Ask AI

Ask AI is the chat assistant inside the side panel. It runs against the same Vapi assistant that powers the Voice Concierge — so the bot's persona, knowledge, and tools are identical whether you type or talk. It's grounded in the current company's cached AI Overview, the user's profile, and the specific question you ask.

Two ways to start

• Type a question. A Perplexity-style auto-growing textarea lives at the bottom of the Ask AI view. Hit Enter to send.
• Tap a quick-start bubble. When you open Ask AI, Claude has pre-generated three to five context-appropriate prompts (e.g. "Track an order", "Dispute a charge", "Cancel my subscription"). One click sends the prompt and renders the response.

What grounds the response

Every Ask AI call sends Vapi:

• The company name and AI Overview (description, HQ, top reasons) as assistantOverrides.variableValues so the assistant prompt can reference {{currentCompany}}, {{phone}}, {{channels}}, etc.
• The customer profile as {{customerName}}, {{customerTier}}, {{customerSentiment}}, etc. — never sensitive identifiers.
• The conversation thread via previousChatId, so the assistant remembers earlier turns within the session. Switching companies starts a fresh thread.
• The question itself as the chat input.

The system prompt — defined in your Vapi dashboard — should:

1. Stay on-topic for the company in context.
2. Cite the source (the company's overview or a published policy) when making factual claims.
3. Refuse questions that require browsing the web or calling external APIs — Ask AI is read-only against the cached context.
4. Render responses as Markdown so the panel can format bullets, bold, and links cleanly. Replies never contain raw *asterisks* or **bold** syntax.

Configuration

Ask AI is gated on three Vapi values entered under Admin → Voice AI:

• Private API Key — required because Vapi's /chat REST endpoint rejects the public key with 401. The public key is scoped to the Web SDK only.
• Assistant ID — same one the Voice Concierge uses.
• Enable voice concierge toggle — must be on (kill-switch).

If any of those are missing the panel shows a one-click jump to the admin screen instead of the chat UI. There is no separate Anthropic API key any more.

The private key can spend on your Vapi account, so the admin warning marks the device as trusted. A future iteration can mint short-lived chat credentials server-side instead of holding the private key in chrome.storage.local.

Privacy notes

• Ask AI does not see Plaid purchase data unless the user explicitly attaches a receipt or transaction to the prompt.
• Conversation history stays in chrome.storage (device-local) by default. Backend recording is opt-in for autonomous-mode rollout.
• Ask AI calls go directly to api.vapi.ai/chat from the panel using your Vapi Public Key. Vapi's logging is governed by your Vapi account settings.

When to reach for Voice Concierge instead

Ask AI is great for one-off questions that have textual answers. The Voice Concierge is the right tool for multi-turn conversations or any task that benefits from speaking and listening (driving, hands-free, hold queues). The Voice Concierge has access to your entire Hub, not just the current company.

Related

• AI Company Overview → — the grounding context.
• Voice Concierge → — the spoken-conversation alternative.