Asking with AI
A plain-language question, with the indicator, a chart, and the calculation.
Ask answers questions about the organisation's data in plain language. The question is written as in conversation, and the agent queries the Qlik app to answer.
Where the question is asked
There are two routes, over the same conversation.
On the Ask page, the question may address every prepared app or one app in focus. With no app in focus the agent selects the one that answers and states which it used, which also serves to locate where a figure lives.
Inside an open app, Ask sits in a side panel, already focused on that app. Every chart carries an "Ask AI about this chart" button, which sends the chart's reference along with the question, with no need to describe what is on screen.
Framing the question
Specific questions produce better answers. "Revenue by region last quarter" delimits both measure and period. "How are sales" delimits neither, and the answer reflects that indefiniteness.
Comparisons require the period to be stated. "This month against the same month last year" is interpretable. "Recently" is not.
Where the administrator registered starter questions alongside the app's queries, they appear as clickable suggestions. That is the shortest route to what a given dashboard answers well.
Refining the answer
The conversation has memory. After an answer, "and 2024 only?", "now by region" or "drop the South" refine the same analysis without restating the question, and the chart keeps its previous form unless another is requested.
Requesting another format also works: "show it as a pie", "and as a table".
When the agent asks back
Faced with a genuine ambiguity, the agent returns two or three clickable options rather than choosing in the dark. It is one question per request: once the choice is made, it executes.
Where the most reasonable reading is evident, it does not ask. It executes and records the interpretation in the calculation, which is where that reading is checked.
Filtering the dashboard through the conversation
In the side panel of an open app, requests such as "filter to 2025", "the South region only" or "clear the filters" change the app's own selections, and the screen follows. Filtering changes the dashboard; asking returns a figure. They are distinct actions, and the agent keeps them apart.
In the opposite direction, the dashboard's filters do not interfere with answers: a question is answered over the app's full data, under the usual permissions. For an answer within the filtered slice, ask for it ("considering the dashboard filters"), and the filters used appear in the calculation.
What the answer contains
The answer carries the text and the chart most suited to the indicator presented. A brain icon sits in the chart's corner and opens How it was calculated.
No part of the processing is withheld. The panel reports the app queried, the dimensions the number was grouped by, the exact expression of each measure, the filters applied, and whether the query was one validated by the administrator or assembled on the spot.
History and rating
Conversations are held for seven days and are visible only to the person who asked.
Every answer may be marked useful or poor. The rating is recorded for the workspace administrator, and that is how a wrong definition in an app's context reaches someone able to correct it.
Limits
The agent queries only the apps the administrator prepared for that purpose. Apps outside that set are not queried, even when named in the question.
The answer considers the data the user has permission to view. Users with different access receive different numbers for the same question (when the app has section access applied).
Each workspace holds a monthly question quota. The portal issues a notice before blocking.
The agent does not generate data absent from the source. Where an answer is not possible, it reports the impossibility rather than presenting an estimate.
A divergence in an answer
Open the brain icon in the chart's corner. In most cases the divergence stems from a definition, such as gross revenue in place of net, and seeing the expression that was summed settles the question without a support request.
A definition unsuited to the business is adjusted by the administrator, under Preparing Ask. The agent uses what is published and does not correct modelling.