Skip to content

AI Glossary · Concepts

What is Conversational Analytics?

Conversational Analytics

Conversational analytics lets a reader query business data in natural language and ask follow-up questions about the results.

Also known as: conversation analytics, conversational BI

· Updated · Chain of Thought

  • 1 explainer
  • 1 episode
  • 3 sources

State what the conversation analyzes

The useful distinction is between asking about a dataset and asking a chatbot to invent an explanation. Starburst’s AIDA documentation describes a concrete implementation: it translates natural-language questions into SQL, executes them against a selected data product and supports follow-up questions (Starburst, AIDA). Its SQL queries stay within the datasets that data product exposes. The same page describes external MCP servers that extend AIDA beyond the selected data product, so check which tools are enabled before assuming the data product is the whole boundary.

The name has a second meaning. Amazon Connect uses conversational analytics for analyzing contact-center conversations, including sentiment and emerging contact themes (Amazon Connect). That is a different job from asking a data interface to calculate churn, so say which one a vendor or a requirement means.

Define the question before checking the answer

Suppose a customer-success lead asks for customer churn for Q2 2026, from April 1 up to but not including July 1, in UTC. Define churn here as the share of accounts active at the start of the quarter that have at least one cancellation event during it. Three accounts start the quarter, and none changes plan or reactivates.

Two accounts are on Pro. One has two cancellation events and the other has one. The third account, on Basic, has none.

PlanAccounts at startCancellation eventsAccounts canceled
Pro232
Basic100

Under that definition, churn is 2 of 3 accounts overall, about 66.7%. By plan it is 2 of 2 for Pro and 0 of 1 for Basic. A query that counts cancellation events instead of distinct accounts would report 3 of 2, or 150%, for Pro. An impossible rate is easy to catch; the same mistake on a larger dataset can produce a plausible 18% that is still wrong.

Test follow-ups against a reference result

BigQuery’s conversational analytics documentation describes data agents with selected knowledge sources, table and field metadata, agent instructions and verified queries for a use case (BigQuery). Starburst’s business rules similarly record how a metric is computed and which date column to use. Both give the model context; you still need to check what it produced.

When the lead follows up with “break that down by plan,” the answer must keep the same quarter, the same account-based definition and the same set of accounts the user may see. Compare the generated query and its result with an analyst’s reference query, and include repeated events, accounts with no events and dates near the quarter boundary in the test set. Ask for clarification when a word such as churn could mean customer churn or revenue churn. A fluent explanation is not the acceptance test.

Describing the conversational interface in his product, Jitender Aswani says, “I can just start asking questions.” (episode 66, 41:44).

For builders, start with a small set of approved questions and record ambiguous ones. Context graph covers the definitions around the data; entity resolution covers whether records identify the same account; data federation covers querying across sources. Do you need to centralize data for AI? develops the same episode’s architecture discussion.

Sources

  • Starburst, AIDA — Describes AIDA translating questions into SQL against a selected data product, follow-up questions, business rules, and external MCP servers that reach beyond the selected data product.
  • BigQuery: Conversational analytics overview — Describes data agents built from selected knowledge sources, custom table and field metadata, agent instructions and verified queries (formerly golden queries).
  • Amazon Connect: Customer conversational analytics — Uses the same name for contact-center analytics: sentiment, conversation characteristics and emerging contact themes in customer conversations.

Go deeper

From the conversation