The Shared Inbox
The inbox brings together all incoming customer conversations across every channel in a single view. You'll find it in the console under Inbox. Every conversation is a thread with the full history, customer context and the associated channel — whether the message arrived by email, via Instagram, WhatsApp or webchat.
At a glance, each thread shows you:
| Element | Description |
|---|---|
| Channel | How the message arrived — email, Instagram, WhatsApp or webchat. |
| Handler | Whether the AI is replying or a human has taken over the conversation. |
| Maturity | The AI's phase for the detected topic: Shadow, Copilot or Autopilot. |
| History | The complete message history, including AI suggestions and sent replies. |
Channels
Alpheris connects the following channels to the inbox. You set up the connection in the console under Email or Integrations.
For email, Alpheris supports three connection methods:
- Outlook / Microsoft 365: connection via Microsoft Graph — the standard method for Microsoft mailboxes.
- SMTP / IMAP: a generic method for any other mail server where no OAuth connector is available.
Incoming mails land as threads in the inbox, and outgoing replies go back out through the same connected mailbox. The AI learning and ticket logic (see Maturity) is designed above all for the email channel.
Instagram & WhatsApp
Instagram and WhatsApp are connected via Meta. You authorize the connection via OAuth; incoming messages reach Alpheris through signed webhooks from Meta. For WhatsApp, this includes a verify-token handshake during setup. Incoming events are verified server-side by signature so that only genuine Meta events are processed — details on this under Webhooks.
AI Reply Suggestions
For every incoming customer message, the AI generates a reply that is grounded in your knowledge base — that is, in the crawled website content, uploaded documents and (for Shopify) your product catalog. The suggestion is produced through hybrid retrieval (full-text plus vector search with subsequent reranking), so that the reply is tied to concrete sources rather than being made up.
For each suggestion, the system determines a confidence from retrieval signals such as source coverage and match quality. This confidence governs whether a suggestion is only shown as a draft or may be sent automatically.
The AI automatically adapts tone and length to the channel:
| Channel | Style | Length |
|---|---|---|
| formal, with a formal salutation | medium | |
| Webchat | balanced, neutral salutation | medium |
| friendly, casual | short | |
| friendly, casual | short |
You configure the personality and reply rules of your AI agent under AI Agent. If a customer replies in another language, Alpheris detects it automatically and answers in the same language — more on this under Multilingual.
Maturity: Shadow → Copilot → Autopilot
Alpheris does not hand full control to the AI from day one. Instead, the AI goes through a three-stage maturity model per topic area (intent cluster, e.g. shipping questions, returns, product questions). Each topic area is assessed and unlocked individually — so the AI can already be running on Autopilot for shipping questions while product questions are still in Copilot.
Promotion to Autopilot is tied to hard thresholds — nothing is unlocked on gut feeling alone:
- At least 50 accepted suggestions in this topic area.
- An acceptance rate of at least 75% (the share of AI suggestions adopted without significant correction).
- The topic area must be qualified: at least 5 cases, and the assessment must not be too old (a 14-day window).
The minimum confidence required per individual reply decreases as the topic area matures: 0.85 by default, 0.78 for established topics and 0.72 for fully matured topics. If a specific suggestion falls below the respective threshold, it is not sent automatically but remains a draft.
Auto-Reply: Gating & Control
Auto-reply — that is, autonomous sending by the AI — is not an all-or-nothing switch. You control per shop whether and how automatic sending happens. Under Inbox in the console, you define:
- Auto-reply on/off as a global switch for autonomous sending.
- Enabled channels — auto-reply can be activated specifically for certain channels only.
- Confidence threshold — only from this value may a suggestion be sent automatically; below it, a draft is created for human review.
In addition, the send mode per channel determines what happens with a finished suggestion:
| Mode | Behavior | Default for |
|---|---|---|
Draft first (draft_first) | The AI presents a draft, and a human sends it. | |
Send directly (direct_send) | With sufficient confidence, the AI sends itself. | Webchat, WhatsApp, Instagram |
You can set the send mode globally or differentiate it per channel via the by_channel mode — for example, deliberately keeping email on "Draft first" while webchat sends directly. This way, several gates interlock before a message goes out automatically.
Quota & Abuse Protection
So that a faulty loop or a wave of incoming messages never leads to a burst of automatic replies, AI sending is rate-limited and safeguarded against abuse. If limits kick in or the system detects unusual patterns, Alpheris falls back to draft mode and hands off to your team instead of continuing to send uncontrolled. Incoming content is also checked for prompt-injection attempts so that injected instructions cannot subvert the AI.
Escalation & Handoff to Humans
The AI is built to hand off rather than guess when in doubt. A handoff to your team is triggered, among other things, when:
- the confidence is too low — the knowledge base does not sufficiently cover the question;
- the topic area is not yet on Autopilot (Shadow or Copilot);
- it is a locked topic area (e.g. complaints);
- quota or protection mechanisms kick in.
Here's how the handoff works in practice:
Inbox via API
The inbox can also be driven programmatically — for example, to mirror threads into your own helpdesk or to reply from an external system. Available operations include listing threads and sending a reply to a thread. API access is limited to the Business and Enterprise plans; authentication is via API key. You'll find endpoints and examples in the API Reference and under Custom API.
For privacy, data location (EU/Germany) and data processing when handling customer messages, see Security & Privacy.