> ## Documentation Index
> Fetch the complete documentation index at: https://docs.yrka.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Use Zavi, the Yrka AI-powered admin chat assistant

> Zavi answers Manual questions, runs permission-scoped read tools, and lets you preview and confirm admin actions without leaving your current surface.

Zavi is the Yrka admin assistant. It answers questions using the official Yrka Manual, runs permission-scoped read tools against your live workspace data, and lets you preview and confirm write actions before anything is committed. You can open Zavi from the admin sidebar dock, the full workspace view, contextual **Ask Agent** links on admin surfaces, or the desktop Spotlight shortcut when it is available.

<Note>
  Zavi is not legal, HR, payroll, or compliance advice. Official Manual content is Yrka-authored guidance about the product. It is separate from your company's own Resources knowledge content.
</Note>

## What Zavi can do

### Answer Manual questions

When you ask Zavi a question, it retrieves answers from official Yrka Manual articles that match your current surface and route. Every answer includes citations — links to the source article — so you can open and verify the original content before acting on it.

If Zavi does not have enough Manual context to answer, it tells you directly rather than guessing. Use the feedback controls to flag answers that feel stale, incomplete, or incorrect. Feedback goes into documentation-quality review and does not create Inbox items or Tasks.

### Read your workspace data

Zavi has permission-scoped read tools that let you ask operational questions without navigating to a separate surface. Available read tools include:

| Tool          | What you can ask                                   |
| ------------- | -------------------------------------------------- |
| **Employees** | Look up employee records and status                |
| **Timecards** | Check timecard summaries for a person or period    |
| **Inbox**     | Get a summary of open work items and review queues |
| **Reports**   | Pull report context and review briefs              |
| **Resources** | Surface resource and training status               |

Report answers include source-linked action rows when the underlying card exposes an owning workflow, so you can move directly from the summary to the right place in the app.

### Preview and confirm write actions

For write actions, Zavi always shows a dry-run preview before anything changes. You review the preview and confirm explicitly. Zavi never commits a write action without your approval unless you have enabled Auto mode for that action type.

Write actions that support preview-and-confirm include:

* Sending messages and creating tasks
* Creating jobs and schedule events
* Requesting documents
* Updating inbox status and task progress
* Drafting and proposing schedule changes
* Reviewing time-off requests

<Warning>
  Confirming a preview runs the write through the normal Yrka domain action. It creates standard records and writes audit history the same way a manual action in the app would.
</Warning>

## Open Zavi

<Steps>
  <Step title="Choose your entry point">
    Open Zavi from any of the following:

    * The **Zavi** icon in the admin sidebar dock
    * **Admin > Zavi** workspace tab (`/admin?tab=agent`)
    * **Ask Agent** contextual links on admin surfaces such as Timekeeping, Messages, and Personnel
    * The desktop Spotlight shortcut when enabled
  </Step>

  <Step title="Ask your question">
    Type your question or request. Zavi uses your current surface and route to retrieve the most relevant Manual context. You do not need to specify which article to look in.
  </Step>

  <Step title="Review citations before acting">
    Open the cited article link in any Zavi answer. Read the source content and confirm it matches your situation before using the answer to make a decision.
  </Step>

  <Step title="Review previews before confirming writes">
    For write actions, inspect the preview — including the action type, affected records, dependencies, and any warnings — before selecting **Confirm**. If anything looks wrong, cancel and adjust your request.
  </Step>
</Steps>

## Use the full workspace

For larger review tasks, open the full Zavi workspace (`/admin?tab=agent`). The workspace gives you:

* **CSV and table action rows** — review bulk action plans with skipped-row details before confirming
* **Thread history** — return to recent conversations without losing context
* **Multi-step action plans** — see the full ordered plan, each step's dependency outputs, and per-step Auto posture before committing

## Auto mode

Auto mode lets Zavi execute eligible, bounded actions without a manual confirmation step for each one. It is off by default and applies only to a defined set of low-risk, reversible actions — such as internal tasks, draft-only schedule proposals, reviewed-source training drafts, and copy-safe employee-visible messages.

<Warning>
  Auto mode is not a general bypass. Mixed-sensitivity action plans — where some steps are eligible and others are not — fall back to preview mode automatically. Auto mode never applies to payroll, billing, access changes, credential actions, or any other sensitive write category.
</Warning>

Enable Auto mode only for actions that show safe preview evidence and meet your organization's guardrail settings. You can review and adjust Auto mode eligibility in **Settings > AI and Agent Controls**.

## Connected apps

Zavi can take governed actions in connected external apps — such as drafting or creating records in a third-party service — when Power Connector is enabled by your organization owner.

Before using connected app actions, review the **External app access** section in Zavi controls. It shows:

* Which providers are connected and their current status
* Global write and auto-execute state
* Provider-level overrides and any pause reasons
* Recent external action history with provider, destination, data class, and policy decision

Every external action goes through connection policy, data classification, and retention rules. Sensitive external writes — including payroll, billing, credential, and access changes — are refused and logged with a safe refusal record so you can see what was blocked and why.

<Note>
  Power Connector requires owner enablement, explicit risk acknowledgement, provider and action allowlists, and live account evidence before any external write can run.
</Note>

## Rate limits and budgets

Zavi has per-workspace usage limits. When you reach the rate limit, the workspace shows the reset time. Heavy analysis tasks — such as large report reviews or multi-step bulk plans — count more toward the budget than simple Manual questions.

If you reach the limit during a session, your thread history is preserved and you can continue once the reset period passes.

## Common issues

<AccordionGroup>
  <Accordion title="Zavi can't answer my question">
    The official Manual may not have enough context for that topic yet. Use the feedback controls to flag the gap. In the meantime, check related Manual pages or reach out to support for workflow-specific guidance.
  </Accordion>

  <Accordion title="An action was refused">
    Zavi refuses actions that fail permission checks, safety gates, provider readiness, or owning-domain validation. Open the refusal detail to see which gate blocked the action and what you need to resolve before retrying.
  </Accordion>

  <Accordion title="I want to use a connected app action but don't see it">
    Confirm that Power Connector is enabled by your organization owner and that the provider is connected with a live account in Integrations. Provider actions require both Power Connector enablement and an active connection before they appear as available options.
  </Accordion>

  <Accordion title="Auto mode ran an action I didn't expect">
    Review the Auto mode eligibility settings in **Settings > AI and Agent Controls** and narrow the eligible action types. Every Auto-executed action writes the same audit trail as a manually confirmed action, so you can inspect what ran and when.
  </Accordion>
</AccordionGroup>
