an assistant to draft rules you can see. A trader describes a strategy to an AI assistant inside a trading platform, which drafts the legs, entries, stops and exits. The trader reads every rule, tests it and switches it on. The rules are visible and can be replicated, which is the white-box category; the trader made each decision.
The question that decides the category is not whether AI was used, but whether the trading logic is disclosed to the user and can be replicated by them.
How AI changes things for each market participant
| Participant | What AI changes | What the rules expect |
|---|---|---|
| Exchanges and SEBI | Faster detection of manipulation, unusual patterns and fake tips | Systems reported and tested; humans in charge of decisions |
| Brokers | Chatbots, onboarding checks, fraud detection, surveillance of client communication | Accountability for outside AI providers; disclosure to clients where AI affects them |
| Advisers and research analysts | Faster research, drafting and client service | The registered person stays responsible for every recommendation |
| Algo providers | Strategy design and signal generation | Empanelment; registration of black-box strategies |
| Individual traders | Help reading filings, drafting rules and code, testing ideas | The same rules as without AI; no registration needed to use AI for one's own trading |
The common thread is accountability: in every case, a person or a regulated firm, not the model, answers for the outcome.
The limits of AI price prediction
There are reasons to be sceptical of any AI product that claims to predict prices reliably:
- Prices already reflect public information. By the time a model reads a public announcement, many other traders and systems have read it too, and the price has usually moved.
- The signal is small and the noise is large. Day-to-day price changes are mostly noise. A model that finds a pattern in noise will find one that does not repeat.
- Markets change. A relationship that held in one period, such as how stocks react to a type of news, can weaken or reverse as more traders act on it, or as the rules of the market change.
- Costs decide the result. Many apparent edges are smaller than the brokerage, taxes and slippage of acting on them, especially for short-term signals that trade often.
- Testing many models inflates the best one. Training many models and keeping the best is the same trap as testing many strategy versions: the winner looks better than it is.
None of this means AI is useless to traders. It means AI is most valuable for tasks with a clear right answer, such as summarising a document or checking code, and least reliable for the one task most products advertise: telling you what the price will do.
What language models can do for traders
Large language models read and write text. For traders, that makes them useful for some tasks and unreliable for others.
| Task | How well it works | Watch for |
|---|---|---|
| Summarising filings, results and circulars | Well | Check numbers against the source document |
| Explaining concepts and rules | Well | Rules change; check the date of the source |
| Writing and checking strategy code | Useful as a draft | Test every line; models invent functions that do not exist |
| Turning a described rule into a strategy | Useful as a draft | Read every rule it produces before testing it |
| Classifying news as positive or negative | Fair | Results depend on the prompt, and on English versus Hindi or regional news |
| Predicting prices | Poorly | No reliable evidence of an edge after costs |
| Giving live prices | Not suitable | Models do not see live data unless connected to it |
Why language models make confident mistakes
A language model produces the most likely next words given what it has read. It does not look facts up unless it is connected to a source, and it has no built-in sense of whether a figure is right. That is why a model can state a wrong lot size, an old expiry day or an invented circular number in the same fluent tone as a correct one. The defence is to connect the model to real data where possible, to ask it to show where a figure came from, and to check anything that affects a trade against the source.
Sentiment from news and filings
Sentiment models turn text, such as news, company announcements and social media posts, into a score. A trader might test whether stocks with very positive news outperform the next day. Building a fair test is harder than it looks:
- Timestamps. The score must use only text published before the trade. A story time-stamped at 9:05 but updated at 11:00 creates look-ahead bias.
- Language. Much Indian market news is in Hindi and regional languages. A model trained on English text may misread it.
- Already priced in. Big news moves prices within seconds, before most retail systems can act.
- Costs. Short-term signals trade often, and costs can exceed the edge. Include them in every backtest.
- Overfitting. Many prompts and thresholds can be tried. Keep data aside for a final test, as our guide to avoiding overfitting explains.
A worked example: testing a news rule fairly
This is a hypothetical example of the process, not of a result.
A trader wants to test a rule: buy a Nifty 50 stock at the next day's open when a model scores that day's company announcement as strongly positive, and sell at that day's close.
- Collect the announcements with the exact time each was first published, not the time of the last update.
- Score each one with the model, using one fixed prompt written before looking at any results.
- Match each score to prices that came after the announcement: the next day's open and close.
- Deduct costs: brokerage, STT, exchange charges, stamp duty and GST, and an allowance for slippage at the open.
- Split the period. Choose the score threshold on the first two years only. Test it once on the third year.
- Compare with a baseline: buying the same stocks on random days, at the same times, with the same costs.
If the rule does no better than the baseline after costs on the held-back year, the model's scores are not adding anything the market had not already priced in. That is the most common result, and finding it out before trading costs nothing.
A worked example: reading a circular with a language model
This is a hypothetical example of good practice.
A trader wants to know what a new SEBI circular on derivatives changes for their option-selling strategy. They download the circular from SEBI's website and give the full text to a language model.
- Ask for a summary of the changes, with the paragraph number for each one. Paragraph numbers make every claim checkable.
- Ask for the effective dates, listed separately. Dates are where summaries most often go wrong.
- Ask what the circular does not change, to catch assumptions the trader or the model might make.
- Check each point against the paragraph it cites. In practice, most points will be right and one or two will be slightly off: a date for one phase given as the date for all, or a margin rule applied to the wrong contracts.
- Write the trader's own note of what changes for their strategy, from the checked summary, not from the model's text.
The model saved most of the reading time. The checking step is what made the result safe to act on, and it took minutes, not hours.
Using AI safely: a short checklist
- Give the model the source document rather than asking it from memory.
- Ask it to cite where each fact came from, and check the citations.
- Never paste your broker password, API secret or one-time codes into any AI tool.
- Treat every strategy an AI tool drafts as untested until you have backtested it on data it was not designed on, and run it in a forward test.
- Keep the final decision, and the switch that turns an algo on, with yourself.
Risks for retail traders
- Deepfake tips. Videos and voice clips that appear to show exchange officials, fund managers or well-known investors recommending stocks have been used in scams. NSE and BSE have both warned investors that their officials do not give stock tips.8
- "AI-powered" return claims. Promises of guaranteed or very high returns from an AI bot are a warning sign, whatever the technology.
- Black-box bots from unregistered providers. Under SEBI's framework, these should not be offered to retail traders through brokers' APIs.
- Confident wrong answers. Language models can state wrong figures fluently. Check anything that affects a trade.
- Herding. If many traders use similar models on similar data, their trades can crowd into the same positions, and leave through the same door at the same time when the models change their view.
- Sharing account access. An AI tool that asks for your broker password, API secret or one-time codes should not get them.
How to check an AI trading product
| Question | What a good answer looks like |
|---|---|
| Who is the company, and is it registered? | A named company; SEBI registration where it gives advice or research |
| How does it reach the market? | Through your own broker's official API, with an empanelled provider where the framework requires one |
| Can you see the rules it trades? | Yes, for white-box strategies; for black-box ones, a registered research analyst behind them |
| What does it claim about returns? | No promised returns; past results labelled as such |
| What does it need from you? | Never your broker password or one-time codes |
| Can it place orders without you? | Only through an algo you have switched on yourself, within your account's limits |
How QuantPower uses AI
QuantPower has two AI features. Both are software, not advisers.
- Quanty, the AI assistant in the QuantPower app, answers questions about the product, explains indicators and rules, builds a strategy from your description for you to check, backtests it, and runs a risk check that flags missing stops or sold options without a hedge. It saves algos switched off. It can read your own account, such as funds, positions and orders, and explain why an algo did or did not trade. It has no tool that places, changes or cancels an order, and it cannot switch an algo on. Every answer is marked as AI, and every figure must come from a tool result or your own words. Asked which strategy is profitable, it says that it is software, not a research analyst, and offers to help you make your own trading systematic and rule-based. It answers in the language you use, from a list of 27 entries.
- Connect AI links your own AI app, such as Claude or ChatGPT, to your QuantPower account through one MCP server, with the permissions you choose. It can read your account and market data, backtest your rules, save an algo switched off and prepare an order as a card in your app. It cannot place, change or cancel an order, switch an algo on, or confirm its own card. Only you confirm, with a tap and the second factor of your login. The rule is: AI prepares, you confirm.
The algos themselves are rule-based. Every rule of an algo is set by you and visible to you, whether you built it, adapted it from a template or had Quanty draft it.
What this means for traders
- Use AI to read, summarise and draft, and check what it produces against the source.
- Treat any AI trading signal as a strategy to test, with the same backtest, walk-forward testing and forward test on live prices as any other rule.
- Prefer strategies whose rules you can see and understand. Rules that are disclosed and can be replicated also fall in SEBI's white-box category.
- Verify who is behind any AI product: SEBI registration, broker partnership and exchange empanelment.
- Ignore tips from videos or messages, however real they look, and report suspected impersonation to the exchange or SEBI.
Our guide to machine learning in finance covers how models are trained and tested on market data.
Where AI rules in Indian markets are going
The direction is visible even before SEBI's final guidelines. Firms that use AI in client-facing services will be expected to have senior management accountable for their models, to test and monitor them, to keep humans able to override them and kill switches able to stop them, and to tell clients when AI is involved. Retail AI trading will continue to sit inside the algo framework, with hidden-logic strategies needing a registered research analyst. And fraud using generated video and voice is likely to keep growing, which puts more weight on investors checking the source of any tip. For traders, the practical result is that AI will be most useful where it helps them understand, build and test their own rules, and least trustworthy where it promises to decide for them.
Start with 500 free tokens
Connect your broker, build a strategy and test it on paper before a single live order goes out.
Frequently asked questions
Is AI trading legal in India?
Yes, within SEBI's rules. Automated orders through a broker must follow the retail algo framework, and a strategy whose logic is hidden needs a provider registered as a research analyst.
Can ChatGPT or another chatbot tell me what to buy?
General chatbots are not registered advisers and do not see live market data unless connected to it. Use them to understand concepts, not for trade decisions.
Has SEBI issued final rules on AI?
Not as of 20 October 2026. SEBI published proposals in June 2025, and on 10 October 2026 its chairman said final guidelines would be issued shortly. Check SEBI's website for the current position.
Does QuantPower use AI to trade for me?
No. QuantPower is trading software that runs rule-based strategies whose rules stay fully visible to you, such as those you build yourself or adapt from its templates. Quanty and Connect AI can help you draft and test rules, but neither can place an order or switch an algo on. QuantPower gives no tips or buy or sell advice.
Can I use a chatbot to write my trading code?
Yes, as a draft. Test every line, check that every function it uses really exists in your broker's API, and run the strategy in a forward test before it trades. If it runs on your own account through your broker's API, the rules for a client's own algo apply.
How can I tell if a video tip from a market official is fake?
Assume it is. Exchange officials and regulators do not give stock tips in videos or messages. Check the exchange's or SEBI's own website for any statement, and never act on a tip that asks you to join a group, move money or share account details.
Related guides
- Algo trading growth in India: how the algo share of turnover has changed
- What is algo trading?: how automated trading works in India
- Backtesting: testing a rule on past data before it trades
References
- A. Vaswani and others, "Attention Is All You Need", *Advances in Neural Information Processing Systems* 30 (2017)
- SEBI, circulars on the reporting of artificial intelligence and machine learning applications and systems by market intermediaries, market infrastructure institutions and mutual funds, 2019 (sebi.gov.in)
- SEBI, consultation paper on guidelines for responsible usage of AI/ML in Indian securities markets, 20 June 2025 (sebi.gov.in)
- SEBI, circular on safer participation of retail investors in algorithmic trading, 4 February 2025, and glide-path circular, 30 September 2025, with the framework in force for every broker from 1 April 2026 (sebi.gov.in)
- Speech by SEBI chairman Tuhin Kanta Pandey at FICCI's 23rd Annual Capital Markets Conference, 19 August 2026, as reported by Outlook Money
- Remarks by SEBI chairman Tuhin Kanta Pandey at the BSE Brokers' Forum Capital Market Confluence, 10 October 2026, as reported by ANI
- Digital Personal Data Protection Act, 2023, and Digital Personal Data Protection Rules, 2025 (meity.gov.in)
- NSE and BSE investor cautions on deepfake videos of their officials, 2024 (nseindia.com, bseindia.com)