The assistant now follows 12 imperatives on every analytical question:

• Ranks lists by impact instead of enumerating in random order.
• Cites the specific data point behind each claim, in compact form.
• Recommends a specific next 24-hour action — never just "you might consider".
• Triages long lists into Now / This Week / Watch / Skip.
• Surfaces non-obvious findings as "Worth flagging".
• Connects across data ("this contact's last activity was on Acme's open deal").
• Forward-looks on consequences ("if no contact by Friday, ...").
• And more — see the full list in the analytical voice block.

Simple lookups ("what's the address of Acme?") still get a 1–2 sentence prose answer — the structured TL;DR / Details / Recommendation shape kicks in for analytical questions only.

After every substantive assistant message, 3–4 chips appear below the response. Each chip is a one-tap follow-up that drills deeper, drafts outreach, schedules a follow-up, or compares alternatives. Chips never appear on simple confirmations or clarifying questions.

When you open the chat panel on a CRM detail page (a company, contact, or deal), the opener chips reflect the entity's actual state — stall time, recent activity, last contact — instead of generic “summarize this”.

Generation is cached per entity for 5 minutes, so revisiting the same page in a short window does not re-run the AI call. Each cache miss is metered against your AI budget alongside other AI features — see Usage Limits for how this counts toward your monthly spend.

Preferences saved via /remember or the AI Preferences settings page apply on every future chat. See the AI Preferences doc for details.