Charlie works as an eight-layer investing AI stack: Your Systems, Portfolio Data, Market Intelligence, LLM, Tools, Client Config, Compliance, and Observability. The assistant in the middle is what the user sees. The stack around it is what makes the experience useful inside a regulated investing product.

The stack in one view

The eight layers are the practical building blocks around Charlie.

Your Systems

This is the institution's own environment: core systems, CRM, PMS, internal APIs, and workflows that already exist. Charlie is designed to connect into that reality.

This is also why deployment can happen in phases. A pilot does not need every system connected on day one. It needs the right systems for the first use case, then a clear path to expand. That rollout path is covered in Deployment options.

Portfolio Data

Portfolio data is the user's investing context: holdings, positions, transactions, performance, account information, and portfolio history. Without it, an assistant can explain investing in general. With it, Charlie can answer questions about this portfolio, this position, this exposure, and this performance period.

Market Intelligence

Market intelligence gives Charlie context beyond the portfolio: news, price movement explanations, research, instrument data, and approved market-data sources. It is what lets Charlie move from “your portfolio is down” to “these holdings moved, and this is the market context behind it.”

This layer should work with providers the institution already trusts where possible. It keeps the assistant useful without turning the model into a source of market truth.

LLM

The LLM is the reasoning and language layer. It understands the question, keeps the conversation coherent, chooses when the stack needs tools or data, and explains the result in a way the user can understand.

It is still only one layer. Financial institutions may prefer Anthropic, Google, OpenAI, Mistral AI, a private deployment, or another model setup depending on compliance, performance, cost, and infrastructure requirements. The model-choice logic is covered in Choose an LLM that fits your requirements.

Tools

Tools are where the work happens. Charlie should not ask the model to invent portfolio values, P&L, exposure, dividend income, concentration, or order-flow state. Those outputs should come from deterministic tools and controlled workflows.

This is the core idea behind The tools do the real work: the model explains, but tools calculate, check, retrieve, and act.

Client Config

Client Config is the white-label layer. It controls brand behaviour, tone of voice, greeting style, answer length, supported languages, topic scope, enabled tools, product boundaries, and the level of formality in the experience.

That matters because Charlie should not sound or behave the same everywhere. A self-directed investing flow, an advisor workflow, and a private wealth experience may need different tone, scope, and output style. The deeper configuration story is covered in Brand and tone.

Compliance

Compliance is the layer that keeps the assistant inside the institution's policy boundaries. It includes advice deflection, required disclosures, guardrails, escalation behaviour, and reviewable order or conversation flows.

The important point is that compliance is not just a disclaimer at the end. It affects what Charlie can answer, how it should answer, when it should ask a follow-up question, and when it should refuse or redirect. The traceability and guardrail setup is covered in Guardrails and full traceability.

Observability

Observability is how teams inspect and improve Charlie after it is live. It includes traces, logs, evaluations, review workflows, dashboards, and API-ready outcomes where an institution wants to connect results into its own tooling.

This layer is often invisible to the end user, but it is one of the layers that separates a demo from a production product. It lets product, compliance, engineering, and support teams see what happened and keep improving the system.