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

# KI-Reife — the road to autopilot

> The KI-Reife tab shows how much autonomy each rule has earned, and lets you graduate the ones you trust to run without asking.

**KI-Reife** ("AI maturity") is the second tab of the [Entscheidungen](/user-guide/inbox/decisions) page — next to **Warteschlange** (the live queue). Where the queue is about deciding on *today's* suggestions, KI-Reife is the overview: for every rule the AI runs, it shows how far that rule has progressed from "still learning" to "ready to run on its own", and gives you a one-click **Jetzt automatisieren** to promote it.

Nothing here fires anything. It's a read-only picture of earned trust plus a deliberate promote button — the opposite of a black box.

## Two ways a rule graduates

A rule ramps up through the same trust threshold whether you promote it from the queue or from this tab — they're two views of one mechanism:

<CardGroup cols={2}>
  <Card title="From the queue" icon="bell">
    While you're clearing decisions, InboxMate nudges opportunistically: after enough approvals it asks **Ab jetzt automatisch?** right on the card. Handy in the moment.
  </Card>

  <Card title="From KI-Reife" icon="gauge-high">
    Here you see *every* rule's progress at once and promote proactively with **Jetzt automatisieren** — without waiting for the next matching email to trigger a nudge.
  </Card>
</CardGroup>

## The four stages

Each rule sits at one of four maturity stages, shown as a badge under **Regeln nach Reifegrad**:

| Stage           | German                   | What it means                                                                               |
| --------------- | ------------------------ | ------------------------------------------------------------------------------------------- |
| Learning        | **Lernt**                | Still gathering approvals, or its drafts are still being edited heavily.                    |
| Reliable        | **Zuverlässig**          | Consistently approved and edits are shrinking — trustworthy, not yet offered for autopilot. |
| Autopilot-ready | **Bereit für Autopilot** | Crossed the trust threshold and cleared every gate. One click away from auto.               |
| Automated       | **Automatisiert**        | Already running in auto mode. Its matches skip the queue entirely.                          |

The thresholds are deterministic, not a vibe score. With the default trust threshold of **5** approvals, a rule reaches **Zuverlässig** at 3 approvals and becomes **Bereit für Autopilot** at 5 — *provided* it also clears the gates below. A **Weg zum Autopilot** progress bar on each card shows how close it is (approvals ÷ threshold).

<Info>
  Both **Ja** and **Anders** (approve, or approve with a correction) count toward a rule's approvals. **Ignorieren** does not — declining an action never pushes it toward automation.
</Info>

## What holds a rule back

Approval count alone isn't enough. A rule is capped below **Bereit für Autopilot** while any of these is true:

* **Drafts are still heavily rewritten.** The **Bearbeitungs-Trend** metric tracks how much you change a rule's drafts before sending. If your recent edits are still large, the AI hasn't learned your voice yet and the rule stays at **Lernt** — no matter how many times you've approved it. (Actions that produce no draft, like archiving, have no edit trend and skip this gate.)
* **The rule has gone quiet.** A rule that earned readiness long ago but hasn't been exercised recently is held at **Zuverlässig** with the note *"Länger nicht bestätigt — für Autopilot erneut freigeben."* InboxMate won't advertise autopilot for a rule it hasn't seen work lately.
* **You dismissed the nudge.** Choosing **Weiter fragen** keeps a rule supervised until you promote it deliberately.
* **The action type can't graduate at all** — see below.

## Reading a rule card

Each card carries the stage badge, the **Weg zum Autopilot** bar, and three metrics:

* **Saubere Freigaben** — approvals with no correction needed (the real trust signal).
* **Gelernt** — how many style/fact exemplars this rule has taught the draft generator. This ties into what the AI has [learned from your edits](/user-guide/inbox/learning).
* **Bearbeitungs-Trend** — whether your edits to its drafts are growing or shrinking (or **keine Bearbeitungen** for action types with no draft).

When a rule is **Bereit für Autopilot**, the card shows **Jetzt automatisieren**. Clicking it opens a confirmation — *"Aktion automatisieren? Ab sofort führt die KI die Aktion „…" für die Kategorie „…" automatisch aus – ohne weitere Nachfrage."* — with a reminder that it's **reversible**: *"Du kannst das jederzeit im Kategorisierer wieder auf „Vorschlagen" zurückstellen."* Confirm with **Ja, automatisieren**.

## The headline

Above the rule list, three summary cards and a chart give the account-level picture:

* **Gesparte Zeit (geschätzt)** — estimated hours saved, *"Geschätzt aus tatsächlich ausgeführten Aktionen der letzten 8 Wochen."*
* **Automatisiert** — rules running fully automatically.
* **Bereit für Autopilot** — rules you could promote with one click right now.
* **Autonomie im Zeitverlauf** — a weekly sparkline splitting executed actions into **Automatisch** (fired on their own), **Freigegeben** (you approved via the queue), and **Manuell übernommen** (you overrode the AI).

<Note>
  The saved-time figure is deliberately honest and auditable — it's a real count of executed actions multiplied by a documented per-action constant (e.g. a drafted reply saves more minutes than an archive), broken down per action type. It is never an extrapolated growth projection.
</Note>

## Not every action can graduate

Eligibility for auto mode is fixed per action type:

| Can auto-graduate                                                                                                                                           | Never auto-graduates  |
| ----------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------- |
| **Archivieren**, **Ticket erstellen**, **Weiterleiten**, **E-Mail senden**, **Automatisch antworten**, **Als gelesen markieren**, **In Ordner verschieben** | **Endgültig löschen** |

**Endgültig löschen** (permanent delete) always stays a suggestion, however often you approve it — it's irreversible, so InboxMate never runs it unattended.

<Warning>
  Promoting a **Weiterleiten** or **E-Mail senden** rule also confirms it may send without further review. That confirmation is stamped the moment you click **Ja, automatisieren**, because those actions can't be recalled once automatic. Customer-facing replies therefore never reach autopilot by accident — only after you've approved that specific rule to the threshold and explicitly promoted it.
</Warning>

## Where this connects

* The decisions you make in the [Warteschlange](/user-guide/inbox/decisions) are what feed every number here.
* The **Gelernt** metric and the shrinking **Bearbeitungs-Trend** come from [what the AI learns from your edits](/user-guide/inbox/learning).
* When a rule is empty of activity you'll see *"Noch keine Regeln zum Reifen"* — approve a few suggestions in the queue and rules start appearing here as they earn autonomy.