Flo — AI Financial Intelligence Assistant

Flo is available from a persistent sidebar on every page. Ask about the platform, your data or a financial concept, and she answers using three sources at once:

1 · Your conversation — Flo remembers the thread, so you can ask a follow-up, correct a misunderstanding or drill deeper without restating context.

2 · This deployment's documentation — before answering, Flo searches the platform's own documentation and grounds her response in it. Answers reflect how your system actually works, rather than a language model's general impression of how ledgers usually work.

3 · Live data — when you ask about a specific account, balance or transaction, Flo queries the real thing. She does it using your own credentials, so what she can see is exactly what you can see. Never more.

After each answer she suggests two to four follow-up questions as one-click chips, so exploring a topic does not depend on knowing what to ask next.

Two enrichment paths run before Flo answers.

Documentation retrieval (purple path) — your question is matched by meaning, not keywords, against the platform's indexed documentation, and the most relevant passages are added to Flo's working context. If retrieval is unavailable, she degrades gracefully and still answers — no error, no dead end.

Live tools (gold path) — Flo is offered a set of governed tools she can choose to call mid-answer to fetch real balances, transaction history or verification status. Results feed back into her reasoning before she replies.

Your credentials travel with every tool call, so each one is subject to the same permissions and record-level visibility as a request you made yourself.

Answers are pushed to the sidebar the moment they are ready — no polling, no spinner waiting on a timer.

Flo keeps a rolling window of your recent exchanges within a session.

What that means in practice:

• Ask her to "explain account AC-1234", then follow up with "what was unusual about last Tuesday?" — she resolves last Tuesday against that account without you repeating it.
• Correct her — "no, I meant the liability account" — and she carries the correction forward for the rest of the conversation.
• Start a fresh session whenever you like. History is kept, browsable and resumable from the Flo History page.

Sessions are scoped to the authenticated user. No conversation is ever shared between users, which matters when the conversation contains customer financial detail.

Retrieval-augmented answering grounds Flo in LedgerFlow's actual documentation instead of a language model's recollection of financial software in general.

How it works: your question is converted into a semantic representation, compared against an index of the platform's documentation, and the closest passages are supplied to Flo as context before she answers.

What is indexed: platform documentation, functional requirements, API reference material, and any custom content you choose to add — so Flo can be taught your internal procedures, not just our product.

Why it matters: the failure mode of AI assistants in finance is a confident, plausible, wrong answer. Grounding is what turns "sounds right" into "is right, and here is where it comes from".

Graceful degradation: if retrieval is unavailable, Flo answers from general knowledge rather than failing. The user is never left staring at an error.

Live tools are what elevate Flo from something that explains the platform to something that answers questions about your actual business.

Flo is given a described set of governed capabilities she may call — account lookups, transaction history, balances, verification status, reconciliation results and more — and she decides when a question needs real data.

A worked example:

1. You ask: "Is account AC-5001 over its daily transaction limit?"
2. Flo determines she needs live data and looks the account up.
3. The lookup executes under your credentials.
4. She answers: "Account AC-5001 has processed 47 transactions today against a profile limit of 50. You are 3 transactions from the daily cap."

The security model is the point. Every tool call inherits your role, your permissions and your record-level visibility. Flo cannot see data you cannot see. She cannot bypass approval workflows, skip audit logging, or reach around any other control — not because she is instructed not to, but because the enforcement sits beneath her, in the same place it sits for every other caller.

An AI assistant whose access is governed by a prompt is a liability. This one is governed by the same machinery that governs your API.

Result: accurate, live, permission-safe answers — not confident estimates.

List and detail pages throughout the platform can surface an Ask Flo link beside individual rows. Click it and the sidebar opens with the row's context already filled in — account number, transaction reference, customer name and whatever else is relevant.

The point is removing the retyping. Nobody investigating an anomaly wants to copy an account number into a chat box; the question you were already asking is simply carried across, in context, with one click.

Because the context is assembled at the moment you click, the link behaves identically whether the sidebar is already open or opening fresh.

After each answer, Flo offers two to four suggested follow-up questions as clickable chips beneath her response — for example "What is the current balance?", "Show recent transactions", "Explain the fraud flag".

Click one and it becomes your next question. No typing.

This exists because the hardest part of using an assistant well is knowing what to ask next. New users in particular do not yet know what the platform can tell them. The suggestions turn a blank prompt into a guided path, and quietly teach the product as you use it.

Chips disappear as soon as you type something of your own — they are a prompt, never an obstruction.