Configure the AI Agent

Where to configure this

You'll find the chatbot configuration in the dashboard under AI (console.alpheris.com → AI). There you set the persona, tone, answer rules, and handoff behavior. For the AI inbox (email and messaging) there's an additional, dedicated agent configuration under Email → AI Agent — see Inbox & AI Replies.

Persona & Tone

The persona defines who the agent is and how it sounds. You control it via three main sliders plus a free-form system instruction.

Tone

The tone of the answers. The options are:

Form of Address

Controls whether the agent addresses customers formally or informally:

• Automatic — the agent adapts the form of address to the customer's language and style (default).
• Formal (Sie) — consistently formal address.
• Informal (Du) — consistently personal address.

Emojis

The emoji setting controls whether the agent uses emojis: Automatic (context-dependent, sparing), On, or Off. For serious B2B presences, Off is the usual choice.

System Instruction

In the free-form system instruction, you describe the persona and fixed behavior rules in your own words. It complements the sliders above and applies to every answer.

You are the customer service assistant for [Company Name].
Answer in a friendly, matter-of-fact way, in short, clear sentences.
For product questions, recommend suitable items from the range.
If you can't find something in the knowledge base, honestly say
that the information is missing – never guess.
Don't promise delivery dates, discounts, or refunds.

Answer Rules

Answer Length

Choose between Short, Balanced, and Detailed. Shorter answers usually work better in web chat, more detailed ones in email support. The length can additionally differ per channel (see channel profiles further down).

Prohibited Statements

In the Prohibited Statements field, you enter things the agent must never claim or promise — such as guaranteed delivery dates, medical efficacy claims, or statements about competitors. These rules take effect in addition to the system instruction.

Clarifying Questions & Sources

• Ask a clarifying question: If the request is ambiguous, the agent asks a targeted follow-up instead of guessing.
• Show sources: Optionally, the agent can link the knowledge sources it used beneath the answer — useful for support-heavy sites, often distracting in sales conversations.
• Conversation length: You can limit the maximum number of bot replies per conversation before a handoff is offered automatically.

Grounding — only from the knowledge base

Every answer is produced by hybrid, grounded retrieval: Alpheris combines full-text search and semantic vector search over your knowledge base and ranks the hits by relevance through reranking. Only the passages found this way serve as the basis for the answer.

You can control precisely which sources feed in. In the AI configuration, the source groups can be enabled or disabled individually:

• Website crawl of your domain
• Knowledge base (FAQ pairs and uploaded files)
• Shop data (products and orders, if connected)

Fallback Behavior

You define what happens when the agent can't find a supported answer in the fallback:

Maturity: shadow → copilot → autopilot

In the AI inbox, the agent learns in stages instead of answering on its own right away. Per topic cluster (e.g. "Shipping", "Returns", "Invoicing") it goes through three phases — from silently reading along to automatic sending. This way you build trust before the agent sends messages itself.

The move from one phase to the next doesn't happen across the board, but per cluster, and it's tied to measurable thresholds. A cluster only qualifies for promotion once enough cases exist (at least five, and no older than 14 days). For Autopilot, two hard conditions are added:

• at least 50 accepted suggestions in the cluster, and
• an acceptance rate of ≥ 75%.

In addition, the answer confidence must be above the cluster's threshold. The default threshold is 0.85; the more mature a cluster, the lower it may go (warm ≈ 0.78, fully mature ≈ 0.72). If confidence is below that, the suggestion stays in copilot mode for manual approval.

Channel Profiles

Tone and length adapt to the channel the message comes in through. The default profiles are:

The global persona from the AI configuration is preserved; the channel profiles only adjust tone and answer length to suit the medium. Which channels are actually available is listed under Inbox & AI Replies.

Escalation & Handoff

You can trigger a handoff to a human in several ways:

• On request: The customer types a keyword like "agent" or "human". You enter the triggering words as handoff keywords.
• On low confidence: If the answer confidence is below the escalation threshold, the agent offers a handoff instead of an uncertain answer. The default value is around 0.68.
• Via the fallback: In the Escalate or Form fallback mode, a handoff happens directly when evidence is missing.

You define the target channel for the handoff (e.g. email or ticket). Through the escalation form, you also collect contact details — optionally as a ticket, a contact request, or a lead. The form's title, subtitle, and button label are freely customizable.

Unfortunately I don't have confirmed information on that.
Would you like to be connected to a team member?
Our team is available Mon–Fri from 9 am to 6 pm.

Test before going live