Where this lives. Sign in and open /ai in the app.
What it does
Three linked pages cover this. AI (/ai) states the ground rule up top: "AI drafts, classifies, and recommends. It never sends, calls, or spends — those actions stay behind human approval." Below that, a Safe automation policy card lists the actions your workspace has marked as requiring approval, enforced by the server regardless of role or AI confidence — an AI-generated draft queues as "awaiting approval" and shows up in Automations and approvals. A Review queue lists low-confidence classifications waiting for a human call, and Prompt templates shows the approved prompts shaping AI drafts, each with a usage count.
AI Models (/ai-models) shows which providers actually have a working key, not a hardcoded claim — the page explicitly notes it used to say "no provider is connected" even after one was added, and now checks live. If more than one provider is connected, the first one answers and the rest are automatic fallbacks if it fails. A separate "Copilot language layer" status shows on or off: when on, a model only helps read how you phrase things and word replies — it never decides or acts on its own, and asking the Copilot "who are you" reports this same state back to you. Below that, Task assignments lists the model each task is written against — for example reply classification, reply drafting (with knowledge-base grounding — see Knowledge), meeting summaries, and campaign copy — as shipped defaults, not something you can reassign per workspace. A Usage card is a placeholder until a provider is connected and AI calls start logging.
Copilot (/copilot) is the conversational operator. Its own description: "Every number is queried from your real data and every action runs the same code the buttons run, confirmed before it happens." A deterministic engine answers and acts on its own for things it already understands; the language layer (when on) only helps with wording. Anything that changes something shows as a confirm block in the chat — a title, a plain description of the effect, and Confirm/Cancel buttons — nothing runs until you click Confirm. Two buttons above the chat open side panels: Training shows resolution, no-AI-needed, and confirm rates measured from real conversations, plus the exact phrases people asked for that it still can't do; What I've learned lists vocabulary it has picked up (teach it directly in chat, for example: when I say "the big one", I mean "Q3 Enterprise") as plain, individually deletable rows, never hidden weights.
Before you start
- Connecting an AI provider key happens in Settings > Integrations, not on these pages — /ai-models only shows the resulting connection state.
- With no provider connected, the language layer stays off (even if switched on) and the Copilot still works for anything its deterministic engine already understands.
- The four task-to-model assignments on /ai-models are fixed in the current build; there is no per-workspace control to change which model handles which task.
Set it up, step by step
- Add an AI provider key under Settings > Integrations.
- Open AI Models (/ai-models) and confirm the provider shows "connected."
- If you want the Copilot to phrase things with the model instead of template text, turn on the language layer where your admin settings expose it, then verify it here.
- Open Copilot (/copilot) and start typing, or click one of the suggestion chips such as "What's wrong?" or "Show my campaigns."
- When the Copilot proposes an action, read the confirm block's effect line, then click Confirm to run it or Cancel to back out.
- Click What I've learned any time to review or forget a taught phrase.
- Check AI (/ai) periodically for items sitting in the Review queue that need a human decision.
What you should see
A connected provider shows a green "connected" pill; an unconnected one shows "no key." The language layer pill reads "on," "on — but no key," or "off." Training metrics show an em dash instead of 0% when a workspace has no conversations yet, so a fresh workspace reads as empty, not broken. A confirm block always names a concrete effect — it never asks you to confirm something it doesn't describe first.
Common problems
- The Copilot answers with generic wording instead of natural phrasing. No provider is connected, or the language layer is off; check /ai-models.
- An action you expected the Copilot to just do instead showed a Confirm button. That is by design for anything with a real effect — nothing outbound or destructive runs without your click.
- A taught phrase stopped working. Someone clicked "forget" on it in What I've learned, or it was auto-learned and later corrected.
Related
Common questions
Is AI-written cold email any good?
Only as good as what it is given. A model with your product material, your pricing and your real customer stories writes something specific. A model with a prompt and nothing else writes the same email every competitor is sending, and recipients recognise it.
Will the copilot send email on its own?
No. Anything that reaches a real recipient, buys a domain or changes infrastructure is put behind an approval. The copilot drafts and proposes; a person decides.
Which AI model should I use?
For copy, the strongest general model you have a key for. For classification work like sorting replies, a smaller and faster model is usually enough and much cheaper at volume.
Read next
Last updated September 19, 2026. Written by the team that operates the platform.