On this page
In short. Ask a question in plain language. The answer is a number and a chart from the same measure catalogue every dashboard uses, or a plain "no data", never a guess. The assistant calls typed tools that already carry organization scope, brand scope and Vietnam time, and it reads no free-form documents.
How it works#
The assistant sits under AI Assistant ("Trợ lý AI") in the left menu. A super admin turns it on per organization. Inside an organization, the menu item needs the Manage AI capability, which owner and admin hold by default.
Each tool wraps one measure definition, so organization scope, brand scope, Vietnam time and the revenue formulas are inside every answer. Tools are grouped by area (shared core, eCom, marketing, CRM, owned channels, market data, customer discussion, organization datasets); see Reference.
On every request the platform checks which planes hold rows and builds the tool set and the prompt from that result. A plane with no rows has no tool, no description and no mention, with no super-admin exception. A question about it gets a statement that the plane is off and a pointer to an administrator, never an estimate. Market data and customer discussion also need their own plugin.
Starter prompts and history#
An empty conversation offers only questions the organization can answer: a real subject, a usable measure and a group with at least two members carrying data. Administrators set the name, greeting, starter prompts, provider and budget in Configure AI and the data assistant.
Every tool call writes a log row with its parameters, result and duration. A failed call becomes a "Tool failed" card instead of breaking the reply, and each reply names the source query it used. Conversations are stored per member under organization and brand scope. Reopening one loads its newest 60 messages, with masked names replaced for the reader exactly as in the live answer; a public share of a conversation is masked for its anonymous reader. Deleting a conversation archives it.
Rating an answer#
Under each answer, the thumbs-up (Helpful answer) and thumbs-down (Not helpful) record your verdict; a thumbs-down asks What was wrong? with Wrong numbers, Missing information, Off topic or Other reason. You rate answers in your own conversations only, and pressing the same thumb again clears it. Ratings are stored with the answer and do not change later answers. The assistant inside a presentation deck has no rating yet.
Where it stops#
- A turn stops after ten tool calls.
- There is no search over documents: an out-of-scope question is declined and redirected to what is answerable.
- A period with no data is answered as "no data", never as zero.
- Names that identity masking hides are replaced before the answer reaches the browser (Identity masking).
- Rollout state: Data assistant, inside Content Intelligence.
Reference#
Tool areas
| Area | Examples |
|---|---|
| Shared core | Catalogue, snapshot or trend, period comparison, ranking, diagnosis, forecast |
| eCom | Order detail, livestream sessions, the operating cycle |
| Marketing | Ad spend, budget pacing, campaign performance |
| CRM | Lead funnel |
| Owned channels | Web, search, social and email performance |
| Market data | Comparison against the market, competitor moves |
| Customer discussion | Mention counts, sentiment, the vocabulary table |
| Organization datasets | Direct queries over the organization's own uploaded data |
Gates before a question runs
| Gate | Rule |
|---|---|
| Feature | The data-assistant plugin is on for the organization; a super admin passes |
| Member | A member of the organization; the menu item needs Manage AI |
| AI provider | Not configured, or a key that no longer decrypts: a pointer to the AI settings |
| Rate limit | 30 requests per minute per member and organization |
| Brand | A member limited to some brands must have one of them selected |
| Chat budget | A super admin may cap the organization's chat tokens per month; once spent, questions are refused until the 1st of the next month, Vietnam time. No cap by default |
| Data scope | Every tool call re-pins organization and brand before it reads |
What it reads and never reads
| Reads | Never reads |
|---|---|
| The measure catalogue of the planes with data | A plane with no rows: no tool, no description, no mention |
| Rows under the organization and brand re-pinned on each call | Free-form documents, or a number no tool returned |