Treat the chat platform as separate capabilities, not one integration. The agent replies in conversations routed to it, reads the visitor's context, hands over to people with a note, and closes chats only when the customer confirms. It cannot read other conversations, delete messages, block visitors or export transcripts. A tested kill switch sends every chat to people.
What does the connection involve, and who owns it?
A live chat platform routes website and app conversations to queues, shows advisers the visitor's context, stores transcripts, and offers functions such as tags, canned replies, file sharing, proactive invitations and visitor blocking. Connecting an AI agent places it in that routing, usually as the first responder, with people behind it.
The digital or support systems team owns the platform. Security owns the agent's identity. Support operations owns queues and handover rules. Compliance owns disclosure wording and transcript retention. The agent takes part on terms set by all four, in line with our guide to the AI support agent that knows when to say no.
What are the steps, and at what depth?
- Person, systems team: registers the agent as a bot participant with its own identity and a named owner, not as an adviser seat.
- Person, support operations: defines routing. Which entry points, pages, languages and hours go to the agent first, and which go straight to people.
- Agent, depth 1: reads visitor context: page, logged-in status, and, after authentication, the customer reference.
- Agent, depth 3: replies in conversations routed to it, opening with the approved disclosure that it is an AI and how to reach a person.
- Agent, depth 3: tags conversations from your approved tag list.
- Agent, depth 3: hands over to a named queue with a note, using the platform's transfer function. The mechanics are in how to let an AI agent hand a chat to a person in your live chat tool.
- Agent, depth 3: closes chats only after the customer confirms the matter is resolved, or after a set period of silence with a closing message.
- Person, support operations: keeps proactive invitations. Starting conversations with visitors who did not ask begins as a held capability, approved by campaign.
- Person, systems team: leaves admin functions out of the role. Reading other conversations, editing or deleting messages, blocking visitors, exporting transcripts and changing routing are not granted.
- Person, security: tests the kill switch. Turning the agent off routes every new and open chat to people or an offline form.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Context | Reads page and login state | 1 | Nothing |
| Reply | Answers routed chats | 3 | Disclosure wording |
| Tag | Applies approved tags | 3 | Tag list |
| Hand over | Transfers with note | 3 | Queue rules |
| Close | Closes confirmed chats | 3 | Nothing |
| Proactive invite | Proposes campaigns | 2 | Approving them |
| Admin | Nothing | none | Delete, block, export, routing |
| Kill switch | Nothing | none | Triggering and testing |
Which controls matter, and what evidence does each leave?
Bot identity, not an adviser seat. Every message the agent sends is attributed to it. Evidence: platform transcript authorship.
No cross-conversation access. The agent reads only the conversation it is in. Evidence: access logs scoped to conversation IDs.
Disclosure at the start. Evidence: the disclosure message in each transcript.
Transcripts cannot be altered by the agent. Editing and deleting messages are outside its role. At the platform level, altering the filed record of a verdict is withheld for every agent at every depth; platform floor: what it means for AI agents explains that idea. Evidence: refused verdicts for any attempt.
Kill switch tested. Evidence: dated test results and the time chats took to reroute.
What should you measure?
- Chats resolved by the agent, against chats handed over, by entry point.
- Wait after handover until a person replies. A long wait undoes the agent's speed.
- Customers asking for a person and time to transfer.
- Accuracy on sampled transcripts, compared with the level you set before going live. How accurate does an AI support agent need to be before going live? covers setting that bar.
- Kill switch tests and real activations.
Where does it usually go wrong?
The agent is set up as an adviser. It then sees all queues, can transfer anywhere, and may be counted in staffing reports. Use the bot participant model.
Handover happens with nobody available. Out of hours, the agent transfers to an empty queue. Routing must know when people are there and offer an alternative when they are not.
Context is lost between channels. A customer who started in chat calls in later and has to repeat everything; see can one AI support agent work across email, chat and phone?
Proactive chat becomes nagging. Invitations on every page irritate customers and fill the queue with idle chats. Keep campaigns under approval.
Uploads are accepted without a plan. Customers send photos, statements and identity documents in chat. Unless file handling is a granted capability with its own retention rules, the agent should not request or process them. Files that arrive anyway are external content and go to a person.
The agent closes chats to look efficient. Measure resolution by whether the customer returned, not by closure.
Where Pryme fits
In Pryme, each chat capability is a separate entry in the agent's role with its own depth. Disclosure, handover and close are certified on your own transcripts through Attest. Handover itself, from the customer's side, is covered in how to hand a live chat over to a person with an AI agent.
Questions people ask
Can the agent and an adviser both be in the same chat?
Yes, if the platform supports it, as an assist mode. The agent then suggests replies to the adviser at depth 2 and does not speak to the customer.
Should the agent see the customer's previous chat transcripts?
After authentication, summaries of recent ones are usually enough. Full transcripts may contain more personal data than the current question needs.
What happens to chats in progress if the agent is switched off?
They should transfer to people with their transcript, not end. Test this as part of the kill switch.