Alongside the onboarding wizard sits a compact assistant that can talk you through setup end to end. It does more than answer questions: it opens each step with a suggestion, drives the work forward when you want it to (drafting your pipeline, loading sample data, staging your imports, inviting a teammate), and finishes by handing you your first real action in the product. Nothing it does reaches your CRM until you approve it.

The manual wizard is always right there. The assistant is an optional concierge over the same steps, never a replacement: you can ignore it, close the rail, or do any step by hand and it stays out of your way. It runs the same Laureo AI as the main app's chat, but with a different persona, a smaller toolset, and stricter guardrails. The copilot is a guide you stay in control of, not an autopilot.

On each wizard step the assistant opens with a short, specific message about that step and one or two suggested moves. The opener is contextual: on the pipeline step it offers to draft stages for your segment, on the data step it offers to connect your inbox or load sample data, on the invite step it offers to add a teammate.

The escape hatches are always visible. Every step keeps its normal wizard controls, so you can decline the suggestion and do it yourself, type your own request instead, skip the step, or collapse the assistant entirely. The opener is an offer, never a gate. You are never forced to reply to the assistant to move on.

Under each turn you get a small set of one-tap reply chips, never more than three at a time, in two flavors:

• Step-driving chips move setup forward. They are the primary set: "Draft my pipeline," "Load sample data," "Skip this step." Tapping one takes the action (through the approval flow below).
• Asking chips answer a question without changing anything. This set is smaller and secondary: "What stages do agencies use?" gets you advice, not an edit.

Chips are shortcuts, not the only way through. You can always type your own message, and typing a request is exactly as valid as tapping a chip. The built-in starter chips answer instantly.

The copilot drives setup through a curated subset of tools: read-only lookups plus a few onboarding-specific writes, each routed through your approval:

• Read your CRM: list companies, people, deals, pipelines, etc. to answer "what's already here?" questions.
• Draft and confirm your pipeline: suggest stages for your segment, which you keep, edit, or replace before anything is saved.
• Mark an onboarding step complete or skipped via the mark_onboarding_step tool.
• Load sample data for the org via the load_sample_data tool.
• Send you to the inbox-connect flow via the request_inbox_connection tool (it opens the sign-in prompt for you).
• Invite a single teammate via the invite_team_member tool (one at a time, with explicit name/email/role).

Everything the assistant writes goes through the same three-beat pattern: propose, confirm, undo. When it wants to change something (create your pipeline stages, load sample data, stage a batch of imported contacts, send an invite), it first shows you exactly what it will do in a confirmation card.

• Propose: the card spells out the change in plain language before anything happens.
• Confirm: nothing writes until you click to approve. You can edit the proposal or reject it outright.
• Undo: after a change lands you can adjust it from the same wizard step, so a wrong turn is never stuck.

This is the same approval discipline as the main assistant. Nothing reaches your CRM, and nothing goes out to anyone, without your explicit OK.

The assistant can move setup forward, but only through the approval-gated actions above. Outside that set it has no destructive or outbound tools:

• It cannot delete any record (company, person, deal, activity, anything).
• It cannot send email, log a call, send an SMS, or push to any external integration on its own. The finish step seeds a draft, but you are the one who clicks send.
• It cannot edit existing records: no field updates, no stage changes, no reassignment.
• It cannot bulk-invite or modify role permissions.

If you ask the copilot to do something it doesn't have a tool for, it will say so and point you to the right surface (usually a wizard step or an admin page). Two independent gates enforce this: the prompt-level filter that hides non-onboarding tools, and a dispatch-site allowlist check that rejects any tool call the prompt shouldn't have produced. The double-gate is defense against prompt-injection from content the AI reads.

The assistant runs on your workspace's existing AI budget, the same monthly allowance the main assistant draws from. There is no separate onboarding fee and no per-prompt charge, and the built-in starter chips answer without drawing from your budget. It is offered to team members with an AI seat.

It also degrades gracefully. If your AI budget is spent, or the assistant is unavailable, setup falls back to the manual wizard: every step still works with its normal controls, so you are never blocked from finishing. The conversation is a faster path through the same wizard, never the only path.

The last step does more than mark setup complete. Instead of dropping you onto an empty screen, the assistant seeds your first real action in the product, for example a first follow-up email or a first task, with a draft already filled in for you to edit.

One click sends or saves it, and the assistant never sends anything on its own. You review the draft, change what you want, and press send yourself. From there you land in the app with real momentum instead of a blank start.