By Hemant Sood, Founder and Managing Director, Findoc Group
Financial markets have run on algorithms for decades, but agentic AI is a different kind of machine participation. A conventional trading algorithm follows a recipe written by a human: when specified conditions occur, it executes an order. An AI agent receives an objective, breaks it into tasks, selects tools, interprets changing information and decides what to do next within its permissions. The shift is not from slower code to faster code. It is from instruction-following systems to mandate-following systems.
In practice this could redraw the investment chain. On an earnings day, a research agent could read filings, compare management commentary with earlier guidance and flag inconsistencies. An execution agent could adapt order slices as depth changes, while a risk agent throttles activity and a compliance agent preserves the evidence behind each step. At Findoc, we already use AI for first-cut research and reconciliation of quarterly financials. The output still needs human review, but the time from a result announcement to a working analytical view has fallen from days to hours. When that capability moves from research into execution, information may begin reaching prices faster than human participants can read it.
Control may arrive before alpha
The most visible promise is better returns, but the first commercial gains are likely to come from better control. Agents can watch portfolios, disclosures, order books and operational systems at once. They could detect that a position is individually within limits yet dangerous when combined with options gamma, sector exposure and collateral concentration.
For retail investors, the most valuable agent may not be the one that trades most often. SEBI’s own study of individual traders in equity derivatives found that more than nine in ten lost money over the three years to March 2024. An agent that understands cash-flow needs, time horizon and existing concentration before suggesting action, and that sometimes adds friction by simulating a sharp fall or declining an unsuitable trade, would serve that population better than one optimised for activity.
Faster intelligence, faster failure
The same architecture changes the nature of risk. A rules-based algorithm is imperfect but bounded; an agent can take a route its developer never scripted. It may act on a false corporate announcement, a manipulated social-media signal or poisoned data before a human sees the error. If many firms rely on similar foundation models, data feeds and cloud providers, apparently independent agents may reach the same conclusion and exit together.
The Financial Stability Board’s 2024 assessment listed third-party concentration, market correlation, cyber risk and model and data governance among the principal AI-related vulnerabilities. Those concerns grow sharper once a model can act rather than merely analyse. The problem is no longer only hallucination; it is hallucination connected to capital, credentials and market access.
Essential limits therefore should not live only inside the model. Position limits, product whitelists, maximum order size, drawdown triggers, trading pauses and kill switches are better kept deterministic and independently enforced. An agent should not be able to rewrite its own boundaries, and every consequential action should leave a tamper-evident record of the data accessed, the model version, the tools called, the limits checked and any human override.
India can build an audit-first model
India begins with real advantages: fully electronic exchanges, near-real-time regulatory reporting and a deep technology workforce. SEBI’s 2025 framework for retail algorithmic trading already brought API-based order flow under broker responsibility, order tagging and exchange oversight, and its consultation material on AI use by regulated entities keeps responsibility for outcomes with the intermediary rather than the model.
The RBI’s FREE-AI committee, reporting in 2025, set out 26 recommendations across six pillars covering infrastructure, policy, capacity, governance, protection and assurance. Its central idea applies beyond banking: innovation and safeguards are designed together or not at all.
For brokers, asset managers and market infrastructure institutions, deployment should progress in stages. An agent can begin in shadow mode, recommending without acting, then move to human approval, then to capped autonomy in tightly defined products, and only after stress testing to wider permissions.
Evaluation should ask not just whether the agent made money but whether it respected limits, behaved predictably when data disappeared, resisted malicious inputs and could be stopped instantly. Regulation, in turn, will likely have to move from reviewing models to supervising systems, since a static model document cannot describe an agent that plans, calls external tools and interacts with other agents.
The next generation of markets is unlikely to be defined by machines operating without people. It will be defined by machines handling more decisions while humans retain responsibility for objectives, boundaries and exceptions. The institutions that benefit will not be those that grant AI the most autonomy, but those that can show, trade by trade, that autonomy stayed inside a system built to protect capital first.
-Hemant Sood is Founder and Managing Director of Findoc Group, a SEBI-registered financial services firm headquartered in Ludhiana. Views are personal.