Ask AI: an assistant that knows your account
Mass puts an AI assistant everywhere you work. A context-aware bar lives at the bottom of the dashboard, a full command center can call real tools across your CRM, content, and funnels, and grounded Q&A answers questions about a lesson or article using only its indexed content. This guide explains each surface, how tool calling works, and how execution stays under your control.
15 min read · The complete Ask AI guide
The context-aware assistant bar
An always-present input that knows where you are and what you're looking at.
The dashboard carries a persistent Ask AI bar — an input that's always visible and expands into a conversation when you focus it. It's context-aware: its prompt adapts to whether you're in the agency view or a specific account, and it greets you with the current account's name so a question is always asked in the right scope. Expand it to see the back-and-forth; collapse it to get it out of the way.
The bar is the lightweight entry point — the place to ask a quick question about the account you're in without breaking your flow. For heavier work that creates or changes things, it leads into the AI command center described below.
- Always present — a docked input bar in the dashboard that expands into a chat on focus.
- Scope-aware — its prompt changes for the agency view vs a specific account and names the current tenant.
- Quick by design — the fast lane for a question about the account you're in, without leaving the page.
- Gateway — leads into the command center for actions that create or modify resources.
The AI command center
A full assistant with tool calling, a command launcher, and route-aware suggestions.
The command center is the full Ask AI surface. It streams responses and can call real tools — not just talk. A Cmd/Ctrl+K launcher opens a command palette that unions slash commands with your recent chat sessions, so you can jump to an action or reopen a conversation in a keystroke (Cmd/Ctrl+Shift+N starts a fresh chat, and a shortcuts panel lists the rest).
It meets you where you are: quick-action buttons and slash commands cover common jobs — search contacts, create a contact, view or create deals, list funnels, build a funnel — and the suggested chips adapt to the route you're on. Behind the scenes the assistant carries your collected account context into each turn so its answers and actions are grounded in your real data.
- Tool calling — the assistant can execute real actions, not only return text.
- Cmd/Ctrl+K launcher — a palette unioning slash commands and recent sessions; Cmd/Ctrl+Shift+N opens a new chat.
- Slash commands & chips — /contacts, /create-contact, /deals, /create-deal and more, with route-aware suggestions.
- Context-carried — your account context travels into each turn so responses and actions use real data.
Tools & the registry
A per-user tool registry across CRM, campaigns, generation, automation, and analytics.
Tool calling is backed by a central registry. Tools are organized by domain — CRM, campaign, generation, automation, and analytics — and are scoped per user via an authenticated client, so a tool only ever acts within your account. The chat endpoint assembles the available tool definitions into the model's prompt; the model invokes a tool by emitting a structured JSON block, which the client parses into a tool call.
Generation tools (punchlines, datasets, themes, templates, hub blueprints, and more) return artifacts that render inline in the chat, so a tool result isn't just text — it's a usable object you can act on. The same tool vocabulary is what the command center's quick actions and slash commands map onto.
- By domain — CRM, campaign, generation, automation, and analytics tools live in one registry.
- Per-user scope — tools run against an authenticated client, so they only touch your account.
- JSON-block invocation — tool definitions are injected into the prompt; the model calls a tool by emitting a structured JSON block.
- Inline artifacts — generation tools return artifacts (punchlines, datasets, themes, templates…) that render in the chat.
You stay in control of execution
Detected tool calls wait for your go-ahead before they run.
Detecting a tool call and running it are deliberately separate. By default the command center surfaces the tools the AI wants to call in a pending drawer rather than auto-executing them, so you review and approve an action before it touches your account. This keeps the assistant useful for mutating work — creating a contact, updating a deal — without handing it unchecked authority.
Tool execution carries credit accounting like the rest of the platform: an Ask AI turn checks and deducts credits, and tools that incur real external cost account for their own usage on success. The result is an assistant that can do real work, on your data, at a cost you can see, only when you say go.
- Detect ≠ execute — tool calls are surfaced for review in a pending drawer instead of auto-running.
- Explicit approval — you approve a mutating action before it touches your account.
- Credit-accounted — chat turns and tools deduct credits, and externally-costed tools account for their own usage.
- Real work, on demand — the assistant acts on your data only when you confirm.
Grounded Q&A: Ask-this-lesson & blog
Public-facing Ask AI answers only from indexed content, with citations.
Ask AI also shows up for learners and readers as grounded question-answering. On a course lesson, a learner can ask a question and the system embeds it, scores the lesson's indexed transcript chunks by similarity, builds a prompt from only the top excerpts, and returns an answer with citations and a grounded flag. The prompt is deliberately locked down: when the retrieved excerpts don't cover the question, the model refuses rather than guesses.
The same retrieval-grounded pattern powers content Q&A elsewhere — for example the AEO blog editor's assistant, whose tools run an audit, refresh rankings, or read a section, with mutating "propose" tools held behind an explicit apply gate. Across surfaces, grounding is the throughline: answers are tied to real, indexed content, not free-form generation.
- Ask-this-lesson — questions are answered from the lesson's indexed transcript chunks, with citations.
- Refuses when uncovered — a locked-down prompt declines instead of hallucinating when excerpts don't cover the question.
- Grounded flag — responses carry a grounded signal based on retrieval confidence.
- Same pattern, more surfaces — the AEO blog assistant and others reuse retrieval grounding, with apply gates on mutations.