AI agent settings: setup, profiles, voice and messages

These settings are part of Cliedesk Pro. Open Settings → AI → AI agent. You need the Manage AI permission (Admins have it). Every change is saved as a settings version, so you can see who changed what and go back.

Setup

The Setup tab walks you through 5 steps: Knowledge, Behaviour, Actions, Test and Deploy. Results shows the agent’s customer turns of the last 30 days.

• Before you start lists what the agent needs: an AI provider, the AI agent switched on, a service user and help content. Each missing item has a link that opens the screen where you fix it.
• Read my site reads your site name and description. On a WooCommerce shop it also reads the currency, your shipping zones and the return window from your refund and returns page. It fills a draft business profile, proposes the Friendly tone and the Ecommerce profile, and writes draft starter policies (returns, shipping, escalation).
• Check the business profile and click Save business profile.
• Starter policies are drafts. They never reach customers until you switch one to Active and click Save starter policies. A new Read my site never changes an active policy.
• Setup never takes the agent live. It stays in shadow mode until you change the channel in Settings → AI → Shadow and rollout. Try questions first in Settings → AI → Test lab.

General

The General tab holds the agent’s main settings: name, service user, limits per turn, the grounding threshold, customer memory and the fixed English messages. Choose a Brand to change a value for one brand only. Values you do not change for the brand come from the default.

Profiles

A profile is a goal and a set of tools for the one agent. Cliedesk has 4 built-in profiles: Service, Ecommerce, Sales and Onboarding. You can create your own, for example “Billing”.

For each profile you set:

• the goal and the success metric,
• the knowledge sources it may read (none ticked = all sources),
• the actions it may use (none ticked = the default look-up tools). Money and other changes still wait for a person’s approval, whatever the profile allows,
• the policies and playbooks it follows (empty = all),
• a tone (or the brand’s tone),
• the intents that switch a conversation to it. Example: when triage finds the intent “refund”, a conversation moves from Service to Ecommerce. Each switch is in the audit log and in the turn’s trace,
• its own limits: steps and tool calls per turn, tokens per turn, the grounding threshold and a monthly AI budget. When the budget is used up, the agent hands the conversation to a person.

Click Clone to copy a profile in one click. Built-in profiles cannot be deleted; switch them off instead.

Scope rules

Scope rules decide which profile a conversation gets. A rule matches when any of its conditions is true. Rules are checked from top to bottom, and the first match wins. If no rule matches, the conversation gets the Default profile.

Conditions can use the channel, the widget or form, brand, product, organization, language, identity level, page address, cart value, WordPress role, segment, email domain and contact fields. A rule can also limit the knowledge sources (for example: the widget on your pricing page reads only files) and set a tone for that audience.

Rules run when a conversation starts and again when the customer’s identity changes, for example when a guest logs in.

Voice

Choose All brands or one brand.

• Identity: the agent’s name, avatar and a short bio. The widget shows the name and avatar, and emails use the name as the sender (“Mia at Shop”).
• AI disclosure: on by default. In chat the first AI reply of a conversation ends with the label (for example “AI assistant”). Emails get a short footer. To switch it off you must confirm a compliance warning: laws such as the EU AI Act require this notice.
• Tone: Friendly, Neutral, Professional, Matter-of-fact or Playful, with a different tone per channel if you want. A tone set on a channel wins over the preset.
• Custom persona: your own description of how the agent talks. Above 3,000 characters you see a warning, but it is saved. Only the first 6,000 characters are used.
• Answer length per channel: Concise, Balanced or Detailed, a short channel instruction and No bullet lists. Chat is Concise and email is Balanced by default.
• Formality per language: one line per language, for example de: formal. Mirror answers in the same form of address the customer uses (du or Sie, tu or vous).
• Words: a greeting and a sign-off for emails (tokens: {first_name}, {name}, {agent_name}, {brand_name}), a glossary (“subscription: plan” replaces “plan” with “subscription”), forbidden words and phrases, and the emoji policy (allowed, off, or only some emoji).

The Live preview shows the same sample answer with your unsaved changes. The voice also shapes copilot drafts and rewrites, and formality is used for AI translations.

Cliedesk checks every AI reply before it is sent: forbidden phrases are removed, glossary words are replaced and emoji are removed when they are off. The sign-off is placed before the signature.

Messages

The agent has 14 fixed messages: welcome, cannot find an answer, handoff, handoff failed, guardrail topic, unsupported language, technical error, after hours, verification request, chat continues by email, offer a person, offer a person when asked, and the two buttons Talk to a person and Keep chatting with AI. They come in 21 languages, and the customer gets them in the conversation’s language. When you set your own text for the chat widget’s Talk to a person button, that text is used in every language. Choose a brand and a language, edit a text and click Save messages. An edited message is sent exactly as you wrote it. Back to the default removes your text.

Rewrite in our tone asks the AI for a new version in your brand’s tone. Nothing changes until you click Use this text and save.

Brand voice

Learn our voice reads your team’s best replies: by default replies of the last 90 days in conversations rated 4 or better that were not reopened. The AI writes a style card: sentence length, greetings, empathy phrases and typical words. Edit the card and click Approve as policy. Only an approved card guides the agent and the copilot. Every 3 months a new proposal is made; it changes nothing until you approve it.

AI memory on the contact page

When customer memory is on (General → Customer memory), the contact page shows an AI memory panel with the short facts the AI keeps about the customer. View and edit opens a page where people who may edit contacts add, change or delete facts.

Channels

The same answer looks different per channel: buttons and source cards in chat, a full reply with greeting, sources and signature in email, plain text within the length limit on SMS, WhatsApp and social channels, and spoken text without links or lists for voice.