Ask useful questions in AI Reporting and verify the answers

Updated Wednesday, October 28, 2026 · Marcel Gravelle

AI Reporting helps you explore supported support data by writing a question instead of assembling every reporting condition yourself. Use it to investigate a focused operational question, such as the distribution of recently created cases by Company and state. The useful skill is specifying the population and checking the result: a plausible-looking chart is not enough evidence for a customer commitment or a performance conclusion.

At a glance

For: support operations leads, supervisors and authorized reporting users. Applies to: Supportbench Core.

Start here: Reporting → AI Reporting

Know when to use AI Reporting · Check access and scope · Write a complete, testable question · Generate your first chart or table · Make the reporting period unambiguous · Try focused B2B operations questions · Validate the result before sharing a conclusion · Save and reuse a prompt correctly · Example: prepare a weekly account-review question · Recover from unsupported requests and empty results

Know when to use AI Reporting

Use it for	Why it fits	Use another route when
An exploratory distribution	A focused count or measure by a supported field can reveal where to investigate.	You need a dependable recurring distribution: use a tested saved view and Scheduled Data Views.
A clearly defined period comparison	An explicit event and time window makes the question testable.	You need a standardized scorecard definition: use the corresponding KPI guide and report.
A quick table to inspect a pattern	Data Table can make categories and values easier to inspect before charting.	You need a complete raw export or an auditable list of every matching record: use the supported view/export process.
A repeatable question you often ask	A saved prompt preserves useful wording for reuse.	You need saved results or historical report runs: the prompt library does not store those.

The page is titled AI Powered Reporting and is marked Preview. It supports a defined set of reporting data and relationships, not every conceivable business question. Treat the examples below as starting questions to test against your actual configuration and result.

Check access and scope

Your active role needs both the applicable case-reporting and AI Reporting permissions. The page and generation requests enforce authorized scope. Division options follow your role; All Divisions is available only where your role has unrestricted Division scope. Team and Agent choices narrow the supported result and do not grant additional access.

Scope detail	What an operator should know
Team and Agent	Team attribution depends on the dataset. For Case/Activity reporting it uses recorded case-team relationships; Agent selection relates to case ownership. Do not assume this always means the person who created an activity.
Article-only questions	Selected Team or Agent is not supported for an article-root report. Use the applicable all-team/all-agent choices and verify article eligibility.
Chat questions	Chat-only reporting requires supported unrestricted all-Division scope. A restricted-Division role or selected Division can make that question unsupported because some chats have no case Division.
Custom fields and tags	Use exact configured labels and verify the result. Only supported fields and relationships are available; not every custom field type or arbitrary combination is guaranteed.

Do not request broader permissions merely to make an unsupported question run. Choose an appropriate reporting surface or have an authorized report owner investigate the required scope. Keep unnecessary customer-identifying narrative out of prompts; the question may be used by AI planning, while the current report query and presentation remain within the application’s authorized reporting path.

Write a complete, testable question

A useful prompt names the record type, measure, date basis, period and grouping. Add a specific condition only when it helps the decision. Prefer one main question over a paragraph combining workload, sentiment, revenue, forecasting and recommendations.

Part	Example	Why it matters
Record type	Cases	Cases and activities are different units.
Measure	Count cases	A count, average duration and percentage need different calculations.
Date basis	Created in the last 30 days	Created, closed and activity dates select different populations.
Grouping	By Company and current Status	Specifies the categories to compare.
Scope	Selected Team/Division controls, plus a supported condition if needed	Makes the intended authorized population explicit.
Output	Choose Data Table in Chart Type	Prompt wording alone does not choose how the result is rendered.

A follow-on question must stand on its own

AI Reporting submits the current question and controls; it does not send the previous question and result as conversation context. Instead of “break that down further,” restate the complete request: “Count cases created in the last 30 days, grouped by Company and current Status.” Recheck filters and generate again.

Generate your first chart or table

Open Reporting → AI Reporting. Confirm the AI Powered Reporting heading and Preview badge.
Set Division, Team and Agent to the intended permitted scope. Changing Team repopulates Agent choices, so recheck Agent afterward.
Choose Timeframe, or choose Set in Prompt and write an explicit supported period. Read the time-window guidance below.
In Report question, enter one complete question under Ask a Question About Your Support Data. The field allows up to 8,000 characters, but a focused question is usually easier to verify.
Choose Chart Type. Start with Data Table when you want to inspect the returned categories and values.
Choose Generate Report. Wait for the result or read the displayed error; do not assume a failed request means zero matching work.
Inspect labels, units and categories. Compare several known records or an independently defined report before relying on the output.
If the question needs refinement, revise the full question and relevant controls, then Generate Report again.

Available presentation types are Bar Chart, Line Chart, Pie Chart, Doughnut Chart, Polar Area Chart, Radar Chart, Scatter Plot, Bubble Chart and Data Table. Choose a form that suits the result: a table for detailed inspection, bars for categories, and a line for a genuine ordered time series. A chart choice does not make an unsupported calculation available.

Make the reporting period unambiguous

Timeframe	How to interpret it
Set in Prompt	Use a concrete supported time expression and verify the resulting population. If no supported expression resolves, the current path falls back to the current calendar month.
Today	The authenticated reporting user’s local-date window.
Last 7 / 30 / 60 / 180 Days	The selected rolling day window applied by the reporting service.
Last 1 Year	The current implementation uses 360 days, not a full calendar year. Use a carefully stated and verified period when calendar-year reporting matters.

Default date bases differ by dataset: case creation, activity creation, chat start/creation and article creation. State the event you mean in the question. “How many cases closed last month?” is not the same as “How many cases created last month are now closed?” Verify that the resolved report answers the intended one.

Avoid contradictory time instructions between the prompt and the Timeframe control. Use one consistent window and inspect representative boundary records. When reporting to another region, state the date basis and reporting time zone rather than assuming everyone’s “today” begins at the same moment.

Try focused B2B operations questions

Operational question	Starting prompt	Verification before use
Which customer organizations are generating work?	Count cases created in the last 30 days, grouped by Company and current Status.	Choose Data Table and verify Company relationships, case-created dates and status values.
Which issues dominate new requests?	Count cases created in the last 30 days, grouped by Issue Type.	Check issue classification on representative cases; an unclassified case may need different handling.
How is current-owner workload represented?	Count cases created in the last 30 days, grouped by current Owner and Priority.	This is a created-case cohort, not necessarily all currently open workload. Add a supported open-state condition only after checking its meaning.
What activity mix accompanies the work?	Count activities created in the last 7 days, grouped by Activity Type.	Verify the activity date basis and authorized case population; do not call activity count a case count.
Is a configured classification useful?	Count cases created in the last 30 days, grouped by the configured case field [exact field name].	Replace the bracketed text with the actual label. Confirm the field is supported and values are interpreted correctly.
Which knowledge records should we inspect?	Show article views grouped by article title.	Use supported all-team/all-agent scope, inspect the meaning of the view measure and article eligibility, and do not assume a rolling period unless verified.

These are prompt patterns, not guarantees that every request resolves or that every account has the same fields. For surveys, response measures, chat and more complex relationships, start with one measure and validate its population. Do not infer a customer’s satisfaction from an unrelated sentiment score or a case’s age.

Validate the result before sharing a conclusion

Read what was returned. Confirm categories, units and date labels. A table answering a nearby question can still look convincing.
Check inclusion. Find a few known qualifying records and compare their relevant fields and dates with the requested scope.
Check exclusion. Confirm that an unrelated customer, excluded Division or out-of-period record is not represented as part of the intended population.
Check definitions. For durations and percentages, identify the measured event and denominator. If these cannot be established confidently, use the dedicated report with a documented definition.
Check completeness. Large results are bounded. The default row limit is 500 and the maximum is 2,000; the current page does not display the report’s truncation flag or warnings. Narrow or aggregate the question and do not describe a large returned list as a complete export.
Check the action. Use the result to select cases or questions for investigation. Record a decision only after reading the supporting evidence.

The current page presents a chart or table. It does not provide a generated-report history or an Export/Download control. Saved prompts preserve wording, not results. For a distributed or repeatable report, move the verified question into a supported saved-view, export or scheduled-report workflow.

Save and reuse a prompt correctly

With a useful full question in Report question, choose the bookmark action Save this prompt.
Enter a clear Title, such as “New cases by Company and status,” then choose Save.
Use Saved Prompts and Search prompts… to find it later. With a question in Report question, choose New to save it as a new titled prompt.
Use the saved prompt’s Use action to fill the question. This does not run a report.
Recheck Division, Team, Agent, Timeframe and Chart Type, then choose Generate Report. Prompt storage contains title and question, not those controls or a previous result.

Prompts belong to the signed-in user. Give the title enough meaning to recognize the question later, and update stale wording when your operational definitions change. A saved prompt is a reusable starting point, not a certified report definition.

Example: prepare a weekly account-review question

An operations lead wants to understand new support demand across customer organizations before choosing accounts for follow-up. They select the permitted team, choose Last 30 Days, ask for cases created by Company and current Status, and use Data Table.

They compare several known cases with the table and check that transferred or closed cases are interpreted according to the question’s current fields and created-date cohort.
They notice one Company with several cases and open its actual case history. The table does not prove all cases have one cause.
They review customer impact, ownership and commitments, then use a focused case view for the account meeting.
They save the full prompt for the next exploratory review, while documenting the filters and date meaning separately.
If a repeat distribution is needed, they build and validate a scheduled data view with explicit recipients rather than assuming the AI result is saved or emailed automatically.

Recover from unsupported requests and empty results

What you see	What to do
This specific report request is not yet supported or the right data could not be determined	Simplify to one record type, one measure, an explicit period and one main grouping. Use exact supported field names and inspect dataset scope limits.
AI Reporting is temporarily unavailable. Please retry.	Keep the full question and controls, then retry after the temporary issue. Repeated failure should be reported with non-secret question/scope and the displayed message.
No data to display in table format	Check Division, Team, Agent and Timeframe, then verify whether qualifying records exist. Do not assume the entire account has no activity.
A chart shows zero values	Check whether real zero buckets are appropriate for the selected scope and period. Distinguish a valid zero from a request failure.
Article-only reporting fails with Team/Agent selected	Use the supported all-team/all-agent scope and verify the article population.
Chat reporting fails with a Division restriction	Use an authorized supported reporting path; case-less chats cannot safely be forced into a case-Division scope.
A short follow-up produces an unrelated answer	Restate the full question. Previous report context is not sent as a conversation.
The table seems incomplete	Narrow or aggregate the request and verify limits. Use a dedicated export for a required complete record list.
A saved prompt produces a different result	Compare current controls, changing records and time windows. The prompt library does not restore an earlier result or filter state.

Was this article helpful?