The pitch is compelling: an AI agent that can read market signals, generate trading strategies, and execute them — all without human intervention between steps. Vendors are lining up to show you prototypes. Goldman Sachs and Jane Street are reportedly exploring how large language models can augment their quantitative strategies. Bloomberg has been integrating LLM capabilities into its terminal ecosystem. Even Infosys is building agentic AI frameworks for its financial services clients.
But here’s the uncomfortable truth. The gap between a working demo and a production-ready trading system is measured in regulatory findings, unexpected losses, and career-ending incidents. If you’re a CIO, CTO, or Chief Risk Officer at an Indian financial institution or fintech, the question isn’t whether agentic trading will arrive. It’s whether you’ll be ready to evaluate it properly when it does.
Why This Is Different From Traditional Algo Trading
Algorithmic trading has existed for decades. What’s new is the “agentic” part — AI systems that don’t just execute pre-programmed rules but actively reason, adapt, and make decisions in real-time. Traditional algos follow if-then logic that humans can audit line by line. LLM-based agents operate through pattern recognition across vast datasets, making their decision paths harder to trace.
This matters because regulators — including SEBI in India — require explainability for automated trading decisions. When an agent makes a trade, you need to answer why it made that trade. “The model thought it was a good idea” won’t survive a regulatory inquiry.
The Four Things to Demand in Every Vendor Demo
First, backtested robustness across multiple market regimes. Any vendor can show you impressive returns during a bull run. Ask for performance data during the 2020 COVID crash, the 2022 rate hike cycle, and at least one emerging market currency crisis. If they only have six months of backtest data, that’s not a product — it’s a hypothesis.
Second, documented behaviour under market stress. What happens when liquidity disappears? When spreads blow out? When the agent encounters a market condition it wasn’t trained on? You need written specifications for circuit breakers, position limits, and automatic shutoffs. If the vendor can’t show you these, they haven’t thought seriously about deployment.
Third, latency guarantees with penalties. In trading, milliseconds matter. An agent that takes 200 milliseconds to make a decision might as well take 200 years in high-frequency contexts. Get contractual latency commitments tied to specific infrastructure configurations. Understand whether those guarantees hold when the agent is reasoning through complex scenarios versus executing simple trades.
Fourth, explainability that satisfies compliance. This is where most current prototypes fall short. Ask vendors to walk you through the decision audit trail for specific trades. Can they show you which data inputs drove the decision? Can they explain it in terms a regulator would accept? If the answer involves hand-waving about “emergent behaviour,” keep walking.
The Regulatory Landscape Is Moving Faster Than You Think
SEBI has been tightening algo trading rules since 2021, requiring brokers to get approval for any automated strategy offered to clients. The European Union’s AI Act specifically calls out high-risk AI systems in financial services. The US SEC has proposed rules that would require detailed documentation of AI-driven trading decisions.
Indian financial institutions operating globally — or serving global clients — will face the most stringent standard among all applicable jurisdictions. Building compliance infrastructure after you’ve deployed an agentic system is expensive and disruptive. Building it before deployment is just good planning.
Your Risk Team Needs New Capabilities
Traditional model validation teams know how to stress-test statistical models. LLM-based agents require different expertise. You need people who understand prompt injection vulnerabilities — where malicious inputs could manipulate agent behaviour. You need monitoring systems that can detect when an agent’s decisions start drifting from expected patterns. You need kill switches that work in milliseconds, not minutes.
Some firms are creating dedicated AI risk functions separate from traditional model risk management. Others are upskilling existing teams. Either approach works, but doing nothing doesn’t.
What This Means for You
Agentic trading will become a competitive factor in Indian financial services within the next two to three years. The firms that adopt it successfully won’t be the ones who moved fastest — they’ll be the ones who asked the hardest questions during vendor evaluations.
Start building your evaluation framework now. Require backtests across market regimes, stress behaviour documentation, latency guarantees, and genuine explainability. Train your risk teams on LLM-specific vulnerabilities. And remember: the vendor who pushes back on these requirements is telling you something important about their product’s maturity.
The goal isn’t to avoid agentic trading. It’s to adopt it on terms that won’t end up in a regulatory enforcement notice.