The Report Builder is an AI assistant that builds and changes reports. You describe what you want in the AI panel of a report you can edit; it plans the work, makes the changes and shows you what it did. To build a report by hand, see Build a report.
Open the Report Builder
Open a report you can edit and choose Edit.
Open the AI panel on the right and pick Report Builder in the segmented row at its top. The row also offers Model AI, which works on the report's semantic model.
Describe what to build or change in the message box and send it. You can also drag an image into the box to hand it to the assistant: the hint reads Drop a screenshot or a sketch, then say where to rebuild it.
The panel remembers your last choice between the two sides; the first time it opens, it starts on Report Builder. A side you cannot use is not shown, and if only one side is left, the chooser becomes a plain label.
Ask for work
Small asks apply to the report directly.
A big job shows a plan card first. A job is big when it adds a new page, makes more than 3 visuals, or rebuilds a report from a screenshot or a sketch.
The plan card lists what the job will do: the pages and visuals it will make, the slicers it will set, the measures, field parameters and calculated columns it will request from Model AI, the custom visuals it will prepare, and anything it will delete or overwrite. You can edit the plan, and the job runs once you approve it.
Each hand-off is a row in the plan under Model AI or Custom Visual AI, with its kind (Measure, Field parameter, Calculated column or Custom visual), an editable Hand-off name and Hand-off intent and a model picker. The plan card carries its own Open Job view.
A job built from a picture shows it on the plan card under Pictures, with numbered marks over it, so you can say what a mark should become. A picture that cannot be loaded reads Picture not available.
Deleting or overwriting something the plan did not list still asks for your approval.
The report is saved as a version before the job starts, and it writes a version whose name begins with Report AI: after each finished page and at the end of the job, so you can always go back. See Save a report, resolve conflicts and restore versions.
Ask Model AI for new fields
The Report Builder does not change the semantic model itself. When a job needs a field the model does not have, the plan asks Model AI for it as a hand-off of kind Measure, Field parameter or Calculated column.
A total, a ratio or another value that adds up rows is a Measure. A value per row of one BI table, such as an age band to put on an axis or a slicer, is a Calculated column on the BI table it belongs to (Calculated columns).
A new field's name must not belong to a measure, field parameter, column or BI table of the model. A calculated column's name holds up to 128 characters.
Model AI builds the requested fields after you approve the plan and before any page work starts. It creates new fields and never changes an existing one: it creates only the calculated columns the plan lists, and changes or deletes only the ones it created for this request.
Model AI works in your acceptance mode, and its approvals show on the hand-off card. Its changes are listed under AI changes in the model's History tab, where an editor of the model can revert them (Model AI).
When Model AI reports back, the hand-off card names what it created, a calculated column with its BI table, such as "Customers"."Age band", and the Report Builder places the new fields like any other.
Choose how changes are approved
Pick the acceptance mode in the message box. It sets how much of the job is checked with you:
Mode
What it does
Auto-accept
The default. A big job shows its plan once for approval and then runs; small asks apply directly.
Manual
Every change waits for your approval.
Unrestricted
Skips the plan and the approvals.
A shield icon shows while Auto-accept or Unrestricted is on.
Follow a running job
When a job starts it says "I'm starting; you can go elsewhere, just don't close the browser". You can move around DataMind while it runs; if the browser closes, the job is parked with the status "paused: no browser" and resumes when you open DataMind again.
When the job finishes, or needs you, you get a notification in the app, a badge on the tab title and, once you allow it, a desktop notification. The permission is asked on the first job that runs in the background, as "Get a desktop notification when the Report Builder finishes?". The notification names the report and what happened, such as "{reportName}: has a question".
The report header carries the job's banner: Working with assistants, {name}'s request is working with assistants, {name}'s request is running, {name}'s request is waiting for them or Paused until DataMind reconnects, with Stop.
The Report Builder tile in the sidebar dock, above your account menu, follows the job; a phone shows the same activity as a floating Report Builder pill with a count badge at the top of the screen, and on a desktop there is no top pill. The tile's subtitle reads Needs your input, {n} running, 1 finished or {n} finished, and opening Report Builder lists your jobs: each row names the report and the job, shows its state, and offers Stop while the job runs or Dismiss once it has finished, plus Job view and Open.
A job's state reads Working, Has a question, Needs your approval, Resuming, Paused: no browser, Working with assistants, Done, Stuck, Stopped or Failed; a question parks the job on the tile as Needs your input.
Every step the assistant takes appears as its own card in the thread. Open a card to read What the Report Builder asked for and what came back; a step that is still running reads This step has not finished yet., a refused step reads Refused, and one that needed nothing from you reads No input. The cards name their step, such as Report plan, Add a page, Set the fields of a visual, Format visuals, Set the field parameter order, Ask Model AI for fields, Hand a visual to Custom Visual AI, Start page assistants, Start the Verifier, Picture pages, Check pages and File the verification.
The Job view holds the whole run. Open it with Open Job view on the plan card or a hand-off card, with Job view on a row in the dock, or from a link in the report's chat. Its Timeline tab lists who did what, and its Graph tab draws the agents and the hand-offs between them. The two tabs carry the tablist label Job view mode, the close control reads Close the Job view, and the header carries Stop. A job with no title reads Report Builder request, with the report's name and the job's state under it.
In the Timeline, steps by the same sender within 5 minutes group into one block. A step shows its model, its {tokens} tokens, its cost as $0.0042 or $0.42, its time as 350 ms, 12 s, 4 min 05 s or 1 h 12 min, and its clock time as Oct 18, 14:05. A node reads Idle, Working, Waiting, Needs you, Done, Failed or Stopped, and every event is a Request, Brief, Question, Answer, Report, Approval, Status or Message.
Message any assistant in the job with @. Under Message an assistant, pick it from Choose an assistant, write in Write to the assistant, or start with @ and its name and send it. The hint reads "The message reaches it when it finishes its current step.", and an assistant that is busy holds your message as Queued until you Withdraw it; a message that meets a waiting question can go with Send anyway. For Custom Visual AI the hint turns amber: Custom Visual AI reads this in the visual studio, which must be open on the tile.
An assistant's own thread is read-only, with the reason "This is an assistant of the Report Builder. Message it with @ in the Job view, or answer its question there.", and Back to the Job view returns you.
A hand-off gets its own card in the thread, headed Model AI · 3 new fields or Custom Visual AI · {tile title}, with its state, its model and the reason for it. Its actions are Open Model AI thread or Open Custom Visual AI thread, Open Job view and Stop hand-off, and stopping one shows Hand-off stopped. Its state reads Waiting for Model AI, Model AI is working, Model AI reported, Model AI is done, Model AI failed or Model AI stopped, or the same lines with Custom Visual AI, such as Custom Visual AI is queued. A hand-off that needs you reads Model AI needs your approval: {title}, with Approve and Reject.
While another assistant runs a round the job asked for, a banner above the chat reads Report Builder is working · round {n} · for {name}, with Stop. Stopping asks Stop this Report Builder round? and explains that the round stops there, anything made so far stays and the Report Builder goes on without it.
A step that previews a custom visual waits until the report's tab is open again ("Custom Visual AI is waiting for the tab").
Anyone who can edit the report can stop a running job, from the Job view or from the dock. Starting the job, answering its questions and approving its work belong to the person who asked.
A large job runs as several agents, one for each page, and closes with an automatic check of the finished pages for overlaps, query errors, empty visuals and number mistakes.
When it is done, the job is kept with the report as how this report was built. A finished turn carries Before this request: version {n} and Undo this request, which asks "Restore the report to version {n}, from before this request? Later changes stay in Version History." and answers Request undone with "Saved as version {n}. Nothing in the history was overwritten."; a version that has since left the history reads Version {n} is no longer in the version history.
The Job view remembers at most 50 jobs and loads their events 500 at a time. The dock remembers at most 100 dismissed jobs, and a job's journal is kept for 7 days.
What the Report Builder changes
Everything inside the open report: pages, visuals of every type, including Text box, Image and Bookmark navigator, field wells, options, limits, slicers of every style, page and visual filters, interactions, layout, which visuals are hidden, report bookmarks, the theme and the report's assets.
Two things stay outside its reach:
A viewer's own state stays with the viewer: expanded rows and drill state, saved column widths and a map's camera are not part of the report the assistant edits.
A custom visual's code is never edited. The assistant can place the visual, bind its fields, change its options and add its assets like any other tile, but the code belongs to whoever wrote the visual; see Custom visuals and the visual library.
You can also start a report from a description with New report with AI in the report list, and it rebuilds a report from a screenshot or a sketch; the dialog is on Build a report.
Bookmarks and stories
The Report Builder works with report bookmarks the way an author does in the Bookmarks panel; see Bookmarks.
It creates, changes, deletes, reorders and groups report bookmarks, with their options and captions, hides and shows visuals, and adds a bookmark navigator. It describes each bookmark's page, slicer selections, field parameter choices and visible visuals; it never copies them from your screen.
Ask it to make a story of the report, and it builds a group of bookmarks, each with a caption on what to notice, and optionally a navigator, ready for Present with bookmarks.
Ask it to show you a bookmark, and it applies that bookmark on your screen. That changes your view only, never the report.
A big job lists its bookmarks on the plan card under Bookmarks, each marked New, Change or Delete. Rename one in its Bookmark name field, or drop it with Remove this bookmark from the plan.
Changing or deleting an existing bookmark, or hiding an existing visual, asks for your approval when the plan did not list it, like any other overwrite.
Its steps appear as cards such as Create a bookmark, Change a bookmark, Delete bookmarks, Reorder bookmarks, Group bookmarks, Hide or show visuals and Show a bookmark.
When the Verifier checks a job that made bookmarks, it pictures each bookmark and reports one that points at a page, visual or slicer that no longer exists. Picturing a bookmark needs the server renderer.
Your own bookmarks, and everyone else's, stay out of its reach: it cannot see, create or change them.
Review my report
In edit mode, choose Review my report in the report header or in the AI panel toolbar. It is described as "Saves your unsaved edits, then the Verifier pictures and checks every page; the review itself changes nothing".
Open Review options and choose Review the whole report or Review this page.
Follow the review as a Review card in the chat. While it runs it reads Reviewing, Reviewing 3 pages or Reviewing every page; when it ends it reads "No problems found on 4 pages." or "1 error, 2 warnings on 4 pages.".
Read the findings, each with the kind of check, the page it sits on and a severity. The checks are Rendering, Overlap, Query error, Empty visual, Slow (30 s rule), Numbers disagree, Design and Other.
Tick the findings you want fixed, each with a tick box labelled Include finding {n}, and choose Fix these (2), whose label counts the ticked findings. The assistant receives them as a message that begins "Fix these findings from the review:" and works through them in the same chat.
The review saves a version named Before review first, so the report is where it was if you dislike the fixes. The card also says where its pictures came from: Pictures from the server renderer., Pictures from this browser., or Visual check not possible: no server renderer and the tab was hidden.
A review that cannot start says why: "You need Use DataPrism AI and edit rights on this report.", "A Report Builder request is still working on this report." or "Add a visual first.". If the report changed elsewhere, the toast reads "Someone else saved this report. Resolve the conflict, then ask for a review again.". Other outcomes are Could not start the review, "The review did not run. The Report Builder's reply above says why.", The review was stopped., "The Verifier got stuck:" with its reason, and The review could not be loaded.
Models
The Report Builder runs on Kartlos Premium by default. You can switch it to Kartlos Standard, and the agents it starts follow your choice. For each custom visual it prepares, it names the model it will use and its reason, and the cost estimate sits on the plan card, where you can change the pick before you approve. The check of the finished pages runs on Kartlos Premium.
Permissions
The Report Builder needs the Use DataPrism AI permission and edit access to the report. You do not need edit access to the semantic model: when a job needs a new measure or calculated column, the plan asks Model AI for it, and Model AI keeps its own gate.
Preparing a new custom visual also needs the Author DataPrism Custom Visuals permission, unless the tile matches a visual already published in the library.
On a custom visual, Interact needs no extra permission: in edit mode, any editor of the report gets it, on the selected tile and on the tile that fills the canvas in Prime. Its tooltip reads Interact with the visual, or Stop interacting with the visual while it is on, and only Edit code beside it still needs Author DataPrism Custom Visuals.
While you view as another user, the Report Builder can read the report with you, but it cannot change anything.
One job runs at a time per report. Everyone who can edit the report shares its thread and can watch the job; a copy of the report starts with an empty thread.
There is no token budget and no cap on how much work runs at once. Two safety rules remain: at most 2 custom visual steps run together, and a job that repeats the same error 5 times, or takes 20 steps with no change to the report, stops itself with "I'm stuck, here's why" instead of looping.
A request that needs a new measure or calculated column goes to Model AI as part of the plan. Without access to Model AI, the Report Builder works with the fields the model already has and says what is missing.
The Report Builder's field list marks which columns are calculated columns and which of them are broken. It places a working calculated column like any other column and never uses a broken one in a visual. People also add calculated columns themselves, in the model editor or from the report's BI tables panel.
In the plan card, the title holds up to 200 characters, the notes 2,000 and a hand-off intent 1,000.
Questions, answered
What is the difference between the Report Builder and Model AI?
They are the two sides of the same AI panel. Report Builder works on the report: pages, visuals, filters, layout and theme. Model AI works on the semantic model the report reads from: measures, columns, tables, relationships and tests. See Model AI.
Can I close the browser while the Report Builder works?
No. You can move around DataMind or switch tabs while the job runs, but closing the browser parks it. The status reads "paused: no browser", and the job resumes when you open DataMind again.
What happens when the job needs an answer from me?
The Report Builder tile in the sidebar dock marks the job Needs your input, and you get an in-app notification, a badge on the tab and, if you allowed it, a desktop notification. Open Report Builder and choose Job view or Open on that row to reach the report at the question; the job waits there for you.
Why did the Report Builder stop and say it is stuck?
It guards against loops: when the same error repeats 5 times, or 20 steps pass with no change to the report, it stops with "I'm stuck, here's why". Read the reason it gives, adjust your request and send it again.
Can two people run jobs on the same report?
One job runs at a time per report, and everyone who can edit the report shares its thread. Other editors can watch a running job and stop it; starting a job, answering its questions and approving its work belong to the person who asked. A copy of the report starts with an empty thread.
Can I make the Report Builder ask before every change?
Yes. Pick Manual as the acceptance mode and every change waits for your approval. Auto-accept is the default, and Unrestricted skips the plan and the approvals altogether.
Does the Report Builder create new measures or calculated columns itself?
No. When a job needs a measure, a field parameter or a calculated column the model does not have, the plan lists it as a request to Model AI, which builds it after you approve the plan. If you cannot use Model AI, the assistant builds with the fields the model already has and says what is missing. See Ask Model AI for new fields.
Does the Report Builder edit a custom visual's code?
No. It can place a custom visual, bind its fields, change its options and add its assets, but it never edits the code itself. To change the code, open the tile's Studio; see The custom visual Studio.
Where can I see how a report was built?
The Job view holds the whole run: a Timeline with each step's model, tokens, cost and time, and a Graph of the agents and hand-offs. Open it with Open Job view on the plan card, with Job view on a row in the sidebar dock, or from a link in the report's chat. The finished job stays with the report as how this report was built.
Which model does the Report Builder run on?
Kartlos Premium by default. You can switch it to Kartlos Standard, and the agents it starts follow your choice. The check of the finished pages runs on Kartlos Premium.
How do I have a finished report checked?
Open the report in edit mode and choose Review my report in the report header or in the AI panel toolbar, then Review the whole report or Review this page. The review pictures and checks the pages and lists what it finds with a severity; tick the findings you want fixed and choose Fix these (2). It saves a version named Before review first, and the review itself changes nothing.
Where do I see the Report Builder's jobs?
In the sidebar dock, above your account menu: the Report Builder tile shows Needs your input, {n} running or {n} finished and opens the list of jobs, each with Job view, Open and Stop or Dismiss. On a phone the same activity is a floating Report Builder pill at the top of the screen.