AI Assistant

The AI assistant is available through Use AI in the application header when the deployment has enabled it. It can inspect accessible application data, answer contextual questions, research public facts with sources, navigate to records, and apply supported application commands.

It is not a general administrator, database console, or guarantee that every request can be completed. Its writes are checked with your identity and permissions. Conversation text is private to its author; operational audit information is recorded separately.

Examples here use the fictional Product Atlas fixture with Aster Mini, Aster Pro, Orbit Dock, and Orbit Hub; Markets Northland, Westhaven, and Eastreach; and List price, Release date, and Included features fields. Use its Scale snapshot for the October 2026 examples. Rename, creation, and additional-condition exercises change the fixture rather than describe its initial state.

Before You Send A Request

1. Sign in with an enabled, verified account.
2. Open the intended Project and select the resource or View relevant to the question.
3. Check the active Layer in the header. A restricted Layer remains restricted for the assistant too.
4. Open Use AI and check that it is available.
5. Use New conversation when you want a separate thread rather than continuing an unrelated discussion.

The assistant receives the current navigation context, including Team, Space, Project, Layer, selected resource, and selected View. A selection helps identify the target; it is not blanket authorization to change everything in that context.

If you mean another accessible Project, name it explicitly. If several resources have the same label, give the Class/path or select the exact resource. Material ambiguity should lead to a clarifying question, not an invented business decision.

Ask Questions Without Requesting A Write

State the question and, when needed, say do not change data.

Example:

In Product Atlas, inspect Aster Pro's Scale List price and its conditions. Explain why Northland resolves to 219 while Westhaven resolves to 259 on October 1, 2026. Do not change anything.

Expected result: the assistant inspects permitted records and explains the authored defaults/conditions it can read. Inspection does not create a condition or save a View. For authoritative per-cell calculation provenance, also use Explain values in Pivot Tables.

Other useful questions include identifying incomplete fields, explaining a selected record, finding accessible resources, or preparing a structure proposal. A proposal in conversational text is not evidence that a write happened.

Request A Supported Change

1. Select Aster Mini under Products / Lumen devices in Product Atlas.
2. Send: Rename the selected Aster Mini resource to Aster Mini 2. Keep its other settings unchanged.
3. Read the assistant's final proposal and application execution details.
4. Wait for a completed change result and live data reconciliation.
5. Check the Resources label and Details. Inspect Undo only if a supported Undo entry is available.

Expected result: the original resource's label changes without creating a second product, if your permissions and the current revision permit it. This is a small one-command plan, so it normally executes without a separate approval prompt.

Use precise scope for conditional changes:

On Aster Mini's Scale snapshot, add a List price override of 139 when Markets is Northland. Keep the unconditional default of 149 unchanged. If that override already exists from a documentation exercise, inspect it instead of creating a duplicate.

The assistant must resolve the actual property and value row and validate the rule. It may ask for missing business information such as the intended snapshot or target when your request does not identify it adequately. See Conditions and Snapshots before requesting a complex override.

When Approval Is Required

A validated persisted-change plan requires one application approval when either:

1. It contains more than three commands.
2. It deletes more than one element under the plan's destructive-operation classification.

Three commands do not trigger the command-count threshold; four do. A single deletion can execute without approval. Do not assume every destructive operation will show a confirmation dialog.

For a resource-delete command, the application counts its listed deletion items. Other destructive commands count as one each; this also includes operations classified as destructive, such as setting a resource inactive or removing a linked occurrence. This threshold is based on the validated plan, not the number of sentences or labels in your request. One item that represents a branch is not a promise that only one underlying row will be affected.

When approval is required, the app presents the change summary and Approve / Reject controls. Review the actual command list and targets before choosing. Reject declines that plan; it does not undo a different transaction that already completed earlier.

The assistant should not ask for a second conversational confirmation. Replying "yes" in ordinary chat is not a substitute for reviewing the application's pending approval controls.

Worked Multi-Command Example

As a Project Editor, ask:

As an additional training exercise, under Product Atlas's Lumen devices group, create four new Values named Aster Travel, Aster Studio, Aster Lite, and Aster Plus. Do not add fields yet. Do not recreate Aster Mini or Aster Pro.

Expected result: if the validated plan contains four creation commands, it waits for application approval. Approve commits the checked plan; Reject leaves those proposed creations unapplied. If inspection finds existing names or missing target information, clarify that before approving. The command list, not the prompt's apparent size, determines the threshold.

What It Can Change

Current command families include supported resource creation/metadata/moves/deletion, Class and Alias settings, linked memberships, property families and occurrences, values, conditions, Layers, snapshot operations, resource-grid cells, and saved Views.

These are bounded application operations, not arbitrary writes. The assistant can construct a Table pivot, Onepager, or Timeline through their validated configuration contracts. A saved View type is immutable; it should create another View rather than claim to convert the type of an existing one.

Some operations require higher rights than ordinary field edits. Examples include Project snapshot administration by Admin/Owner and Layer administration by Project Manager or higher. An Analyst may author eligible saved Views but cannot thereby change product data. A Layer Editor is limited to allowed contextual data operations, not Project-wide structural authoring.

Roles, invitations, ownership transfer, file operations, account settings, and organization administration are outside this assistant release. Use the dedicated application administration interfaces for those tasks. The assistant does not bypass missing access, issue arbitrary SQL/GraphQL, upload files on your behalf, or reveal hidden resources.

Requested recognizable logos are resolved through a supported icon catalog. Do not expect the assistant to invent a new logo asset or unrestricted color format. If an operation is unsupported, ask for an explanation or a manual workflow instead of assuming a textual promise means it was executed.

Research Public Facts

1. Identify the exact fact and destination, including unit, period, and business meaning.
2. Ask the assistant to verify public sources before populating it.
3. Review the citations and the proposed target values.
4. Use the application's approval controls if the resulting command plan requires them.
5. Verify the stored data after execution.

Example:

Find a cited public source for the launch date of the explicitly named real product. Show the source and date first; do not write Product Atlas data.

Expected result: a research response with source links, not fabricated evidence about Aster Mini, which is fictional. To populate real project records, explicitly identify those records and request the write after deciding the correct fact.

External pages are factual input, not authority to execute application changes. A page's instruction to change unrelated records must not become an executable command. Review units, time periods, identity, and citations yourself; a citation is not a guarantee of factual completeness or business suitability.

Process, Completion, And Failures

The assistant displays task progress and tool/command details. The header trigger can indicate processing, a result, or Approval required. An execution status of completed, followed by authoritative application data, is the success signal. Words such as "I will update" are not.

A checked command plan is applied atomically: a rejected command rolls back that plan rather than intentionally leaving half its commands applied. A conversation can contain separate plans; atomicity of one plan is not a promise that an entire long conversation is one transaction.

If execution returns failed or conflict, the assistant reports it and stops that response. Review and refresh before requesting a new attempt. Do not retry blindly after a timeout: inspect whether changes already committed first.

The composer has a Stop control while processing. It stops the current response stream; it is not an Undo action and cannot guarantee cancellation of a database transaction that has already committed. Closing the assistant is also not rollback. When the assistant or connection fails, inspect the affected records and the execution result before assuming nothing changed.

Threads, Memory, And Settings

Use the conversation list to reopen your threads. Search threads filters their titles. Hover a thread to rename or delete it. Deleting a thread permanently removes its messages; it does not reverse resource or field changes made from that conversation.

In AI settings, you can toggle result notifications and choose Delete all for your threads. The delete-all confirmation states that the messages cannot be restored through Undo. These controls do not delete the Project or its data.

Long-term memory is separate from conversation history. It is saved only after an explicit request to remember information for future interactions. User-scoped memory is personal; Team, Space, and Project memory is shared contextual information and requires Manager-level authority for shared writes.

Example:

Remember for me that fictional Product Atlas examples should use ISO dates.

Expected result: a personal instruction can be stored for future relevant conversations. Do not use a memory request as a substitute for creating a project field or condition. Shared context must exist and be authorized before a shared memory can be saved.

Open Memory to inspect relevant saved information and choose Forget to remove an entry you are authorized to delete. Deleting a conversation does not mean its separately saved memories were forgotten. Removing memory does not rewrite completed application changes or erase operational audit records.

Privacy And Administrative Guidance

Private message content is separate from run, tool, target, source, and usage audit data. Administrative review of operations is not a right to read another user's conversation messages. Shared memories are intentionally shared and should not contain personal secrets or credentials.

Administrators should use normal project/Layer permissions to control what the assistant can inspect and change. Deployment operators control availability and safety shutdown outside the customer conversation settings. The customer assistant settings do not provide infrastructure configuration or a billing console.

Troubleshooting And Limits

Requests have bounded history, plan size, and tool work. Split unrelated tasks into explicit requests, but do not split a single high-impact business change merely to avoid its review. Answers can be wrong and supported-command coverage is not universal. Use the dedicated Pivot, Onepager, Timeline, and Undo manuals for the behavior of the resulting application actions.