Model AI

Model AI is an assistant panel beside a semantic model. Ask it to read, check and change the model's measures, columns, tables, relationships and field parameters, decide how its changes are approved, and undo them when needed.

Open Model AI

1. Click DataPrism in the sidebar and open a saved semantic model.
2. On the right edge of the screen, click the Model AI strip. The panel starts folded, and DataMind OS remembers whether it is open in this browser.

Model AI also opens beside a report, for the model that report reads from. It appears only when a saved model is selected. Each model has one chat, shared by everyone who can edit the model.

Choose how changes are approved

Two settings control Model AI:

Click the mode chip in the message box to change the acceptance mode. The picker explains each mode:

• Auto-accept: "The AI saves new measures, edits of unused measures, folders, formats and descriptions without asking; changes to measures that charts use, deletions, relationship and fact-table changes, and changes to tests a person wrote wait for your approval."
• Manual: "Every change waits for your approval."
• Unrestricted: "Every change is made without asking, except removing a BI table."

Your choice is saved with that model's chat, and each editor of a shared chat sets their own.

Review and approve a change

1. Type your request in the message box. To show Model AI something, attach a screenshot; only images can be attached.
2. When a change needs you, its card shows Model change needs your approval.
3. Click Approve to apply it, or click Reject. In the reason box, add a reason if you want Model AI to know why, then click Confirm reject. Click Cancel to close the box without rejecting. A rejected change is not made: "An editor of this model rejected this change, so it was NOT made."

An approval shows what it will change before you decide:

• the object and the operation, listed as {objectType} · {op}, for example measure · update;
• the fields that change, as {field}: {before} -> {after}, with none for an empty value;
• the impact, Impact: {n} reports, {m} charts, with any reports and charts you cannot open counted in brackets after it;
• the reports and the dependent measures, listed by name;
• for measures, a numbers table titled "{measureName}" before and after by {groupBy}, with Before and After columns, so you can compare the values first.

Long lists are cut to their first 20 entries. While a change waits for your approval, you cannot send a new message until you approve or reject it.

Model AI can change eight kinds of object: measures, columns, BI tables, relationships, the model itself, measure folders, field parameters and measure tests. Its changes are creates, updates, deletes, additions, removals, renames and replacements. To learn about the objects it changes, read Measures and Relationships and the core table.

See what Model AI is doing

While a turn runs, the panel streams what Model AI is doing, for example Reading the model outline, Reading your editor, Checking the model, Running measure tests, Querying the model, Previewing "T", Running SQL, Estimating the cost of a query, Creating 3 measures and Replacing a BI table source. A tool the panel does not know shows its own name.

One turn runs at a time in the chat. While a turn runs, the chat names the person behind it and shows Thinking Ns while it waits, then Thought for Ns. A message sent in the meantime is refused: "Your previous turn in this chat is still running. Stop it or wait for it to finish, then send again."

Read the result of a change

Each change card shows what happened to every object it touched:

• The summary counts the results, for example 3 saved, 1 skipped.
• Every object carries a status chip: saved, failed, skipped or not attempted. Hover a chip that is not saved to read its message.
• Each changed field is listed as {key}: {before} -> {after}; long values are shortened at 120 characters and end with ....
• For a change to the model itself, extra lines name what moved, such as the Type column and Version flags.

Read result cards

A reply can carry the results of work Model AI did for you: a table result card for a query it ran, and a tests card for measure tests it ran.

A table result card shows the rows a query returned.

• The badge above the grid shows the size of the result: n of m rows when the total is known, and n rows when it is not. truncated appears when the run did not return the whole result.
• The grid shows the first 20 rows. Click Show N more rows to reveal the rest.
• A cell with no value reads null. Long values are trimmed in the grid; hover over a cell to see the full value.
• Click SQL to show or hide the query behind the result. The view starts open and can be copied.
• Click Show as chart to view the result as a chart; the button then reads Hide chart. Switch the view between Bar, Line and Table. The chart controls appear only when the result has a group column and a value column.
• The RLS applied or RLS not applied badge tells you whether the run used row-level security. See Row-level security and View as.
• The n ms badge shows how long the query took.

A tests card summarises a measure test run.

• The summary line counts the run: n pass, n fail, n error, n not run. It ends with (not saved) when the measure is not saved yet.
• Each row names its test and shows its status chip: pass, fail, error or not run. Under the name, the row shows the measure and the kind of test, as {measure} · {kind}.
• A failed test shows expected {expected}, actual {actual}, then up to 10 lines, one for each key that differed.
• When the run has no results to list, the card shows the summary line alone.

To run tests yourself and follow their saved results, see Measure tests.

Revert an AI change

1. Under a change card, click Revert. To undo everything from one reply, click Revert this turn under that reply. It appears once the reply has finished.
2. In the dialog, click Revert. The dialog warns that changes are undone in reverse order and that objects someone changed since are skipped and listed.
3. To find older changes, open the model's History tab and go to the AI changes section.

Reverts run in reverse order. If someone changed an object after the AI did, that object is skipped and listed in the result, and the summary reads, for example, Reverted by {name}: 2 reverted, 1 skipped, 0 failed. If a revert fails, a Revert failed message appears. Reverting needs the same permissions as using Model AI.

A revert never overwrites work done after the AI's change, so it leaves some objects alone; each skipped object carries its reason, and the result lists every object as reverted, skipped or failed. The usual reasons:

• someone edited the object since the AI's change, for example changed since the AI's change: name is now Metric 3;
• charts, report filters, relationships, dependent measures, row-level security groups or field parameters now use it, for example used since the AI's change: 1 charts, dependents: "Double";
• a person wrote tests for it since the AI's change;
• the object, or its measure, no longer exists, or it is no longer part of this model;
• the table now carries the model's fact rows, or a row-level security group targets it;
• the two tables are joined another way since the AI's change;
• a replaced source can only be put back while its record is kept, and the model keeps the newest 20 replacements;
• the change was too large to keep its image, so it cannot be undone;
• the object changed while the revert was running;
• the change is the removal of a BI table, which is permanent.

Some changes cannot be reverted. The chat says Nothing to revert in this turn. when a reply made no changes, This change was already reverted. for a second attempt on the same change, This change was rejected, so nothing was made. for a rejected change, This change has nothing to revert. for a change with nothing to undo, No AI changes were found for this semantic model. on a model with no AI changes, and This change cannot be reverted: removing a BI table is permanent. for a removed BI table; a change that cannot be found reads AI change not found, so reopen the section and try again. Further refusals: Changes cannot be reverted from an impersonated session. while you view the platform as another user, and A turn is still running in this model's AI chat. Try again when it finishes. while a turn runs. While a change waits for approval, reverting is refused until you decide. For the change history of a model, see Check, document and trace a model.

Check a query's cost

Ask Model AI to check a query's cost. The result card rates the query Fast, Watch or Slow, and shows Rows read, Bytes read and Peak memory. The card marks its figures as approximate when the database cannot give an exact estimate. For the full cost check, see Query cost.

When a Model Builder is working

A Model Builder is a builder Mind that works on the model. While it runs, a banner appears above the chat.

• The banner names the round, for example Model Builder is working · round 3 of 8 · started by {name}. Between rounds it reads Model Builder is waiting to start its next round, Model Builder is loading full history or Model Builder waits for answers in the Warehouse Builder chat.
• If you can stop the builder, the banner shows Stop. Click it, then click Stop in Stop the Model Builder?. Anything the builder has already made stays in the model, and the person who started it can start it again from the Warehouse Builder chat.
• Messages you send meanwhile show Queued · delivered after the round, and messages other editors queued are listed by name. Click Remove to cancel a queued message.
• When the Model Builder hands a round to the Model AI, the reply starts with a hand-off row that names the round and the person behind it, with a link reading From {name}'s Warehouse Builder chat.

To learn how builder Minds work, read Warehouse Builder and Lake Builder. For the session card in the builder chat, see Model Builder sessions; for how a builder's changes are approved and its questions answered, see Model Builder approvals.

Permissions

Model AI needs all three of these permissions. Without any of them, the panel does not appear.

You also need edit rights on the particular model you are working on. Everyone who passes these gates shares the model's one chat, and any of them can approve or reject a change. For the full list, see Who can do what in DataPrism.

Limits and edge cases

• Model AI works only on a saved model.
• Without Use DataPrism AI, the refusal reads: "You do not have permission to use AI in DataPrism. Ask an administrator for the "Use DataPrism AI" permission."
• If you can open a model but cannot edit it, Model AI refuses your request with: "You cannot edit this semantic model, so you cannot use its Model AI."
• Removing a BI table always asks for your approval, even in Unrestricted.
• If Model AI refers to an object that is no longer in the model, it says: "<name>" is not in this model any more.
• If the chat cannot load, the panel shows: "The Model AI could not be opened."
• The chat belongs to the model and is shared by its editors. It cannot be made private or deleted, and it cannot generate reports or presentations.
• While you view as another user, you can read the chat, but you cannot change the acceptance mode or revert changes.
• On a model that tokenises personal data, the conversation shows tokens, not clear values.

Questions, answered

Which mode is safest?

Manual is the safest. Every change waits for your approval.

Will Auto-accept change measures my charts use?

No. Changes to measures that charts use wait for your approval. Deletions, relationship and fact-table changes, and changes to tests a person wrote also wait.

Can I undo a change Model AI made?

Yes. Click Revert under a change to undo that change, or Revert this turn under a reply to undo all its changes. If someone has changed an object since the AI changed it, that object is skipped and listed. Older changes live in the model's History tab, under AI changes.

What do "skipped" and "not attempted" mean on a change card?

skipped means the change was left out on purpose, and not attempted means Model AI stopped before reaching it, for example because an earlier object failed. Hover either chip to read the message behind it.

What does the impact line on an approval count?

The reports and charts the change touches. Counts in brackets, as in Impact: 3 reports, 12 charts (2 reports and 5 charts you cannot open), are the ones you cannot open yourself.

Can two people use one model's Model AI at the same time?

They share the chat, but one turn runs at a time. While a turn runs, the chat names the person behind it, and a message sent meanwhile is refused until the turn stops or finishes.

I do not see the Model AI panel. Why?

You need Use DataPrism AI, View DataPrism Semantic Models and Edit DataPrism Semantic Model, and you must have a saved model selected. Ask your administrator for any permission you lack.

Can the AI delete a BI table without asking?

No. Removing a BI table always waits for your approval, even in Unrestricted.

Why did my message not send while a builder was working?

It is queued instead. The chat delivers it when the round ends; to cancel it, click Remove under the queued message.

Why do I see only some of the rows in a result card?

The table card shows the first 20 rows of the result. Click Show N more rows to see the rest. If the card shows truncated, the run itself was cut short and did not return every row.

Why does a result card say "RLS not applied"?

It warns that this run did not apply row-level security, so the card's rows are not narrowed the way they are in your reports. Treat its numbers as wider than what you normally see. See Row-level security and View as.

What does "(not saved)" mean on a tests card?

The tests ran against a measure that is not saved yet, so the run's result was not stored with the measure. Save the measure and run the tests again if you need the result kept.