Analyze conversations

In Harriet Desktop and cloud workspaces, Harriet can analyze every conversation on a website widget or public Slack or Teams channel you manage. Ask questions like “What is Harriet failing to answer on our support widget?”, “What do customers ask for most?” or “Are employees using the internal widget the way we intended?” Harriet starts from a summary, then downloads the transcripts into your workspace and works through them, even when there are thousands.

Who can use it

People with the Manage embeds permission can analyze the channels they can configure. Website widgets also need the website widget feature on your plan. Private Slack channels can only be analyzed by account owners. A Teams channel you have configured is included for anyone with Manage embeds: Harriet cannot check whether that channel is private or who its members are. Private chats with Harriet, such as web chat and direct messages, are not included.

Enable the skill

Channel conversation analytics is off by default. An admin assigns the skill to an Endpoint AI profile or to individual people under Skills. Assigning it does not give everyone access: each person still needs Manage embeds for the channel they ask about.

The skill works in Harriet Desktop and cloud workspaces, including for collaborators and scheduled runs in a workspace. It is not available in web chat, Slack, Teams or website widgets.

Ask a question

In Harriet Desktop or a cloud workspace, name the widget or channel and what you want to know. Harriet first shows how conversations ended (answered, left without an answer, passed to your team, or no question asked), recent examples of what people needed, common topics, thumbs up and down, and how many people used the channel. That summary often answers the question by itself.

For a closer look, Harriet exports just the transcripts the question needs. For example: only conversations that ended without an answer, ones with a thumbs down, visitors who clicked New case, or only identified visitors. For a large channel, Harriet can take a random sample. Dates are inclusive. The usual window is 30 days, and a single export covers at most 366 days and 50,000 conversations. If that would take too long or the file would be too large, Harriet keeps the conversations it has already written and tells you the export was shortened.

The export is downloaded into a folder in your workspace. Harriet then reads it in batches, often handing batches to helpers that work in parallel, and combines what it finds into themes, counts and example conversations.

What an export contains

Conversations are grouped the same way as in your widget results: one topic per conversation, with a new conversation after a break of 4 hours or more. A conversation appears a few hours after it finishes. If your organization sets a message retention period, older conversations are not exported. Stats and exports also exclude conversations that began before a person cleared their chat history, including conversations that span that reset.

Privacy and security

Transcripts are verbatim and can contain personal information. Signed-in widget chats run with each employee's own permissions, so they can include HR or pay details that Harriet showed that employee. Give Manage embeds only to people who should see that.

💡

For charts rather than transcripts, see See how the widget is helping on the Website widget page.