Analytics Agents¶

Status: Current | Last reviewed: 2026-10-08

Analytics Agents let you explore transcripts through a branded chat, exact conversation counts, prepared findings, trends and source evidence. Your administrator must enable Analytics with FEATURE_ANALYTICS_AGENTS=true before it appears. It is off by default in each deployment, independently of Scene agents. Disabling it hides its creation option, filters and saved entries, and makes Analytics links unavailable without deleting saved data. Everyone with read access in the tenant can use published agents. Builders and Admins can edit them.

Open Agents → Filters → Type → Analytics Agent to find Analytics agents. Select a card to open its maximized preview; the back arrow returns to the filtered library. Builders can use the card's Edit action to open the canvas. There is no separate Analytics navigation item.

Create and edit¶

1. Open Agents → New Agent → Analytics Agent, choose Type → Call Drivers, and enter a name. Call Drivers is the only available type.
2. Select any of the four preset canvas nodes: Data, Preparation, Analysis, or Experience, or click its pencil icon. Its settings open in the left sidebar. Each card describes its purpose and shows a summary, such as the selected dataset or analyst model.
3. In Data, choose a dataset and select Manage dataset, or select Create dataset. Mark synthetic demonstrations separately from real data.
4. Configure preparation, upload sources and wait for a ready data version.
5. Select Preview in the header to open the saved draft maximized. Use Minimize preview to place it beside the editor, or close it to edit.
6. Select Publish, optionally add version notes, and publish the shared draft.

Edits autosave through Studio collaboration. Presence and save status appear in the header. Preview and Publish wait for the shared draft to reach the server. If the connection is unavailable, edits remain on screen; reconnect before closing the page. Click the agent name or Settings to change its name and description, view the Type dropdown or inspect published versions. The same Settings dialog contains Tags and Save tags. Tags apply across versions and save separately from the Analytics draft.

Call Drivers analyzes why customers contact you, trends, root causes and conversation evidence. Existing Analytics agents use this type automatically. The type determines which canvas steps and configuration panels are available.

The header keeps Settings, Preview and Publish on the right. The gray canvas uses the same zoom and Fit view controls as voice agents. Nodes stay in a fixed order and cannot be dragged or rearranged. You can pan and zoom the canvas. Close the settings sidebar, or click empty canvas, to make more room. Its header stays in place as you scroll settings. On smaller screens, settings overlay the left side of the canvas; close them to return to the canvas.

Manage datasets¶

Data → Manage dataset opens the dataset manager:

See Datasets and intent hierarchies for operational CSV/JSON imports, version carry-forward and the searchable tree editor.

Keep existing files and add / replace selected files copies retained sources into a new build. Select an existing file before uploading a replacement with the same filename. To remove selected files, build without replacement uploads. Replace all files builds a version from a complete new upload. Older versions and existing chats keep their original data. At least one file must remain.

Limits are 10,000 files, 100 MiB per file and 2 GiB for the resulting batch, including retained files. JSON/JSONL, Contact Lens/CCIQ JSON, CSV, Netomi XLSX, WAV, MP3, M4A, FLAC, OGG and WebM are supported. Each build must have unique conversation IDs across its files. If source retention was disabled, or an imported version has no original files, upload a complete replacement.

If an upload fails or you reload, reselect its original files. Resume upload skips completed files. Retry incomplete stages keeps successful compiler work. Cancel preparation prevents the version becoming available. Provider request budgets remain consumed across retries; rebuild to change settings.

JSON and CSV¶

Download the example matching JSON format. Studio JSON uses this structure; JSONL contains one conversation object per line:

[{"id":"example-1","channel":"chat","startedAt":"2026-10-01T12:00:00Z",
  "turns":[{"role":"customer","text":"How do refunds work?","startMs":0},
           {"role":"human","text":"I can explain the policy.","startMs":1200}]}]

CSV uses one turn per row. Map conversation ID and transcript text columns; optionally map speaker role, conversation start, channel and turn offset. Rows with the same conversation ID are grouped in file order. Use ISO 8601 dates, milliseconds for offsets, and quoted fields for commas or multiline text. Download CSV example & use mapping to start with working columns. Clear an optional column mapping when that column is absent. Map source role labels such as agent to Agent; unmapped speakers remain Unknown.

Missing dates remain unknown. Contact Lens exports retain source IDs, channel and declared start time. Netomi XLSX evaluation labels and summaries do not become newly compiled model findings.

Audio¶

Enable Audio transcription and automated PII redaction before uploading audio. Choose service provider, service, integration, then supported model and options. Language labels show full names. Advanced · speaker roles maps provider speaker labels to Customer, Agent, Bot, System or Unknown.

One audio file becomes one conversation, identified by its filename. Audio alone does not establish a conversation date or identify which speaker is the customer. Files must be in a format and language supported by the chosen provider; provider credentials, regional access and storage requirements still apply. Only prepared, redacted text reaches search and annotation.

Preparation capabilities¶

Enable and configure capabilities individually:

PII services are Amazon Comprehend, Google Sensitive Data Protection and Azure AI Language. Entity labels are human readable and map to each service’s native values. All detected entity types uses the detector’s defaults. Tenant policy can require stronger redaction or additional types. Existing versions using Studio’s structured rules retain that mode; choose a cloud detector for names and addresses. Already redacted uploads require an explicit attestation.

Annotation and embedding models can use OpenAI, Amazon Bedrock or Google Vertex AI integrations. Generation model lists load from the integration; embedding choices list implemented models and supported vector sizes. Vector dimensions cannot be entered arbitrarily. Semantic queries always use the model and size saved in the selected data version.

Drivers, intent and other vocabulary¶

Drivers describe the issue that generated contact. Root causes belong to a driver and describe why that issue occurred. Intent describes what the customer wants to accomplish and can have category/subcategory parents. For example, a billing driver might have a duplicate-charge root cause while the customer’s intent is to request a refund.

Edit vocabulary & intent hierarchy opens a wide modal with separate views for drivers/root causes, intent, friction and escalation. The intent tree supports search, branch moves, collapse/expand and import/export for up to 10,000 entries. Friction describes obstacles; escalation describes triggers for additional assistance. Stable IDs keep classifications consistent across builds. Unknown remains a valid outcome when evidence is insufficient.

Prepared insights¶

This replaces the label “Offline reports.” It means work performed in the background so chat can reuse it: exact aggregates, interval summaries, larger trend windows and dataset-wide driver/executive summaries.

Configure each hierarchy level’s duration and which smaller level feeds it. Parents must be exact multiples of child windows. Child coverage controls how much completed input is required. Counts always come from the full population; missing input does not mean zero activity. Written trend and dataset-summary prompts and models are independently selectable. Without a narrative model, exact aggregates still compile.

Advanced settings contain build bounds, evidence samples and cost limits. The preparation request budget limits provider calls for a data build; chat has separate answer budgets. Advanced search settings select English word matching or multilingual exact matching and searchable speaker roles. Original prepared turns remain available when inspecting a conversation.

Changing preparation requires a new data version. Rebuild from retained sources in Data versions, then preview or publish that data selection. Changing chat models, prompts or branding does not rebuild transcripts.

Analysis and prompts¶

Select the Live analyst provider, service, integration and model. Analyst prompt exposes the full task prompt. Enable tools using capability cards: charts, prepared insights, anomalies, intent drift and scenario estimates. Warehouse queries, transcript retrieval and citations are always available. Composition labels, anomaly sensitivity and scenario assumptions use visual controls. Evidence retrieval can be enhanced with semantic or hybrid search.

Answer review checks draft claims against evidence and can use its own model and full prompt. Opening briefing has its own prompt, optional separate model and refresh schedule. Basic transcript mode has separately editable search planning and grounded-answer prompts/models. Reset a prompt to restore the application default. Authorized data scope, tool contracts and output validation remain enforced around these task prompts.

The current date is always today in UTC. Historical demo dates no longer change “today” or “last week.” Ask for explicit historical dates when comparing old demos. Unknown dates remain unknown. An opening briefing may use the latest completed report period and labels that period.

Vanguard’s existing demos retain GPT-5.4 with low reasoning for chat and GPT-4.1 mini for newly generated annotations and report narratives. Existing OpenAI embeddings remain text-embedding-3-small, 512 dimensions. Imported CallSight annotations and reports keep their original outputs and provenance. The conversation and briefing defaults derive from CallSight; the annotation prompt uses Studio’s structured evidence contract.

Experience and the Analytics app¶

Experience applies an existing tenant branding template or agent-specific branding: logo, icon, primary color, background and chart palette. Applying a template copies its values into the draft; later template changes do not alter published agents. Configure the welcome and starter questions here.

Open Agents, then Filters → Type → Analytics Agent. Opening a published card starts its maximized preview. Edit opens the canvas. Draft cards open saved draft previews for builders. Minimize preview and Maximize preview retain the current conversation. Copy preview link gives someone a URL that opens directly into the same published or draft mode; draft links require builder access. Each user retains their own chat history.

Use Back to Analytics agents to return to the filtered library. The maximized header includes tenant and account controls; switching tenants clears the current chat and checks access in the selected tenant. Studio SSO is reused. Older Analytics bookmarks continue to work, and there is no separate Analytics sidebar entry.

The opening briefing uses configurable cards, prompts and a refresh interval; secondary findings can start collapsed or expanded. See answers and opening briefings. Responses compose selected charts, metrics, tables, reports and contact cards. Show details expands report sections. Evidence opens a five-contact picker; Show more expands it. Selecting a contact opens the transcript drawer without losing the conversation. Recordings or business facts that are absent from the source are not invented.

If a briefing fails, select Retry briefing. Check its model access and ready prepared data if it fails again. Your chats is private history; New chat uses the currently published configuration and data selection. Existing chats retain their versions, subject to current tenant membership and data access. Legacy selected-user audience lists no longer restrict tenant readers.