Skip to navigation

Export conversations

View as Markdown

When you need to analyze Atom conversations outside Atomicwork, export them from the Conversations page in Analytics. The export includes the knowledge Atom cited, the forms it started and how they ended, the tickets it filed, and a full transcript that also covers activity on those tickets.

To open this page, go to Analytics > Atom > Conversations. Each row shows the Conversation, Deflection, Conversation rating, and Source.

Organization owners, organization admins, workspace admins, and workspace analysts can export conversations. Organization owners and organization admins can export across workspaces. Workspace admins and workspace analysts can export only from the workspaces they have access to.

Filter the conversations you want to export

The export includes conversations that match the Workspace and Duration filters, Search, and any filters you add, such as Channel, User segment, Deflection, Conversation rating, or Subtheme. You can filter to specific workspaces or include all workspaces you have access to. There is no maximum date range.

Each selected conversation is exported in full: every message, plus all activity on tickets created from it, up to the time of the export, whatever the date of that activity.

Export the files

  1. Go to Analytics > Atom > Conversations.
  2. Set the Workspace and Duration filters, enter any Search text, and add any other filters you need.
  3. Click Export.

Atomicwork emails you a download link when the export is ready. The link expires in 7 days. The download is a ZIP file that contains:

  • README.txt: A description of the files, how they join, and what each column means.
  • conversations.csv: One row per conversation that matches your search and filters.
  • messages.csv: One row per message in those conversations, including activity on tickets filed from them. Rows are grouped by conversation and sorted by time.

Join the two CSV files on conversation_id. Every conversation_id in messages.csv matches a row in conversations.csv.

Note

Re-exporting the same date range later can add rows, because ticket activity is included up to the time of the export.

What’s in conversations.csv

ColumnDescription
conversation_idThe conversation’s ID, in the form CONV- followed by the conversation number, for example CONV-3188676. Use this to join rows to messages.csv.
created_atWhen the conversation started, in UTC ISO-8601, for example 2026-11-18T10:15:30.123456Z.
channelWhere the conversation took place, as shown in the product: Portal, Email, Slack, Microsoft Teams, Universal Agent, API, or Atom.
workspaceThe workspace the conversation is linked to. Not applicable if there isn’t one.
subjectThe conversation subject.
themeThe classified theme, or Evaluation pending.
subthemeThe classified subtheme, or Evaluation pending.
conversation_outcomeHow the conversation ended: Resolved, Request created, Dropped, or Evaluation pending.
deflectedtrue if Atom resolved the issue without a human, false if it did not, or Evaluation pending.
judge_reasoningThe AI judge’s reasoning, or Evaluation pending.
knowledge_usedtrue if Atom cited your organization’s own knowledge (documents, connected sources, or verified answers) in a reply. Sources Atom searched but didn’t cite don’t count.
conversation_sourcesEvery source Atom cited in the conversation, listed once each in the order first cited, as Type ID: title — link. Web results start with Web. Verified answers have no link. Sources that have since been deleted show as Type ID: Deleted. Not applicable if Atom cited no sources.
form_surfacedtrue if Atom started a request, either by showing a form (a catalog item, or the general request or incident form) or by collecting the details in chat. Mentioning a catalog item without starting a request doesn’t count.
formsEach request Atom started and how it ended, one per line, as Form name: status. Status is Request created, Dropped, or Evaluation pending. Not applicable if Atom started no form.
ticket_createdtrue if at least one ticket was created from the conversation. A ticket can be created even if Atom didn’t start a form.
ticket_idEvery ticket created from the conversation, one per line, for example ITSR-25707. Tickets Atom only referred to aren’t included. Not applicable if no ticket was created.
ticket_typesThe current type of each created ticket, one per line, as Ticket ID: type, for example ITSR-25707: Service request. Types match those shown on the ticket in the product. Not applicable if no ticket was created.
conversation_ratingHelpful, Not helpful, or No response.

What’s in messages.csv

messages.csv is the full transcript of each exported conversation: every message with Atom, plus all activity on tickets filed from the conversation, including private notes and internal ticket updates. Rows are grouped by conversation and sorted by timestamp.

ColumnDescription
message_idA unique ID that stays the same across exports. Conversation messages start with CN-, ticket notes with RN-, and ticket activity with RA-.
conversation_idThe parent conversation’s ID, matching conversations.csv.
timestampThe message time, in UTC ISO-8601. The number of fractional seconds can vary.
message_originConversation for messages in the conversation with Atom, or Ticket for activity that originated on a ticket, even if the requester also saw it in the thread.
ticket_idOn ticket rows, the ticket the activity belongs to. On the row where a ticket was created, that ticket. Otherwise Not applicable.
authorThe sender’s display name. Atom, system, and service account messages use AI Agent. Not available if the name is missing.
author_typeThe kind of sender: AI Agent, System notification, Service account, Human agent, or Employee. The person who started the conversation is always Employee. Atom and other AI agents are AI Agent. A note Atom adds on the requester’s behalf counts as the requester’s (Employee).
message_visibilityPublic if the requester could see the row. Private if they couldn’t, such as private notes and internal ticket activity like SLA updates.
textThe message text, as plain text.
message_sourcesSources Atom cited in this reply, one per line, as Type ID: title — link, matching the citations shown in the product. Web results start with Web. Deleted sources show as Type ID: Deleted. Not applicable on other rows, or when nothing was cited.
form_nameThe form involved in a form event on this row: a catalog item name, General Request, or General Incident. Not applicable if the row has no form event. Not available if the catalog name is missing.
form_idThe ID of that form. Not applicable for general request or incident forms, and when the row has no form event.
form_eventOffered if Atom showed the form or started collecting details, Request created if a ticket was created from it, or Offered; Request created if both happened in one step. Offered appears on Atom’s reply. Request created and Offered; Request created appear on the ticket-created notification. When several forms are offered in one message, each gets its own line. Otherwise Not applicable.
ticket_typeThe current type of the ticket created on this row, as shown on the ticket in the product, for example Service request. Not applicable on other rows.
message_ratingOn Atom’s replies in the conversation, Helpful, Not helpful, or No response. On every other row, including ticket activity, Not applicable.

Labels used in the files

No cell is left blank. These labels have fixed meanings:

LabelMeaning
Not applicableThe column doesn’t apply to this row.
Not availableThe value couldn’t be retrieved, for example a missing name.
No responseThe requester didn’t rate.
Evaluation pendingThe conversation hasn’t been evaluated yet.

Cells that hold several values list one value per line. Labels match those shown in the product.

theme, subtheme, conversation_outcome, deflected, and judge_reasoning come from an evaluation that runs once every 24 hours. Conversations that started after the last evaluation run show Evaluation pending until the next run. A ticket can already appear in messages.csv while conversation_outcome still says Evaluation pending.

Example: reviewing a cab opt-out conversation

An employee in the IT workspace messages Atom in Slack to opt out of the office cab service. Atom shows the Cab opt-out form, the employee submits it, and ticket ITSR-25707 is created for the travel desk.

In conversations.csv you’d see channel as Slack, form_surfaced as true, forms as Cab opt-out: Request created, ticket_created as true, ticket_id as ITSR-25707, and ticket_types as ITSR-25707: Service request. After evaluation, conversation_outcome is Request created and deflected is false.

In messages.csv, rows in time order include the employee’s request (Employee, Public, Conversation), Atom offering the form (AI Agent, with form_event Offered), the request-created notification (System notification, Ticket, with form_event Request created), and later public replies and private notes from the travel desk agent.

If you export before evaluation runs, theme, subtheme, conversation_outcome, deflected, and judge_reasoning show Evaluation pending, even though messages.csv already shows the ticket.

Learn more