When AI Starts Saying “What to Buy,” How Far Is a Research Assistant from Investment Advice and Trading?
Reading note: This article discusses product and engineering responsibility, not legal advice, securities investment advice, or any claim of a financial-services license. Applicable rules must be assessed for the actual entity, service, and jurisdiction.
Series: Fintech and Compliance, with architectural implications.
Moving a financial Agent from explaining a fund to recommending a purchase and then executing an order is not one more prompt instruction. It changes the service promise and the potential economic consequences for the user. Decide which service the product is allowed to provide before implementing its tool surface.
That is why ETF Agent separates read-only investment research from potential trading effects. Explaining a public index, comparing allocations using private holdings, and executing a real order require different data permissions, business obligations, and action approvals. Calling them all “AI investment analysis” hides the contract the product is actually making.
1. Similar Questions Can Require Very Different Permissions
Consider three requests.
“How are the CSI 300 and CSI 500 indexes constructed differently?” This is generally explainable using approved public sources.
“Compare the risks of two allocations given my holdings and time horizon.” This needs authorization to use current personal holdings and investment goals, along with explicit dates, assumptions and limitations in the result.
“Buy according to that plan on my behalf.” This requests a change to the state of an external financial account. Research permission, a statement of buying intent, and the existence of a trading API do not individually grant execution authority.
In ETF Agent, we separated the action a model understands from what the Host admits. Native research does not inherently authorize trades. The design described here is a read-only research-system choice, not evidence of a securities advisory license, and not proof that adding a disclaimer makes personalized recommendations legally acceptable.
2. The Legal Boundary Cannot Be Guessed from Keywords
The China Securities Regulatory Commission's Interim Provisions on Securities Investment Advisory Business, effective from 2011, address securities investment advisory business. The Measures for the Administration of Suitability of Securities and Futures Investors impose obligations on covered operators supplying securities/futures products or services, including investor understanding, classification of offerings, risk disclosure and suitability matching.
Who and what a rule applies to matters. Not every public financial knowledge answer automatically constitutes regulated securities advisory business; it also does not follow that every AI output must be referred to a human. The provider's status, content of the service, business model, jurisdiction and specific regulatory framework all matter. Business and legal teams must assess the applicable classification.
A production system should define the product's permitted services first and implement the corresponding admission flows. The model must not bypass those flows by inferring that the user is “just chatting.” Equally, a handful of regular expressions containing “buy” or “return” should not indiscriminately reject legitimate financial education questions.
3. From Financial Facts to External Action: At Least Three Separate Decisions
A reasonable research pathway can be decomposed as follows:
Read the relationships in groups; the complete diagram remains available below.
- → Define research scope
- → Authorize public sources and portfolio reads
- → Calculate exposures and state assumptions
- → Publish analysis within research scope
- → Separate business and action authorization? (decision)
- Not authorized → Stop at analysis; no external effect
- Independently authorized → External action under separate contract
Complete relationship diagram · enlarge for detail
First: data admission. Which public files can be read? Is private portfolio access authorized for this user? Is the data fresh? A model may propose queries, but saying “I have permission” does not create permission.
Second: interpretation and computation. Portfolio look-through exposures, fees, daily returns and drawdowns need an explicit data window and calculation method. Deterministic calculation can reduce particular numerical errors, but it cannot decide an appropriate allocation on the user's behalf.
Third: the action boundary. Research text, general risk warnings, personalized investment advice and actual order execution do not inherit the same approval and confirmation just because they use the same report. Activities requiring distinct qualifications or higher authorization must pass through a separate business contract. If the system has no such authorized action, its model-facing tool surface must not pretend one exists.
The final “external action” node is a conceptual boundary, not evidence that today's ETF Agent can place trades.
4. Compliance Must Not Make Every Useful Answer Disappear
If a user asks about a public fund metric and authorization for private data fails, the public explanation need not stop. If one source's eligibility is unknown, that should not automatically mean the model cannot answer anything. Through repeated work on publication and authorization, we learned to apply restrictions along actual data dependencies and action types rather than letting one rejection erase an entire research task.
There are risks in both directions. A permissive system may dress unsupported, ineligible or unauthorized claims as professional advice. An excessively restrictive one may emit only empty disclaimer templates, withholding otherwise legitimate and useful factual explanations.
Real runs have still shown incomplete research evidence, incomplete numerical reasoning, and citations not consistently matched claim by claim. We can deliver the verified parts alongside known gaps. We cannot claim that rejecting uncertain facts improved investment returns.
5. Regulatory Entities and Engineering Practices Are Different Things
On June 18, 2026, China's National Financial Regulatory Administration published its Guiding Opinions on the Secure Development and Application of AI in Banking and Insurance. These emphasize user-side responsibility and risk-tiered governance for the banks, insurers and other entities within their stated scope. They cannot automatically replace rules for the securities industry. Lessons may transfer technically without making all financial institutions subject to identical legal duties.
Engineering should establish what service the product may provide before deciding which powers to expose to the model. Before moving a research assistant toward personalized advice or trading, the team should answer:
- Service commitment: What explanations, recommendations, or actions can this entity and product actually offer? Which separate product, compliance, and human-confirmation workflows are required?
- Data eligibility: Who authorizes public and private portfolio inputs, and which time, definition, and use restrictions must travel with the output?
- External effect: Is the model only proposing a plan, or will account state change? Who owns approval, idempotency, execution receipts, and unknown outcomes?
These decisions belong in product contracts and protected execution, not in a generic “invest carefully” prompt. A read-only research scope is not an engineering weakness. It is an appropriate product choice until advice and trading obligations are explicitly established.