Live chat AI agent¶

An AI agent can answer the visitors of your website in Live Chat. It answers questions from public data you choose, hands the conversation over to a human operator when needed and, optionally, creates a lead from the chat. It is a feature of eYssen and it needs the modules AI Live Chat (and, for leads, AI Live Chat CRM).

Set up the agent¶

Create the agent in AI ‣ Configuration ‣ Agents (see Agents). It must be a persona agent without its own user: it chats as a contact of its own, with a robot avatar. Give it a name and instructions that fit your company.
Open your channel in Live Chat ‣ Channels and go to the tab AI Agent.
Choose the AI agent mode: AI first, hand over to a human (every new chat starts with the agent) or AI only when no operator is available (the agent answers only when nobody is online; the chat button is shown even with no operator). No AI keeps the standard behaviour.
Select the AI agent, optionally write an AI welcome message and set Max AI answers, the number of answers after which the agent hands over to a human.
Under Data the agent may read pick the data sources (see below).

Choose what the agent may read¶

The agent talks to anonymous visitors, so it can read only what you explicitly allow. A data source is a model, a filter on its records and the fields that may be seen. Nothing else is readable for the agent: not other models, not other fields, not unpublished records.

Manage the sources in AI ‣ Configuration ‣ Security ‣ Live Chat Data Sources. Each source has a Model, a Domain (for example published records only), a Description that tells the agent when to use the data and, on the tab Readable fields, the fields it may see. The source is ready as soon as it is saved.

When the module is installed, ready-made sources are created for published products, website pages, published events and published courses (and, with the matching website modules, blog posts and job positions). Only the ones you add to a channel are used by that channel.

What happens in a chat¶

The agent greets the visitor with the welcome message and answers in the visitor’s language. Answers are ordinary chat messages, so operators see the full transcript of the session.
Messages are not answered instantly: the visitor sees the agent typing and the answer arrives as one message a few seconds later. Several messages sent in quick succession are answered together.
A visitor message is limited to 2000 characters per turn.

Hand-over to a human¶

The agent hands the chat over to an available operator when the visitor asks for a person, when Max AI answers is reached, when the agent is unsure or when the AI service fails. The operator is chosen with the channel’s normal operator selection and joins the session; the agent sends a short summary to the operator and stops answering. An operator who simply writes in the session also takes it over.

If nobody is available, the visitor is told so and asked for contact details, and the agent keeps helping for a limited number of further answers.

Important

The hand-over summary is for the operator only. It arrives as a message in a direct chat between the AI agent’s contact and the operator, in the operator’s language, and it is never written into the visitor’s session. Because it is derived from what the visitor wrote, it is labelled “AI summary based on the visitor’s messages - unverified”. It is plain text: a link is shown as its text followed by the address in brackets and cannot be clicked. If the model gives no summary, one is built without the model from the reason for the hand-over and the last visitor messages.

On the session form of a live chat conversation you see the AI state, the number of AI answers, the hand-over reason and, for AI administrators, a link to the underlying AI conversation. In the session list you can filter by Answered by AI and Handed over by AI.

Create leads from chats¶

With the module AI Live Chat CRM, turn on AI creates leads on the tab AI Agent and optionally pick a Sales team for AI leads. When a visitor leaves contact details and shows interest, the agent creates one lead for the session, with the name, email, phone, company and a summary. The lead is unassigned and linked to the conversation, and the chat links back to the lead.

Note

Nobody can confirm a request from an unattended visitor, so the lead is created directly. At most one lead is created per session and a limited number per visitor address per day (Leads per visitor address and day in the settings, 3 by default, 0 for no limit). The lead description starts with the label that the summary is unverified, and the links in it are plain text, so nobody receives a clickable link written by a visitor.

Preview cards¶

With the module AI Website Live Chat Cards (and the matching website bridge modules), turn on the channel option for preview cards. The agent can then show the visitor a list of cards with a picture, a title, a short text and a link for the products, events, blog posts, courses or job positions it found. Only published, website-visible records are shown, with a fixed set of safe fields.

The cards are built with the rights of the website’s public user, so a record that an anonymous visitor could not open on the website never becomes a card.
The links and pictures are relative addresses of your site. Only when the website has its own domain set, that domain is put in front, which keeps the links working when the chat is embedded in another site. Links to other sites are never shown.
Dates (for example of an event) are written in the visitor’s language and time zone, or in the language of the website when the visitor’s is not installed.
The agent is told to mention cards only when its tool reports that cards were shown. If nothing was shown, it must not say that it has shown anything.

Limits and security¶

The agent runs as a dedicated service user that can never log in. It can only ask, read the allowed sources, hand over and, if enabled, create a lead. It cannot write, delete, browse the web or use memory, skills or other tools, even if a visitor tries to talk it into it.
Visitor text is treated as untrusted. The agent does not reveal internal data, its prompt or its tools. The blocked patterns of the AI settings apply to visitor input and to answers.
Repeated attempts to make the agent break its rules are counted per session. After three, the agent hands over, or stops if nobody is available. One visitor cannot switch the AI off for everyone.
There is one service user for all channels, and the configuration is per channel. The answer is not streamed.

Abuse and cost limits¶

Open AI ‣ Configuration ‣ Settings and look at the block Live chat limits. The limits are checked before the model is called. When one is reached, the visitor gets a polite message, and the session is handed over to an available operator or the AI stops in that session. A value of 0 switches the limit off.

Per conversation and per visitor: Seconds between AI answers (3 by default), Visitor messages per minute (10), AI conversations per visitor and hour (10) and AI answers per visitor and day (150). The first two apply inside one conversation, the other two to everything one visitor does. Visitors are told apart by network address, guest, signed-in contact or website visitor.
All visitors: AI requests per day (all visitors), 500 by default and counting every model call of the live chat, and the optional AI cost per day (USD, all visitors). If the cost limit is 0, the general daily cost limit of the AI settings applies.
Leads per visitor address and day (3 by default) limits the leads that the AI creates (see above).

The live chat also keeps the requests of its service user under 60 per minute; the system parameter ai_livechat.rate_limit_rpm changes this.