01 Investing AI
The Investing AI Agent Playbook
A strategic map for launching and scaling AI across digital investing experiences. It helps wealth, product, CX, and AI transformation leaders deploy faster, scale with confidence, and turn investing AI into measurable business value.
Playbook
4.2 Deterministic tools
The easiest way to misunderstand an investing AI agent is to focus too much on the chat bubble.
The chat bubble is what the client sees. It is the place where the investor types, Why is my portfolio down? or I want to buy 3 Apple shares. But the useful work is not done by the words alone. The useful work happens when the agent can select the right tool, retrieve the right data, calculate the right number, apply the right policy boundary, and move the user to the next valid step without guessing.
That is why the tool layer matters so much. In Charlie, the model is the reasoning and explanation layer. The tools are the operating layer. They are what let Charlie behave less like a general-purpose chatbot and more like a bank-deployable investing experience.
The model talks, the tools prove
A language model is very good at understanding intent, asking clarifying questions, and explaining something in a tone humans can actually follow. That is valuable. But in investing, a fluent answer is not enough.
If an investor asks why their portfolio is down, the agent needs more than a plausible market summary. It needs the actual portfolio. So the useful question is not, Can the model answer? The useful question is, Which tool must the agent call before it is allowed to answer?
Layer | What it is good at | What can go wrong if it works alone | Charlie design response |
|---|---|---|---|
Language model | Intent detection, explanation, summarisation, conversational flow | It may sound confident while missing portfolio data, policy context, or exact calculations | Use the model to reason and explain, but ground important answers in tool outputs |
Portfolio tools | Fetching holdings, balances, transactions, performance, composition | Without them, portfolio answers become generic and unverifiable | Retrieve real user-specific data before explaining portfolio state |
Market and news tools | Prices, instrument data, benchmark context, recent verified news | Without them, market explanations become stale or vague | Connect portfolio movements to current instrument and market context |
Execution tools | Order state, quantity, cash checks, order creation, workflow progression | Without them, the assistant can only talk about action, not support action | Carry eligible user-led flows through controlled steps and confirmation |
Compliance tools and guardrails | Advice-boundary control, topic scope, risk detection, auditability | Without them, a helpful answer can drift into regulated advice or unsupported behaviour | Check inputs and outputs against product-specific rules and log the path |
The tool principle in the project scope
For every user question, Charlie needs to choose the most fitting tool. That is the control mechanism. A tool does not answer the user by itself. It returns relevant data. The agent then uses that data to produce a response that is understandable, brief where needed, and grounded in the retrieved facts.
A practical tool inventory
The current Charlie scope exists of a concrete set of functional tools. Each tool exists because it maps to a recurring investor question or workflow.
Tool or capability | Investor job it supports | Why it matters |
|---|---|---|
| “What is the status of my investments?” | Gives the investor a grounded snapshot instead of a generic reassurance |
| “Is my portfolio diversified?” “What is my sector breakdown?” | Turns holdings into interpretable structure using metrics such as NEF and HHI |
| “How is my portfolio doing?” “Which stocks contributed most today?” | Explains return, contribution, gainers, losers, and market status with real calculations |
| “What have I traded recently?” “When did I last buy or sell this instrument?” | Lets the assistant answer account-specific operational questions without sending the user elsewhere |
| “Why did this instrument move?” “What news affected my holdings?” | Connects market events to actual positions, with cited sources rather than loose commentary |
| “What does ETF mean?” “Explain P/E ratio like I am 15.” | Supports education without pretending that every question requires account data |
| Questions requiring instrument prices, returns, or reference data | Keeps instrument answers data-backed and current enough for the use case |
| “I want to buy 3 Apple shares.” “Place a limit order for 1 Tesla share.” | Moves from intent to controlled workflow: clarify, check requirements, confirm, then route |
| Open orders, order history, dividends, cash movements, committed cash | Covers the practical account questions that otherwise become support tickets |
| Free internet search outside the usual news source | Shows that not every attractive capability should be enabled by default in a regulated flow |
FAQ
Why do tools matter so much in Charlie?
What are examples of Charlie tools?
How do tools improve accuracy?
What is the plugin layer for?
Where does MCP fit?
What should be visible after a tool-based answer?
Meet Charlie. The Investing AI Agent.
Give your clients an entire investing experience, rebuilt from the ground up as a conversational AI agent.
