AI settings

The AI settings tab controls how the Mitzu AI assistant (the chat agent that answers questions, builds charts, and reads your catalog) behaves inside this workspace. There are three knobs:

1. Which Claude model it uses.
2. A persistent Custom instructions entry that is included in every conversation — your workspace's "system prompt."
3. Whether the agent is allowed to search the web in real time while answering.

Only Workspace Admins can change these settings. The AI settings tab itself only appears when the AI feature is enabled for your organization.

AI model

Pick which Claude model handles requests in this workspace.

Option	When to pick it
Auto (recommended)	Mitzu routes each request to the best-fit model for the task — heavy planning to Sonnet, quick lookups to Haiku. Best default for most teams.
Sonnet	Always use Sonnet. Stronger reasoning on complex SQL, multi-step chart construction, ambiguous questions. Slower and more expensive.
Haiku	Always use Haiku. Faster and cheaper; good for straightforward questions against a well-documented catalog. Less reliable on complex analysis.

If you're unsure, leave it on Auto.

Custom instructions

Standing context added to every AI conversation in this workspace. Use it to teach the agent domain language, reporting conventions, and business rules that aren't obvious from the catalog alone.

What to put here

Good custom instructions tend to cover:

• Terminology and acronyms. "Refer to monthly recurring revenue as MRR.", "In this workspace, 'deal' always means a closed-won opportunity."
• Calendar conventions. "Our fiscal year starts in April.", "The business week runs Monday–Sunday."
• Defaults for ambiguous filters. "Treat the trial property as false unless the user explicitly asks about trials.", "Exclude internal test accounts (user_email ending in @mitzu.io) by default."
• Reporting style. "Prefer weekly granularity unless the user asks for daily.", "Always break down by country when a geographic cut is possible."
• Safety rails. "Never join raw payment tables into public dashboards."

What not to put here

• Secrets, API keys, or credentials — the prompt is visible to every workspace admin.
• One-off questions — use a normal AI chat for those.
• Giant rulebooks — the prompt is capped at 4,000 characters. The counter under the box turns red when you go over, and the Save instructions button is disabled until you trim it.

Saving

Edits to the Custom instructions are not auto-saved. Click Save instructions to persist your changes; you'll see a success notification at the top right of the screen. Closing the tab with unsaved edits discards them.

How it composes with other context

The Custom instructions are one layer of context the agent receives. The full order is:

1. Mitzu's built-in system prompt (how the agent reasons, which tools it can call).
2. Your workspace Custom instructions (this setting).
3. The user's individual message in the chat.

Instructions at lower layers can refine but not override the ones above. For example, if the Custom instructions say "always exclude test accounts", a user can still ask "include test accounts this time" — the workspace default is a default, not a hard rule.

Web search

Off by default. When enabled, the AI agent can use Anthropic's built-in WebSearch tool to look up information on the public internet while answering — useful for questions that depend on current external context (industry benchmarks, public company news, third-party documentation).

Things to know before flipping it on:

• The agent decides what to search for. To do its job it will send the user's question (and any context it considers relevant — including text from your catalog or chat history) to a third-party search provider.
• Web results are untrusted content. Treat anything the agent quotes from the web the same way you'd treat a random page on the internet, not authoritative workspace data.
• Responses become non-deterministic. The same question can return different answers as the underlying web pages change.
• Only Workspace Admins can change this setting, and the toggle applies to every AI conversation in the workspace until it's turned off.

If your workspace handles sensitive data or you need reproducible answers, leave it off.

Tips

• Keep the prompt short and declarative. One rule per line is easier for the model to follow than long prose.
• When you add a new rule, re-run a few representative questions to make sure the agent honors it.
• If the agent starts ignoring part of your prompt, check whether the catalog contradicts it (the catalog wins for factual questions).