Where the legal line sits
In the European Union, MiFID II defines investment advice as a personal recommendation to a client about transactions in financial instruments. In the United States, the Investment Advisers Act covers advice about securities for compensation, with the personalisation and the compensation both mattering. In India, SEBI regulates investment advisers and research analysts separately, with research analysts publishing general analysis and investment advisers giving client-specific advice. The common thread is personalisation: general information and analysis presented to many people is treated differently from a recommendation tailored to one person's situation.
None of these regimes exempts software. If a tool outputs "buy EURUSD now" to a named client whose portfolio it knows, the fact that a model wrote the sentence changes nothing. This guide is engineering guidance, not legal advice; the position you take needs counsel in each market.
Design rules that keep the assistant on the information side
- Answer questions with analysis, not instructions: what the data shows, how the tools scored it, what would change the picture.
- Present scores as data. A conviction score is a measured output of named tools, shown with its inputs, not a recommendation to act.
- Do not personalise to holdings or risk profile unless you hold the permission to advise; keep the assistant blind to the client's positions by default.
- Show the tool trace. Every answer names the tools called, the data window and the parameters, so a reviewer sees which model produced which number.
- Separate research from execution. The assistant cannot place orders; execution requires the client's own action or an operator-configured engine with caps and a kill switch.
- Make disclaimers true. "Not investment advice" means nothing if the output is a tailored buy instruction. The wording must match the behaviour.
- Log everything and keep it. Prompts, answers, traces and the version of each tool, retained under your recordkeeping policy.
- Govern the models. Document what each tool does, test it before release, re-test on change, and keep a human owner for each.
- Train the desk. Staff who relay an assistant's output to a client are giving the client something; they need to know what they may say.
- Review weekly. Sample conversations, check for drift into advice-like language, and tune prompts and guardrails with the desk.
Say this, not that
| Avoid | Prefer |
|---|---|
| You should buy EURUSD before the announcement. | EURUSD volatility regime score is elevated into the announcement; the shock-detection tool scored the last three releases above threshold. |
| This is the best trade for your account. | Across your watchlist, these three symbols rank highest on the criteria you set: volatility, liquidity and order-flow texture. |
| The model predicts a 2% rise. | The forecasting tool's central estimate over the next session is positive; the tool trace shows the inputs and the historical error band. |
| Guaranteed signal. | Pattern consistency score 0.82 over the sample, meaning the moves that followed this pattern were one-sided and low-entropy in the window shown. |
A ten-point deployment checklist
- A written regulatory position per market, from counsel, that names what the assistant may and may not output.
- A list of the tools the assistant may call, with an owner and a test record for each.
- A guardrail configuration that blocks instruction-style language and personalisation to holdings.
- Tool traces visible to the client and stored for review.
- No execution path from the assistant; execution behind separate controls.
- Retention of prompts, answers and traces under the firm's recordkeeping policy.
- Disclosures in the interface that describe what the tool does in plain words.
- Staff guidance on relaying outputs to clients.
- A weekly review with sampled conversations and a change log.
- A kill switch for the assistant itself, with a named person who can use it.
What two regulated brokers disclose about their assistants
The design rules above match what regulated brokers now publish. Questrade's disclosure for its AI Assistant, read on 3 November 2026, says the Assistant provides general market information and educational content; that it does not provide financial, investment, legal or tax advice; and that no information it generates constitutes a recommendation or solicitation to buy, sell or hold any security or to follow any particular investment strategy. It places a verification duty on the reader: all information provided by the Assistant requires independent verification, through the user's own review and official sources such as prospectuses, financial filings or the broker's primary trading platforms. And it separates research from execution in the strongest form available: the Assistant is a standalone information tool, not connected to the order execution or trading systems, and all trading activity must be performed manually by the user through standard platform tools.
Robinhood's support article for Cortex, its conversational assistant for Gold subscribers, read the same day, carries the position in capitals: Cortex is not an investment advisor and does not provide investment advice, recommendations or endorsements; all information, responses, insights and content it generates are provided solely for informational and educational purposes; it cannot give buy or sell recommendations or predict future price moves. Its execution model differs from Questrade's: Cortex can create an order, but the user must review and submit it. That is the other acceptable shape of the execution rule: the assistant may prepare, the client must act.
Read against the checklist, the two disclosures cover three of the ten points: general information rather than a personal recommendation, a verification duty on the client (explicit at Questrade), and no autonomous execution, whether the assistant is disconnected from trading systems or prepares an order the client submits. A broker or prop firm deploying an assistant to its own clients can use those three commitments as the skeleton of its disclosure, then add what neither text covers: the tool trace behind each answer, retention of prompts and answers, and the weekly review.
How SignalDesk is built against this line
SignalDesk is a conversational research assistant that calls EntropyX's proprietary quant tools (multi-asset forecasting, volatility regimes, news shock scoring and entropy-based pattern detection) and answers with a conviction score and a visible tool trace. Operators choose which tools it may call, set conviction thresholds and define what it must never do; logs feed a weekly review. It does not place orders. Where a firm wants systematic execution, FX Auto-Pilot runs versioned strategies with a dry-run sandbox, exposure caps, drawdown limits and kill switches that the operator, not the assistant, configures.
Sources and further reading
- MiFID II (Directive 2014/65/EU), definition of investment advice
- SEBI (Investment Advisers) Regulations and (Research Analysts) Regulations, in SEBI's list of regulations
- US Investment Advisers Act of 1940 (statute compilation, govinfo.gov)
- Questrade AI Assistant: important disclosures and risk information — Read 3 November 2026: general information only, no advice, no recommendation or solicitation, independent verification required, not connected to order execution, trading performed manually
- Robinhood Cortex: support article with the disclosure — Read 3 November 2026: not an investment advisor, no advice, recommendations or endorsements; informational and educational only; cannot give buy or sell recommendations; can create an order that the user must review and submit
This guide is engineering and product guidance from EntropyX Labs, a technology provider. It is not legal, regulatory or investment advice.