Ask AI is the chat screen in the console. You type a question, the assistant reads the platform’s records to answer it, and it says what it read underneath the answer. The screen needs the organisation’s provider and key to be configured; if they are not, it says so. See AI assistant for the setup.

​
What it can look at

The assistant has four read capabilities. All of them read; none of them write. That list is cut down to the person asking. In the session of somebody who cannot see cases, cases are never described to the model — which is why the phrasing of a question cannot change what is reachable.
The rule-set capability is not there by accident. The answer to “why are there so many declines this week” is usually a threshold somebody moved; without it the assistant would give you the number and no cause.

​
Threads and follow-ups

Conversations are listed on the left and are private to you. A follow-up — “and last month?” — only makes sense with what came before, so the last turns of the thread are re-sent with every question. Not the whole thread: making the twentieth question cost twenty times as much does not make the answer better. Under each answer are the capabilities that turn called, the name of the model that answered, and the tokens it spent. When a number is not what you expected, that is the first place to look — the question may have gone to a different source.

​
From a question to a report

When a question asks for a count or a breakdown, the assistant runs the report and shows the table under the answer. Scroll back up the thread and an earlier turn’s table re-runs with Show the table — what is kept is not the table but the question that produced it. Saving a report under a name and sharing it with the organisation is the AI Reports screen’s job; the chat screen does not save reports.

​
Things to ask

The assistant reads. It does not close cases, publish rules or change settings. Turning a number into an action happens on the screen that owns that action, with that screen’s controls.

​
What to trust, and what to check

The numbers come from the platform’s own records; they are not invented. The room for error is not in the number but in how the question was understood: “last month” and “the last 30 days” are two different ranges and both are legitimate. The capability list under the answer, and the heading of the table that comes back, say which range and which dataset were used. If a figure is going to be reported onwards, confirming it on the monitoring screens is a minute’s work.

AI reports

The screen that turns a question into a table and a chart, and the datasets you can ask about.