1. Docs
  2. Guides

AI budgets, costs & auto-tags

The AI assistant runs on your own Anthropic or OpenAI key, so you pay your AI provider directly. TalkingDot shows what it costs and lets you cap it.

See what it costs

Every AI request is priced from the tokens it used and your model’s price per million tokens (built-in estimates; change them under Settings → AI → AI assistant → Budget & costs → Price per million tokens if you have special pricing). Reports → AI answers shows the AI cost for the period, the cost per AI conversation, the cost per resolution and how much of this month’s budget is used. Settings → AI → AI assistant shows Spent this month.

Set budgets

  • Monthly budget: when it is used up, the AI hands new questions to your team and the ✨ helpers pause until the 1st (UTC). Owners and admins get an inbox notification and an email at 80% and 100%.
  • Budget per conversation: one long chat can’t run up the bill — past it, a teammate takes over.

These come on top of the daily limit and the replies-per-conversation limit.

What visitors see

  • Answers as they are written: words appear in the messenger as the AI writes them (WhatsApp and Telegram get the finished answer).
  • The AI bar: while the AI answers, a bar above the message box says You’re chatting with Nova, an AI assistant with a Talk to a person button, and AI messages carry an AI label. In many countries telling people they talk to an AI is required — keep it on.

Auto-tags & visitor mood

Turn on Settings → AI → AI assistant → Help for your teammates → Auto-tags & visitor mood and, a minute or two after a conversation is closed, the AI reads it once and:

  • adds up to 3 of your tags that clearly fit — they fill Reports → Top tags, the tag folders in the inbox and the conversation.tagged webhook. It never removes a tag a teammate added. With May create new tags on, it may add one new short tag when none of yours fit;
  • notes the visitor’s mood — 🙂 positive, 😐 neutral or 🙁 negative — shown in the conversation details, in Reports → Visitor mood (with the latest unhappy visitors to follow up) and as sentiment in the REST API.

It is one small request per closed conversation and counts toward the daily limit and monthly budget; when either is used up, conversations are simply left unlabelled. A reopened conversation is looked at again only if new messages arrived. Conversations where the visitor never wrote are skipped.

Last updated November 20, 2026 Something unclear? Tell us