Skip to content

Black Friday:

60% off

Cliedesk Pro

Every plan, every site count. Ends Cyber Monday.

00 days 00 hrs 00 min 00 sec
Claim 60% off

AI

AI agent

This guide covers a feature of Cliedesk Pro. The free plugin is a complete help desk; see what Pro adds.

The AI agent is part of Cliedesk Pro. It answers customers on your channels from your help articles and the customer’s own data, asks a question when something is unclear, prepares actions such as refunds for your approval and hands the conversation to a person when it should. One agent serves every channel and brand.

Shadow mode first

On a new site the AI agent is in shadow mode on every channel. It writes what it would answer as a private note (an AI draft) and sends nothing to customers. Compare the drafts with your team’s replies in the AI test lab, then go live channel by channel. See AI test lab.

The agent needs:

  • the switchboard feature AI agent turned on in AI settings,
  • a service user: the agent acts with this user’s permissions. Without one it only writes drafts.

How a customer message is handled

  1. Triage fills the fields, then your rules (workflows) and SLA run.
  2. The agent checks your exclusion rules and screens the message for unsafe topics.
  3. It finds the right help articles and reads the customer’s orders and profile.
  4. It writes an answer and checks that every claim is in its sources. A weak answer is never sent.
  5. It sends the answer, asks one clarifying question, or hands off to a person.

An answer with an email address or phone number the agent was not given (for example another customer’s) is never sent. A person takes over instead.

If a rule already sent an auto-reply, the agent does not greet the customer again. A rule can also say AI agent skips this conversation or AI agent handles (choose the profile).

Safety

  • Protected topics always go to a person at once with your safe message: self-harm, threats, harm to minors and high-risk legal, medical or financial advice. They cannot be switched off. You can add your own topics.
  • Exclusion rules keep the AI out, for example for VIPs, partners, some tags, email domains or keywords. The rule “Don’t let AI answer VIPs” ships switched off. Excluded conversations show the state Excluded from AI with the rule name and go to your team as usual.
  • Actions: the agent only looks things up on its own. Refunds and other money actions, and every change to an order or account, wait for a person’s approval. The customer must be signed in or verified first.
  • Messages that try to change the AI’s instructions are logged. They cannot make the AI run an action.

Answers and sources

  • Public help articles are linked: inline in emails, as source cards in chat.
  • Information from articles that may not be shared is used without a link. The answer then says that information that is not public was also used.
  • The agent reads screenshots, photos and PDFs (up to 5 images per conversation). It copies error codes into the field AI error code.
  • When the customer means an order (“the blue mug from last week”), the agent finds it and fills AI order. If two orders match, chat shows a short list to choose from and email asks one question.

AI state and views

Each conversation the agent touches has the field AI state: AI processing, AI answered, Waiting for the customer, AI consulting a person, Handed over, Excluded from AI, AI action done, AI action failed or Resolved by AI. The views AI handling, AI waiting, Handed over by AI, Excluded from AI and AI to review use it. When a person replies, the AI stops for that conversation. Every handoff fills Handoff reason and pins a handoff note.

Customer memory

Memory is off until an admin turns it on. Then the agent keeps short facts a customer shared (for example “prefers email”) for 180 days. Agents with contact permissions can see, edit and delete them. In this version that is through the REST API (contacts/{id}/ai-memory); a screen on the contact is coming in a later version. The WordPress privacy export includes the facts and erasing the person removes them. Each brand can switch memory off.

The agent never reads other customers’ conversations when it answers. Past answers help only after you turn them into reviewed articles or saved replies.

A settings screen for the AI agent is coming in a later version.

Related: AI settings, AI test lab, AI triage.