Chat with EnderGPT lets you ask questions about your call center conversations in natural language. The AI analyzes your conversation data and returns detailed, research-backed answers. You can ask about topics like complaint trends, agent performance, customer satisfaction, and more.

Each chat session keeps a full history of your questions and answers, so you can ask follow-up questions to dig deeper into the results.

Open EnderGPT from the main navigation sidebar. The chat area has two parts:

If you have access to both regular chats and scheduled chats, the sidebar shows two top tabs:

Selecting Schedule opens the schedule list in the main area; selecting Chats brings you back to the chat workspace. If you only have access to one of the two, that single view is shown without a tab strip.

EnderGPT requires an LLM provider to be configured by your organization's administrator. If EnderGPT is not configured, the create screen shows an informational card instead of the question input. The card includes a link to the privacy notice and a button to submit a support ticket requesting activation.

The AI processes your question through several stages. As soon as you submit, an assistant processing bubble appears in the message area with a thin progress strip and the current stage. The question input stays anchored at the bottom of the chat so it does not jump while the response is being prepared.

The processing stages are:

The current stage is shown on its own at first. If the same stage continues for an extended time without changing, a rotating word (for example, Pondering..., Thinking...) is appended to that line so you can see at a glance that the system is still working — even during a long phase.

A duration counter in the assistant byline ticks up while the response is being generated. For longer research tasks (over 30 seconds), the completed response also keeps a "Researched in" badge showing how long it took.

If EnderGPT cannot finish an answer, the assistant bubble switches to a red alert icon and, instead of the answer text, shows a short technical explanation of what went wrong and how far processing got before it failed. For example:

AI provider rate limit (429) while doing research. Retrying may help.

Final answer validation failed while validating the final answer.

Unknown error (TimeoutError) while writing the final answer.

Each message combines two pieces of information:

When the failure looks temporary — a rate limit, a provider-side server error, or a connection or timeout problem — the message ends with Retrying may help. If that sentence is absent, generating the answer again is unlikely to change the outcome on its own.

Two things are worth knowing about these messages:

If you contact Ender Turing support about a failed answer, a screenshot of the failed message is usually enough to identify the failure class and the stage it happened in — you do not need to collect timestamps or technical logs.

Under a failed answer, the chat owner sees a Generate new response button that runs the question again with the same filters.

You can also simply ask the question again as a follow-up if you prefer to rephrase it.

AI responses are displayed as formatted text with support for headings, lists, tables, code blocks, and other rich content. Each response shows the timestamp and identifies the sender as either you or the AI assistant.

AI responses can include references back to your data so you can verify any claim:

When the AI quotes a transcript, the platform checks that the quoted wording closely matches what is actually in the transcript. If the match is below the confidence threshold, an amber warning icon appears next to the quote. Hovering the icon shows the message:

Quote match is below the confidence threshold. Verify this quote in the transcript.

This warning means the quote may have been rephrased or partially fabricated. The quote itself is not clickable; if the response includes a conversation link for the same source, open it and check the transcript before relying on the wording. Quotes without a warning passed the similarity check.

Once the AI has finished responding, you can type a follow-up question in the streamlined composer at the bottom of the chat. The follow-up composer is a single-line input that grows as you type, with an inline submit button next to it; it does not show filter or scope controls because follow-up questions use the same filters that were set when the chat was created.

Depending on your role, you can perform the following actions on each completed AI response:

EnderGPT is built for analyzing calls and chats. It answers from the conversations you point it at, so the quality of an answer depends more on which conversations are in scope than on how long the question is.

The filters on the create screen decide which conversations the AI can look at. The chat create screen offers the filters you already know from the Conversations view, so you can narrow by Topic, Tags, agent, team, queue, direction, duration, and the rest — in addition to the required date range.

Narrow the scope to the conversations that can actually answer your question:

The chat filter list is not identical to the one in Conversations — it also contains three filters that exist for Enders and are not supported as a conversation scope: Commented by, Reviewed by, and Review scores. They fail in two different ways: Review scores leaves the match count unable to run, while Commented by and Reviewed by match nothing at all without reporting an error — the count simply drops to 0, which looks like a scope that is merely too narrow. If the count falls to 0 the moment you add one of these three, remove the filter rather than widening the rest. Use the Commented and Reviewed filters instead, which are the conversation-level equivalents.

Use the match count next to the submit button to sanity-check the scope as you build it. A very large count usually means the sample still contains conversations unrelated to your question; a count of 0 turns amber and blocks submission until you widen the filters.

A bigger sample is not automatically a better answer. Unrelated conversations add noise the AI has to work around, so prefer a focused scope over a large one.

Filters are fixed once the chat is created. Follow-up questions always reuse the filters the chat was created with, and there is no control for changing them on an existing chat. If you need the same question over a different sample, use Duplicate (see Duplicating a Chat) — it opens the create screen pre-filled so you can adjust the filters before asking again.

Write down what you actually want to learn before you type it. What are the most common reasons customers are dissatisfied with this process? produces a far more usable answer than Analyze everything.

For analysis questions, it helps to spell out five things in the question field:

If you would rather start from a working example, use the example questions on the create screen and edit one to fit your case.

Check that the response actually addresses the question you asked, then verify anything you plan to act on:

If the answer is not useful, review the conversation scope before making the question longer — an unfocused sample is the more common cause. Remove unrelated topics or tags, tighten the date range, or make the question more specific, then duplicate the chat and ask again.

Once the AI surfaces a pattern worth exploring, drill into it with follow-up questions rather than starting a new chat, so the thread keeps its context.

For exact figures, prefer the reporting features built for counting: see Charts, Dashboard, and Funnels. EnderGPT is strongest for trends, recurring patterns, and qualitative insight across many conversations.

The chat title appears in the top bar when viewing a session. If you own the chat and have interactive Chat access, click the edit (pencil) icon next to the title to rename it. Press Enter or click Save to confirm, or press Escape to cancel.

Click the duplicate icon in the top bar or select Duplicate from the chat's context menu in the sidebar. This opens the create screen pre-filled with the original chat's first question and filters, letting you start a new chat based on the same setup.

Any user with permission to create chats can duplicate.

Only an interactive chat owner can delete a chat session. Select Delete from the chat's context menu in the sidebar and confirm the deletion. Scheduled result chats cannot be deleted.

Projects let you group related chats into folders in the sidebar.

Drag a chat from the sidebar and drop it onto a project folder. You can also drag a chat out of a project and drop it on the Recent section to unassign it.

You can only move chats that you own and can interact with, and you can only drop chats into projects that you own. Scheduled result chats cannot be moved between projects.

Click the three-dot menu on a project folder to:

Only a project owner with interactive Chat access can rename or delete a project.

The sidebar groups your chat history into two collapsible sections:

Other sidebar controls:

Multiple chips combine: a chat must match every active chip to appear. When at least one chip is on but no chats match, the Recent section shows "No chats match the active filters" instead of the regular empty-state message.

The sidebar loads your most recent chats first so the list becomes usable within a second or two, even if you have a very large chat history. You do not have to wait for every chat to load before you can start working.

Each chat row in the sidebar can show the total LLM cost accumulated by that chat next to its date (for example, Jul 14 · $0.0142). The cost is only shown for chats that have already consumed LLM usage.

You must have the Automation manage permission to see cost values. Users without this permission (including most agents and reviewers) do not see them. Reporting is always enabled and does not require an organization-level switch.

Interactive chat owners can share their chats with other users or teams so they can view the conversation.

To revoke sharing, switch back to Private and save. All recipients are removed.

Shared recipients can view the full chat history, copy message text, and copy the chat link. They cannot send new messages, rate responses, regenerate responses, rename the chat, or change sharing settings. A "read-only" indicator is shown at the bottom of the chat for shared recipients.

Users who can access the Chat area but cannot create new chats see a read-only landing page. If chats have been shared with them, those chats appear in the sidebar and can be viewed.

Interactive Chat access is granted by the Own scope in the Chats row of the role editor (or the Turing-family bypass). Creating a regular chat or project additionally requires analytics data access.

By default, each user only sees chats they own plus chats shared with them. An administrator can extend a role with the All scope for the Chats row in the role editor (under EnderGPT Chats) so that users with that role can see every regular chat in the organization — including private chats owned by other users.

This scope is not granted to any built-in role out of the box. To enable it, create a custom role with the All chat view scope ticked and assign that role to the users who should have cross-user visibility. See Roles & Permissions for the role editor walkthrough.

Cross-user viewers see other users' chats in the Recent section of the sidebar and can open them to read the full history and copy message text and the chat link. They cannot send messages, rate, regenerate, rename, delete, or change sharing on chats they do not own — the same actions remain reserved for an interactive chat owner. Ticking the All scope automatically also ticks the Own scope, so a cross-user viewer can also create and own their own chats.

Some chats appear as scheduled results, generated automatically by the system on a recurring schedule. These chats have more limited actions than regular chats.

Chat owner actions:

Read-only owners and other viewers can view the full chat history, copy message text, and copy the chat link. They cannot send follow-up questions or perform any other actions. A read-only indicator is shown for these viewers.

When a scheduled Ender Chat result finishes processing and is ready to review, all users who can see that result receive a notification.

In-app notifications are always delivered. You will see a notification in your notification list with the message "A scheduled Ender Chat result is ready to review." Clicking the notification opens the result chat directly.

Email notifications are optional. If you have email delivery enabled for app activity notifications in your notification settings, you will also receive an email with a link to the result chat. Emails follow your configured email delay setting (for example, "send after 15 minutes" if the in-app notification has not been read).

To configure email delivery for scheduled result notifications:

Each user receives at most one notification per scheduled run, even if they qualify to see the result through multiple access paths.

Notifications are not sent for scheduled runs that fail or produce results that are not visible to any users.