Three AI surfaces, three models, three jobs. Each chosen for the work: ephemeral cache for chat, audit-critical reasoning for deal scoring, fast tagging for the inbox.
A sidebar that knows everything the record knows. Open the chat on any company, contact, or deal and Sonnet 4.6 answers your questions against a fully-cached system prompt — related rows, the last 30 activities, the last 20 emails for contacts. Ephemeral prompt caching keeps it fast. No hallucinating about empty columns, because the cache makes it obvious what's blank. Use it for "summarize what's been going on with Acme", "draft an outreach based on the last quarter", "what's the status of the open audit," and the questions you'd otherwise ask an analyst.
A background worker scores up to twenty open deals per user per cycle, writing an AI score and a list of risk flags directly onto the deal. Audit-critical work gets the audit-critical model — Opus 4.7, not Sonnet, because a wrong call on a six-figure deal has consequences and we want the most defensible read. Scores show up in the inbox as suggestion cards if a deal moved across a threshold, so you don't have to go looking. Open the deal and read the flags — every one is grounded in a specific activity, email, or stage change.
Every inbound email tagged with one of six intents: intro, objection, buying signal, scheduling, churn risk, unrelated. The four reply-worthy intents (intro, objection, buying signal, scheduling) get a draft reply ready to go in the inbox. Sonnet 4.6 here because we triage volume — fifteen emails per worker tick — and we need fast and reliable, not deeply considered. The contact detail page renders an EmailList with cinnabar intent badges, so high-signal threads jump off the page.
Every AI surface tells you which model wrote the output. Your records never train models — Anthropic's policy by default. We use Anthropic exclusively; no model routing, no cheaper-when-we-can-get-away-with-it. The model selection is locked: Sonnet for chat and triage, Opus for scoring. If we change a model in the future, we'll tell you on this page first.