# Set up the assistant

> Every setting in the AI Configurator Assistant group, from the API key to house rules, permissions and the journal.

Source: CBX documentation, version 3.x (released). Canonical page: https://docs.configbox.at/docs/admin-guide/ai-assistant/set-up-the-assistant. Last updated 2026-08-11.

---
> **Audience:** store owners and shop managers · **Scope:** the settings that switch the
> assistant on and shape its behaviour · **Last reviewed:** 2026-08-11

All settings live in **Settings → AI Configurator Assistant**. The assistant appears on your
configurator pages as soon as an API key is saved and the feature is enabled; no other setup is
required.

## The API key

Create a key in the Anthropic Console (console.anthropic.com) and paste it into the **API key**
field. The key stays on your server and is only used for the assistant's own calls. Costs are
billed by Anthropic per message — typically single-digit cents per visitor message, depending on
the model.

## Choosing the models

Two model choices, two different jobs:

- **Chat model** — answers visitors live. Latency and cost matter most here. Claude Sonnet is the
  recommended balance; Claude Haiku is the fastest and cheapest; Claude Opus is the most capable
  and most expensive.
- **Model for the journal analysis** — writes the per-conversation reviews and the Insights
  report. These run rarely and their quality decides what your team works on, so a bigger model
  than the chat one usually pays off.

## Persona, greeting, house rules

- **Persona** sets the assistant's stance: a neutral **advisor**, a **sales agent** that moves
  towards completing the purchase, or a precise technical **consultant**.
- **Greeting** is the first message in the panel. Leave it empty for the default.
- **House rules** are your own standing instructions, one per line — shop-specific facts, what to
  emphasise, what never to promise. They outrank the built-in guidance, so this is where you put
  things like "we deliver to Austria and Germany only" or "never promise a delivery date".

## Permissions

Three switches decide how far the assistant may go. All three are on by default, and each one only
ever lets the assistant do something the visitor could already do themselves on the page.

- **Assistant may add the configuration to the cart** — with this off, it advises and selects,
  but the visitor completes the purchase themselves.
- **Assistant may navigate the configurator** — lets it switch pages and point at questions. With
  this off it can only talk.
- **Assistant may change the cart** — lets it change a line item's quantity, remove it, copy it, or
  reopen it for editing when the visitor asks — the same things the buttons on the cart page do. It
  always asks before removing anything, and it can never touch a cart that has become an order.

The last one is its own switch on purpose: agreeing that the assistant may complete a purchase is
not the same as agreeing that it may empty the basket. Reopening and copying a line also need
**may navigate the configurator**, because both hand the visitor back to the configurator to finish
the job — with that off, the assistant tells them to use the cart's own buttons instead.

## Spend control

**Messages per visitor per hour** caps how many messages one visitor can send. It is a lid on
API spend, not a security feature.

## The journal settings

- **Keep the diagnostic journal** — records every conversation for the
  [journal](https://docs.configbox.at/docs/admin-guide/ai-assistant/read-the-conversation-journal). Visitors' words are stored verbatim, so mind your
  privacy policy.
- **Journal retention days** — conversations older than this are deleted by the scheduled
  script. Retention is a promise about visitor data; the deletion only happens when
  [the script actually runs](https://docs.configbox.at/docs/admin-guide/getting-started/schedule-the-recurring-scripts).
- **Analyze conversations automatically** — lets the scheduled script write a lessons-learned
  review for each conversation once it has gone quiet. You can always analyze single
  conversations by hand from their journal page instead.
- **Also write the journal to a log file** — a plain-text mirror for developers; normally off.

## Feeding the assistant knowledge

The settings above shape behaviour — but what makes the assistant *good* is product content:

1. Write real **descriptions** on products, questions and answers.
2. Fill the **detail panes** with specifications, delivery terms and warranty conditions — the
   assistant quotes them when visitors ask.
3. Watch the [wishlist](https://docs.configbox.at/docs/admin-guide/ai-assistant/work-the-wishlist): every `content` item there is a visitor question
   your data could not answer, named concretely enough to fix.
