Skip to main content
The AI Assistant is the front door to every piece of project intelligence Alleato can see. Ask it a question in plain English; it pulls answers from your budgets, contracts, change orders, commitments, invoices, RFIs, submittals, meeting transcripts, Teams messages, emails, and accounting data — at the same time — and writes back a single grounded response with citations. This article is the overview: where to find it, what it looks like, what you can do inside the chat. For how the AI actually picks answers, see AI Architecture and How AI Retrieval Works.

​
Where to find it

Open the AI Assistant at /ai-assistant — there’s a permanent link in the main sidebar, and a global widget button you can click from anywhere in the app to pop the chat open without leaving the page you’re on. There are two entry surfaces:
  • Full chat page (/ai-assistant) — the dedicated workspace with conversation history, side panels, voice input, and the trace panel.
  • Global widget — a compact floating chat that opens on top of any page. Conversations from the widget appear in your history just like full-page chats.

​
The chat surface

​
Conversation panel (left)

Every conversation you’ve had with the assistant is listed in the left sidebar, most recent first.
  • New chat — start a fresh thread. Existing thread is saved automatically.
  • Rename — click the conversation’s menu and choose Rename to give a thread a clearer title (useful when you want to find it later).
  • Delete — also in the conversation menu.
  • Search — type to filter your history by title.
A new conversation gets an auto-generated title from your first question. Rename whenever the title isn’t specific enough.

​
Chat area (center)

The main thread. You type at the bottom, the assistant streams its response above. Each AI response can include:
  • Plain prose — the answer.
  • Source citations — clickable references to the underlying meeting, email, document, or record. Click a citation to open the source.
  • Tool calls — when the assistant pulls live data (e.g. queries the budget, looks up a commitment), the tool calls render inline so you can see exactly what data the answer is based on.
  • Widgets — for some questions the assistant returns a richer card (a daily-update widget, a project status panel, an insight card) instead of plain text.
  • Artifacts — documents the assistant creates or edits open in a side panel so you can read or copy them without losing the chat.

​
Side panels (right)

Opened from the toolbar at the top of the chat:
  • Trace panel — shows the live model trace: which agent handled the question, which tools fired, how long each step took, retrieval results. Use this when an answer looks off and you want to understand why the assistant said what it said.
  • Workspace — pinned context for the current conversation (a project you want every answer scoped to, a document you’ve attached for follow-up Q&A).
  • Artifact panel — opens when the assistant generates a document. Lets you read, copy, or download without leaving the chat.

​
What you can ask it

The assistant is built for whole-business questions across multiple data sources, not for one-tool lookups. Some examples it handles well:
  • “What’s the projected over/under on the Union Collective project, and what’s driving it?”
  • “Summarize what happened on the Vermillion Rise job this week.”
  • “Did Brandon send anything about the steel delivery in the last 10 days?”
  • “Pull up every RFI on the East 96th project that’s still open past its due date.”
  • “Draft a reply to the latest message from Acme Mechanical letting them know the COR is approved.”
  • “How is the Smith Group invoice tracking against the commitment?”
  • “What’s our cash position across all active projects?”
It’s also fine to ask it about itself — “what tools do you have?”, “what data sources can you see?”, “how do you decide which agent answers?” — and it’ll pull from the docs (these articles) to answer. For questions it doesn’t currently support well, see AI Capabilities Index.

​
Conversation features

​
Attachments

Click the paperclip in the chat input to upload an image (screenshot, photo, plan markup). The assistant can describe what’s in it, extract text, and answer questions about it. Document attachments (PDFs, contracts, transcripts) attach through the Workspace panel instead of the input bar — that pins them to the conversation so every follow-up question can reference them.

​
Voice input

Click the microphone in the chat input to dictate your question instead of typing. Voice input uses your browser’s speech recognition; speak naturally, then click again to stop. The transcript drops into the input box where you can edit before sending. Voice output (the assistant speaking responses) is available via the avatar surface — see AI Voice.

​
Suggested prompts

The welcome screen on a new conversation shows shortcut cards grouped by category — common workflows like “project status”, “draft an email”, “what’s overdue”, and “summarize this meeting”. Click any card to drop the prompt into the input ready to send (or edit first).

​
Feedback

Below every AI response there’s a thumbs-up / thumbs-down. Use it.
  • Thumbs-up marks the response as a good example and feeds the quality model.
  • Thumbs-down opens a brief feedback form. The flagged response is reviewed and used to improve future answers.
There are separate feedback paths for specific things the assistant produces:
  • Task feedback — when the assistant created or surfaced a task that was wrong, accept/edit/reject from the task itself.
  • Email-draft feedback — when the assistant drafted a reply for you to send, marking it as “used as-is”, “edited”, or “rejected” trains the email model.
  • Packet card feedback — same pattern for the insight cards that show up in daily packets.
See AI Feedback & Learning for what each signal actually does.

​
How long it remembers

Two layers of memory:
  • Within a conversation — the assistant sees the whole thread. Reference earlier messages freely (“expand on that”, “and what about the second project?”).
  • Across conversations — the assistant keeps a long-term memory of facts about you and your projects (preferences, named people, ongoing situations). You can review and edit what’s stored at /ai-assistant → memory settings, or ask the assistant directly: “what do you remember about me?”
Full reference: AI Conversation Memory.

​
Which agent is answering

Under the hood the assistant is a team of specialized agents — Strategist (router), CFO (financial), COO (operational), CHRO (people), CMO (marketing), CRO (revenue), VPBD (business development). The Strategist looks at your question and either answers directly or hands off to the specialist. You’ll usually see this in the trace panel: “Routed to CFO — financial query”. For most users it’s invisible — you just get the right answer. When you want to force a specific lens (e.g. “answer as the CFO”), you can say so directly. Full reference: AI Architecture.

​
Privacy and access

  • Per-user scope. You see your own conversations only. Admins do not see other users’ chats.
  • Per-user data. The assistant only retrieves data you already have permission to see in the app. A user without Budget Read won’t get budget answers; a user not on a private commitment can’t ask about it.
  • Tracing. Every response is traced for quality improvement. Trace data is internal — see AI Tracing for what’s logged.

​
Common Questions

Why doesn’t the assistant know about a meeting I just had? The meeting needs to finish syncing through the pipeline first — transcript ingest, embedding, then Teams compiler. The full path runs every 30 minutes. If a meeting is more than ~45 minutes old and still missing, check the AI Source Health page. Why does the assistant sometimes say “I couldn’t check X”? A data source was unreachable or timed out for that question. The assistant is built to keep going with whatever else it has and tell you what it couldn’t reach — instead of silently dropping that source from the answer. Treat the message as a transient warning; try the question again a moment later. Can I scope a question to a specific project? Yes. Pin a project in the Workspace panel and every follow-up question is automatically scoped to it. Or just include the project in the question (“on Union Collective…”). How do I get a longer / more detailed answer? Ask for it. “Give me the full breakdown”, “show me every line”, “explain the reasoning step by step”. The assistant defaults to concise; verbose is on request. Can I export a conversation? Not yet from the UI. If you need a thread captured, copy the messages directly or ask the assistant to summarize the thread into a document — that creates an artifact you can download. What model is the assistant using? Depends on which agent answers. Strategist runs on openai/gpt-5.4; the CFO and most specialists run on openai/gpt-5.4-mini. Synthesis steps use openai/gpt-4.1. All routed through the Vercel AI Gateway. See AI Models & Providers. Why is the assistant suddenly slower? The most common cause is a question that triggers many tool calls (e.g. a cross-project rollup that hits 20 data sources). Watch the trace panel — you’ll see which step is taking the time. If a single tool is hanging, source health may be degraded. Can I undo a delete? Conversations are soft-deleted for 30 days. Contact an Admin to restore.