AI Assistant

What it does

The assistant answers questions about the group you are working in — "Why did the gross margin fall in September?", "Which accounts are still unmapped?" — by calling the same read tools a connected assistant uses: statements, ratios, statistics, budgets, mapping coverage, drill-downs and the group structure. It never computes a figure of its own: every number in an answer is checked against what the tools returned, and an answer that quotes a figure the tools did not return is withheld. The tools used are listed under each answer.

Open it from AI Assistant in the sidebar or from the Boardroom's Open the assistant link. The conversation stays for 24 hours in your session and belongs to the group you are in; New conversation starts again.

Type your question in the box under the conversation: Enter sends it, Shift+Enter starts a new line. An empty conversation offers four example questions — click one to ask it. When an answer is withheld because it quoted a figure the tools did not return, the group's admins see which figures under For administrators; the assistant has already tried once more, asking for those figures through the tools.

Changes it can make

Only where the group's admin has turned on assistant writes (Settings → Connected AI assistants) and only within your own role, the assistant can propose or make the same changes a connected assistant can — mappings, adjustment journals as proposals, notes, review points and summary views. In propose mode it prepares them for a person to accept; in apply mode it writes them through the application's own writers, and the page says so above Ask. Every change is marked as the assistant's, can be reversed and is listed with its audit row under the answer. It never locks, deletes or accepts anything, and never writes in a locked period. A viewer's assistant and a sample group's assistant only read.

Allowance

Each group has a monthly allowance of questions (300 unless agreed otherwise). An admin sees the questions used, the allowance and the day it resets on Settings → AI usage. Only answered questions count — a question the assistant could not answer is not charged. When the allowance is used up the assistant says so and the reports work as before.

Which AI model answers

A group's admin chooses the models on Settings → AI models: which ones the assistant may use, the default, and the order in which the others answer when the default has no capacity, times out or fails. Each model shows where the data is processed and its sub-processor; a model outside the EU (Claude through Anthropic's own API) can be allowed only once Anthropic is on the group's list of sub-processors. Admins see Answering with: … above the question box, with a link to change it. Until the models are chosen, Home and the Connected AI page remind the group's admins with Choose which AI models your team may use →; meanwhile the assistant answers with the platform's default model.

When the group allows a Thorough model, the question box offers Standard and Thorough; a thorough answer is slower, goes deeper and uses 3 questions of the allowance. The label under every answer names the model that answered and, when the default failed, which one it stood in for.

When it is available

The assistant is part of the Beta plan. It runs in the sample groups and in a client group once the platform has enabled AI for it; client data is processed by AI models in the EU, unless the group has agreed to Claude through Anthropic's own API, which is then on that group's list of sub-processors.