Arcvue AI — User Guide
Arcvue AI is a chat assistant that lives in the corner of every screen. You ask a question in plain English about your numbers — revenue, contracts, pipeline coverage, cash, debt — and it answers from your firm's own data, on whatever page you happen to be on. This guide explains what it is grounded in, then walks the exact controls the panel gives you.
For: anyone in the platform · Time: ~5 minutes · You'll need: access to the Arcvue FP&A workspace (the assistant appears on every FP&A page).
If your account is a Division Lead, the assistant scopes every data answer to your division — both in the SQL it runs and in its instructions. Ask "what is our pipeline?" and you get your division's pipeline, not the company total. Users without a division restriction see everything.
This changed on 2026-07-26. Before that, scoping silently never applied and every user's query ran across all divisions. If you used the assistant before then and the numbers now look smaller, that is the scoping working — not data going missing.
Part 1 — The ideas you need first (read once)
It is one floating panel, not a page. Arcvue AI is not an item in the left nav. It rides along on every screen as a small circular button in the bottom-right corner. Open it, ask, read the answer, close it — your place on the page never moves.
It answers from your data, not from the internet. Every answer is built on a snapshot of your firm's numbers that the platform assembles fresh for each question: your year-to-date P&L, your active contracts, your time-phased pipeline coverage, your indirect rates, your cash and AR/AP, your debt instruments, active M&A deals, and active pricing proposals. Generic government-contracting knowledge (what a wrap rate is, how DSCR is defined) is fair game, but any specific claim about your contracts, deals, dollars, or people must trace to a line in that data snapshot. If it isn't there, the assistant is instructed to say "I don't see that in my context" rather than invent a plausible-sounding number. That guardrail is deliberate — a made-up contract number is worse than "I don't know."
It sees only what your role can open. That snapshot holds the parts of the list above that your own screens show, and nothing more. A section drawn from a screen your role is denied (the P&L, cash, debt terms or M&A deals, for example) is left out of your snapshot, so the assistant answers without it rather than from figures you cannot see; a data question can read only the tables behind screens you hold, and the contract health narrative is refused to anyone without the Contract Forecast. Two people asking the same question can therefore get different answers, and both are right for the person asking.
It knows which page you're on. When you open the panel from a given screen, that screen's context travels with your question. Ask "what's the trend here?" on a contract page and it understands "here." That's also why the empty-state prompt changes per page.
It reads, it never writes. The assistant can only look up data — it can never change a number, post an entry, or edit a record. Nothing you ask it can alter your books.
Your conversation is your own. Chat history is per user. You don't see other people's conversations, and they don't see yours. What the assistant learns from conversations follows the same access rule: you are told only what was learned from screens you can open. Notes and learned insights never leave your company.
Three ways a question gets answered. Behind the scenes the assistant sorts each question into one of three lanes and pulls the right kind of context:
- a data question ("what were our FY revenue and margins?") pulls numbers from your database;
- an architecture question ("how does the forecast engine work?") pulls the relevant pages of the user guides;
- a knowledge / general question ("what is SCA labor?") answers from domain knowledge and any notes you've added.
You rarely need to think about this — but you can force a lane when it guesses wrong (Part 2, "Steer a question that routed wrong").
Ask for a report and it builds one. A request for a report, chart, table, breakdown or trend is answered with the report itself, rendered in the panel, with a Basis line under every block saying what period it covers and what it left out (Part 2, "Ask for a report, chart or table").
Part 2 — How to use it
The note form is a toggle and the button names the action available: + Add Note opens it and Cancel closes it again without saving.
Step 1 — Open the assistant
On any FP&A page, look at the bottom-right corner for the round teal button with a sparkle icon (hover text: "Ask Arcvue AI"). Click it. A chat panel opens anchored to that corner; its header reads Arcvue AI.
Before you've asked anything, the panel shows a short invitation. On most pages it reads "Ask about revenue, contracts, pipeline, cash position, or debt."; on a specific module it narrows to "Ask about your <page> data." Below it are a few suggested prompt buttons tailored to the page — click one to drop it straight into the input box, then edit or send as-is.
It is not the Help button. The Help button sits in the top bar — a
circled question mark, keyboard shortcut ? — and opens the written guide for
the screen you are on. Reach for Help when you want the documented answer,
and for Arcvue AI when you want one computed from your own data. If Help
opens on the wrong topic, it is keyed to the page you are standing on, so
navigate to that screen first and press ? again.
Step 2 — Ask a question
Type into the "Ask a question..." box at the bottom and press Enter (or click the teal send button). The answer streams in live, word by word, with a blinking cursor while it's still writing.
If you want to cut a long answer short, the send button becomes a red stop button while the answer is streaming — click it to halt.
Good questions are specific about what and when:
- ✅ "What were our largest indirect costs this fiscal year?"
- ✅ "Compare last year's overhead to this year's forecast."
- ✅ "Which contracts are at risk of expiring without a recompete?"
- ✅ "Is our cash position healthy for the next 90 days?"
- ❌ "How are we doing?" (too vague to ground)
Follow-ups work — the assistant remembers the last several turns of the current conversation, so "and by division?" after a revenue question is understood.
Step 3 — Ask for a report, chart or table
Ask the way you would ask an analyst — "chart booked revenue by division for this fiscal year", "table of open payables by vendor and how late they are", "trend gross margin by quarter for the last two years" — and the assistant builds the report in the panel instead of describing it. The panel widens itself to make room.
What you get, and what it means:
- A short sentence, then the report. KPI tiles for headline numbers, a chart for a trend or a ranking, a table for the detail. Every table carries the grid's export control, so the rows go to Excel like any other grid.
- A Basis line under every block. It states the period actually covered, that actuals stop at the last accounting close, which ledger answered which months once your books run natively, and anything the report left out and why (a measure your role cannot see, or a year where contract-level actuals cover only part of booked revenue). Read it before you forward the numbers.
- The numbers are computed by Arcvue, not written by the assistant. The assistant chooses the shape (which measures, which breakdown, which period); the platform runs the query. A figure in a report is the same figure the module's own page shows for that period.
- Save to library under the report keeps it. Arcvue stores the question, not the numbers, so the saved report runs fresh every time it is opened; press Open to see it in Custom Reports, where you can share, export or delete it.
Refine it in the same conversation: "same thing by quarter", "add gross margin", "just one division" change the last report rather than starting over.
If a request cannot be built from the measures the platform knows, the assistant says which measure or breakdown is missing instead of guessing. That is a request for the platform team, not a dead end.
Step 4 — (Optional) steer a question that routed wrong
If a numbers question comes back as a design explanation (or vice-versa), you can force the lane by starting your message with a bracket tag:
| Start your message with | Forces | Example |
|---|---|---|
[DATA] | a database lookup | [DATA] what does our pipeline look like |
[ARCHITECTURE] | a platform/how-it-works answer | [ARCHITECTURE] how does the forecast engine work |
[KNOWLEDGE] | a domain-knowledge answer | [KNOWLEDGE] what is SCA labor |
Plain English works too: begin with "This is a data question:" and the rest of your message is treated as a data lookup.
Step 5 — (Optional) give the assistant a note it should remember
Open the Knowledge Base (the book icon in the panel header). It slides open a small panel titled Knowledge Base listing any notes you've saved. To add one, click + Add Note, fill in the Title and Content fields, pick a category (General, Financial, Contracts, Pipeline, or Pricing), and click Save. From then on the assistant can pull that note in as context when it's relevant. Empty to start — it reads "No knowledge notes yet. Add one to give the AI more context." until you add the first one.
Step 6 — Resize, reset, or close
The panel header carries four controls, left to right:
- New Conversation (the circular-arrow icon) — clears the current chat and starts fresh.
- Knowledge Base (the book icon) — toggles the notes panel from Step 5.
- Expand / Collapse (the corner-arrows icon) — widens the panel for reading long tables, or shrinks it back. The panel widens itself when a report arrives.
- Close (the ✕) — hides the panel back to the corner button. Your history is preserved; reopening the button brings it back.
Part 3 — When something looks wrong
"It says it doesn't see something I know we have." The assistant answers only from the data snapshot the platform hands it for that question — it is built to refuse to guess past that boundary. If it says "I don't see that in my context," the underlying record most likely hasn't synced or been loaded yet (for example, a contract that isn't marked active, or a period that hasn't posted). Check the relevant module first; the assistant will see it once the data is there.
"I got a how-it-works explanation when I wanted numbers" (or the reverse). The router
guessed the wrong lane. Re-ask with a [DATA] or [ARCHITECTURE] tag on
the front of your message (Part 2, Step 3) to force it.
"The number looks stale." Answers reflect the last sync and computation, not a live wire into your source systems. If you just changed something upstream, give the sync/compute a moment and ask again.
"It won't change a number I asked it to fix." By design — the assistant is read-only. Make the change in the relevant module (or via the Bookkeeper for accounting), then ask the assistant to confirm the new figure.
"I hit a rate limit on the demo." The public demo caps AI chat to keep it responsive for everyone; you'll see a message asking you to wait a minute and try again. This limit does not apply to your own tenant.
"An answer got cut off." If you clicked the red stop button, that halts the response mid-stream — just ask again to get the full answer. On a genuine hiccup the panel shows a brief error; retry the question.
One-line summary
Click the sparkle button in the bottom-right corner, ask about your revenue,
contracts, pipeline, cash, or debt in plain English, and read the streamed answer
— grounded in your own data, read-only, per-user, and page-aware. Ask for a
report, chart or table and it builds one on the spot, with the Basis line
under every block. Force a lane
with [DATA] / [ARCHITECTURE] / [KNOWLEDGE] when it routes wrong, and add
notes via Knowledge Base to teach it your firm's context.
Related
- Ask about a specific page → open the assistant from that page so its context travels with your question.
- Teach it a durable fact → Knowledge Base → + Add Note.