# Personalizing advice with robo and LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → copilots
A client opens her banking app at 11pm and types: "I got a $40k bonus. Should I pay down my mortgage or add to my portfolio?" The app answers in seconds, referencing her goals, her tax bracket, and her existing holdings. That is the promise of the AI wealth copilot. The danger is that the same system, asked the same question by someone else, might confidently invent a fund that does not exist.
This lesson shows you how to design a copilot that personalizes goals-based advice, and where it breaks.
A robo-advisor is an automated platform that builds and rebalances a portfolio using rules, usually from a short risk questionnaire. Think Betterment or Wealthfront: you answer questions, you get a model portfolio, the software keeps it on target.
An LLM copilot adds a large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → (a system trained on text that can converse and reason over documents) on top. Instead of a static questionnaire, the client talks. The copilot explains why, drafts a plan, and answers follow-ups.
The two are complementary. The robo engine does the math and executes trades. The LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → handles conversation, explanation, and nuance. Keep those jobs separate. This is the single most important design decision in the lesson.
LLMs are good at language and terrible at arithmetic you cannot see. If you ask an LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → to "rebalance to 60/40," it may produce plausible-looking allocations that do not sum correctly or ignore tax lots.
So the architecture looks like this:
The LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → proposes intent. The engine decides and executes.
Goals-based investing organizes money around specific objectives (retirement, a home down payment, a child's tuition) rather than one blended "risk tolerance." Each goal gets its own time horizon and its own portfolio.
This is where personalization gets real. A copilot can hold context:
When the client asks about the bonus, the copilot can reason across all three. Short horizon goal is underfunded, so the deterministic logic favors it. The LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → then explains the tradeoff in plain language.
Here is a simplified structure for how the LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → captures a profile and hands it to the engine. Note that the LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → outputs structured data, not decisions.
{
"client_id": "anon_4821",
"goals": [
{"name": "down_payment", "horizon_years": 3, "target": 80000, "priority": "high"},
{"name": "retirement", "horizon_years": 25, "target": null, "priority": "high"}
],
"constraints": {"exclude_sectors": ["tobacco", "firearms"]},
"risk_signals": {"stated_tolerance": "moderate", "loss_reaction": "sold_in_2022"}
}That loss_reaction field matters. A client who says "moderate" but panic-sold in a past downturn has a revealed behavior that contradicts the stated answer. Capturing both is better personalization than any single questionnaire score.
Traditional risk questionnaires collapse a person into a number from 1 to 10. That fails in predictable ways.
Capacity versus tolerance. Risk *tolerance* is how much volatility you say you can stomach. Risk *capacity* is how much loss your finances can actually absorb given your goals and timeline. A 28-year-old with stable income has high capacity even if she feels nervous. A retiree drawing income has low capacity even if he loves stocks. Good copilots separate the two.
Stated versus revealed preference. People misreport. A copilot with transaction history sees the truth: did they add to the market in a downturn or flee?
Context drift. A job loss, a new child, an inheritance: these change everything and a static profile misses them. A copilot can prompt for updates when it detects a large cash inflow.
The SEC's Investor.gov site has plain-language material on assessing risk that is worth reading to see how regulators frame these concepts for retail investors.
Rebalancing means buying and selling to return a portfolio to its target mix after markets push it off course. The copilot's value is not doing the trade (the engine does that). It is explaining and timing.
Good copilot behaviors:
An LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → is a confident, fluent answer that is factually wrong. In wealth management this is not a quirk. It is a compliance and liability event.
Real failure modes to design against:
Retrieval, not memory. Use retrieval-augmented generation (RAG): the LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → answers only from a controlled document store (your approved fund list, current tax tables, the client's actual holdings). It cites the source. If the answer is not in the store, it says so.
Constrain numbers to the engine. Any figure shown to a client (allocation, projected balance, fee) comes from the deterministic system, injected into the response. The LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → never generates a financial number freehand.
Refusal and escalation. For anything outside scope (estate planning, specific tax filing, distressed markets), the copilot hands off to a human advisor. Design the handoff, do not treat it as failure.
Confidence and disclosure. Label output clearly as educational unless a licensed process approves it as advice. Log every interaction for audit.
def safe_to_show(response, approved_products, engine_numbers):
# Reject any ticker not on the approved list
for ticker in response.mentioned_tickers:
if ticker not in approved_products:
return escalate("unapproved product mentioned")
# Reject any number the engine did not produce
for num in response.financial_figures:
if num not in engine_numbers:
return escalate("unverified figure")
return responseSimple, but it enforces the rule that the LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → proposes language while the engine and the approved list own the facts.
Knowledge check
1. In the recommended architecture for an LLM wealth copilot, which component should be responsible for calculating allocation targets and generating trades?
2. Why does the lesson insist that the LLM should 'propose intent' while the engine 'decides and executes'?
3. What is the key conceptual difference between a traditional robo-advisor and an LLM copilot?
4. Select ALL correct answers. Which tasks are appropriately assigned to the LLM layer (not the deterministic engine) in this architecture?
Select all the correct answers.
5. Select ALL correct answers. What are genuine risks or reasons for separating the LLM from the deterministic engine?
Select all the correct answers.
In the US, giving personalized investment advice for a fee generally triggers duties under the Investment Advisers Act, including a fiduciary duty (acting in the client's best interest). A copilot does not remove that duty. The firm still owns it.
Regulators have signaled concern about AI washing (overstating AI capabilities) and about conflicts where an AI nudges clients toward products that benefit the firm. Design for transparency: disclose that AI assists, keep a human accountable for advice, and avoid optimizing the copilot for firm revenue over client outcomes.
Practical stance for 2026: treat the copilot as an assistant that expands what a human advisor can cover, not as an unsupervised advisor. Retail scale, human accountability.
A well-designed wealth copilot:
1. Talks to the client and captures goals, constraints, and revealed behavior as structured data.
2. Passes that to a deterministic engine for scoring, allocation, and tax-aware rebalancing.
3. Uses RAG and number injection so every fact is sourced and every figure is engine-generated.
4. Escalates to humans at defined boundaries and logs everything.
The magic clients feel (an app that seems to understand them) comes from context and clear language. The safety comes from never letting the language model be the source of truth.