How to Ask Better Questions in Outlined's Analytics Chat

Ask Outlined's internal analytics chat a specific question with a defined unit, date range, and scope. Then inspect what it actually returned before making a decision. A confident natural-language answer can still reflect an unexpected filter, missing data, or an unsuitable counting rule.

The application labels this area Data Query, available from the dashboard navigation when your permissions allow it. It is an internal natural-language data interface, not a chatbot replying to customers.

Start with one business question

Choose a decision you can describe without software terminology: which complaint categories deserve review, which product questions recur, or how workload changed between comparable periods. Avoid beginning with “Tell me everything about our customers.”

Specify whether you want messages, conversations, customers, formal tickets, or another supported measure. Those are different units. If the tool cannot establish the requested measure, narrow the question or use an appropriate report instead.

Build a precise question

Not every combination is necessarily supported. The examples below are prompts to evaluate, not published test results or guaranteed outputs for every workspace.

Improve vague prompts

Instead of “What are our biggest problems?”, try a narrower question about formal ticket categories during an explicit period. Instead of “How did the team do?”, ask for a defined workload measure for comparable dates, then review quality separately.

A reusable prompt structure is: “For [explicit period], show [supported measure] grouped by [available dimension], limited to [scope]. Explain the unit, filters, and important limitations.” Check the resulting interpretation rather than assume the wording alone guarantees those conditions.

For product interest, distinguish requests from mentions. For complaints, distinguish tickets from all negative messages. Those concepts may use different available records and definitions.

Inspect the result and explanation

The current page can display result tables, SQL when available through “View SQL Query,” and explanatory information where returned. Check whether the shown columns, dates, sources, and filters match the question.

If you cannot assess the SQL, inspect the plain-language explanation and ask an authorised analyst to review a consequential result. Do not copy generated queries into a database console or treat displayed SQL as an instruction to run it yourself.

The ability to view a query improves inspectability; it does not prove the query is semantically correct.

Watch for date and denominator errors

A date might refer to message creation, ticket creation, completion, or an order event. Make sure the chosen timestamp fits your question. Ask whether a rate uses the appropriate denominator rather than accept an unexplained percentage.

AWS's natural-language analytics guidance discusses the importance of the date field and semantic setup in its own product. The general lesson is to inspect interpretation; AWS-specific features should not be assumed for Outlined.

Cross-check a small result

Use an available report with matching dates and definitions, or inspect a few authorised source records. If totals disagree, check scope, units, timestamps, missing data, and whether the result is limited to a displayed subset.

Zero rows can mean no matching data, an overly narrow filter, insufficient access, or an unsupported interpretation. It does not automatically mean the business event never happened.

Keep a question log

Record the business question, exact prompt, filters, date, result, interpretation, and caveat. Reuse prompts only after checking that product and data definitions remain compatible.

Use the page's feedback control where appropriate to identify an unhelpful answer. For important findings, apply the AI-insight validation workflow and bring the verified conclusion into the weekly report.