Why we don't let the LLM trade for you

A lot of retail investing platforms launched in the last 18 months promise “AI-powered” trading. Some of them mean it. Most of them are running a chatbot on top of a feature-flagged broker integration that, technically, could execute trades on your behalf. A small number do execute trades.

We thought hard about this and chose not to. Sthira ships an LLM, but the LLM diagnoses, explains, and suggests. It never places an order. This post explains why.

The three tiers of trading automation

If you map the space, retail-investing automation falls into three buckets:

• Tier 1: Standing instructions. Systematic Investment Plans (SIPs), recurring purchases, fixed-schedule rebalancing. You decide the rule once, the system executes it on schedule. The system has zero discretion. This is what mutual funds have done forever.

• Tier 2: Pre-authorized event response. You set thresholds — “if my portfolio drifts more than 5% from target allocation, rebalance back” — and pre-authorize the response. The system executes when the threshold is crossed. The system has bounded discretion: it knows what to do, it knows when, you signed off on both.

• Tier 3: Open-ended agent execution. The LLM looks at market state, your goals, news, sentiment, whatever, and decides what to buy or sell. You set high-level intent (“grow my portfolio”); the agent picks the trades. The system has unbounded discretion within whatever guardrails the prompt sets.

The first two are well understood, well regulated, and broadly safe for retail investors. The third is genuinely new ground, and it’s where most “AI trading” pitches live.

Why Tier 3 isn’t a feature, it’s a different product

When an LLM places a trade on your behalf, three things shift:

The license boundary moves. In India, executing trades on behalf of someone else — even with their consent — at any meaningful scale falls under the Portfolio Manager regulations. That’s a separate license, separate capital adequacy requirement, separate compliance regime, and a separate target audience (HNI, not retail). It’s a real category, not a feature flag. Retail platforms that quietly ship agent execution are operating in a regulatory grey zone that probably doesn’t survive contact with an enforcement action.

The behavioural model changes. A SIP user who watches their portfolio go down knows it’s their plan that’s underwater. An agent-execution user who watches their portfolio go down has to wonder whether the agent made a bad call. The distance between user and decision affects whether they hold through volatility. The data on this is decades old: investors who feel ownership of a decision hold longer; investors who feel they were steered into one panic-sell sooner. Agent execution maximises the second mode.

The failure modes get worse. A buggy SIP buys 5 of the wrong unit; recoverable. A buggy agent system buys 50 units of the wrong stock during a flash drawdown; not recoverable. The blast radius of automation scales with how much discretion the automation has. The retail audience cannot meaningfully audit an LLM’s reasoning before it acts.

Where we drew the line

Sthira does Tier 1 and Tier 2 and stops:

• SIPs that the user configures and reviews — standing-instruction automation, regulated and understood.
• Drift rebalancing: the user sets a target allocation, a drift threshold (e.g. 4%), and pre-approves the corrective action. If equities drift to 64% of a 60% target, the system queues a rebalance — and the user gets a notification before the trade, with a 24-hour window to cancel. Bounded discretion with a human-in-the-loop affordance.

The LLM in Sthira reads your portfolio and explains things. It can tell you why your concentrated holding is volatile, what’s driving an unusual move in a stock you own, or what a fund’s expense ratio means for your long-term returns. It cannot place an order. The order button is yours and yours only.

This is roughly the same line responsible US-market platforms drew (Wealthfront, Betterment): clear advice, fiduciary framing, no Tier-3 agent execution despite having the LLM capability to do it.

What “Sthira insight” actually does

A concrete example: you open the app, the LLM has noticed your portfolio is up 8% this quarter but the gain is concentrated in one mid-cap holding that’s now 22% of your total allocation. The LLM surfaces:

“Your portfolio is up 8% this quarter, but ABC Industries is now 22% of your holdings — up from your target of 8%. If you’d like to trim it back, the tax-aware rebalancer can show you the lots to sell to minimise LTCG impact. Or, if your conviction in ABC has changed, you can update your target and the drift alerts will adjust.”

Three things happen in that prompt that are important:

1. It tells you something true and specific about your situation.
2. It offers a next action that you choose to take.
3. It surfaces the tax implication — because for an Indian retail investor with LTCG and STCG complexity, the tax cost is often bigger than the move itself.

What it doesn’t do is press the button.

The conviction behind the line

Two beliefs underpin this:

1. Retail trust is built by restraint, not by capability. The platforms that have lasted in retail finance — index fund providers, fee-only advisors, the older Indian distributor model done well — are ones that consistently refused to do the slightly-more-aggressive thing. The platforms that did the more aggressive thing usually got their market cap one quarter and a regulatory hammer the next.

2. LLMs are not yet trustworthy as fiduciaries. Even setting regulation aside, the failure modes of LLM agents (hallucination, prompt injection, training-data anomalies) are not yet well-bounded enough to put real money behind them autonomously. They will be, eventually. We’d rather adopt the capability when the failure modes are characterized than ship the capability and learn the failure modes on user money.

If, later, the regulatory framework opens up and the safety profile of agent execution becomes auditable, we’ll consider it — as a separate product, sold to a different audience, with the disclosure framework that requires. For Sthira’s actual users — laymen building long-term portfolios — Tier 1 + Tier 2 is the right ceiling.

What this means if you’re evaluating retail apps

A practical checklist if you’re choosing between apps:

• Ask whether the app has authority to place trades on your behalf without per-trade confirmation. If yes, you’re in Tier-3 territory regardless of how the marketing pages describe it. Understand what the agent can do, what bounds it operates under, and what the failure procedure is.

• Ask what license the app is operating under. Brokerage license + advisory license is different from portfolio management. The distinction matters.

• Ask what the LLM is allowed to say vs do. “It explains my portfolio” is different from “it executes trades it thinks are best.” If the marketing copy doesn’t draw that line clearly, that’s intentional.

We built Sthira for the layman investor who wants disciplined investing without being micromanaged by an agent they can’t fully trust. If that describes you, we’d be glad to have you.