Configuring the assistant — integration, model, prompts, and costs

Step by step: enabling the assistant for your company

1. Open the assistant using the top-bar icon (side panel) or by going to Tools → AI Assistant (full page).
2. Ask any question. If you get “AI is not configured for this company.”, an AI integration is missing — continue to step 3. If you get “No plugins available.”, the integration works, but your role does not grant access to any area; contact the roles administrator (Assistant limitations and common misunderstandings).
3. Connect an AI integration as you would any other integration (Integrations list — my connections and the available catalog). Without one, the input field is disabled, while the knowledge base continues to answer (The assistant and the knowledge base — why it prefers silence to guessing).
4. Return to the assistant's full page. The AI model list is in the upper-right of the heading.
5. Choose a model. Default (integration) means the model configured for the integration. Your choice applies to this conversation, not globally.
6. Enter a question in Type a message... and send it. You can interrupt a response while it is being generated using the button that replaces the send button, but a cost may already have been incurred.
7. Track costs in the session summary and on the screens under Reports → AI.

A connected AI integration is required

Without one, the assistant reports that it is not configured for this company, and the input field is disabled.

Important: the knowledge base keeps working. The message directs you to the documentation, which answers “how do I do this?” questions (The assistant and the knowledge base — why it prefers silence to guessing).

Connect an AI integration as you would any other integration (Integrations list — my connections and the available catalog).

Choosing a model

The assistant's full page lists the available models and their visible token costs. Your choice applies to this conversation, not globally.

Prompts

The assistant can use its own system prompt from an AI prompt template. It is appended before the built-in prompt, so it does not replace it; it adds context. On the assistant's full page, choose it from the Template list beside the model list.

Create an assistant template under Settings → Integrations → AI instructions: click Add custom instruction, choose the AI Assistant tile, then start from a copy of the default or from a blank template. Once saved, the instruction appears in the assistant's Template list (AI instructions — tailor AI tools to your company). The assistant sends only AI Instructions (fixed) — its instruction has no “Content with data” field.

Leaving the AI Provider and AI Model fields empty in the instruction is not an error: it will use the company's AI integration and that integration's default model. The AI Instructions screen states this in a banner above the list (for example, “OpenRouter · openai/gpt-4o-mini”); the Model column shows that value in gray with “(default),” and the form repeats it below the empty field.

Useful purposes for a template include your company's terminology, priorities, and reminders about local practices.

Response temperature is deliberately set very low and cannot be changed. The assistant should be predictable, not creative.

A conversation expires after 30 minutes

A session expires after half an hour without activity, and the next question starts a new one. The context retains the last 10 messages.

In practice, during a long analysis the assistant will not remember the beginning of the conversation. Repeat important decisions in your question.

Costs

For every answer, the cost and token count are tracked separately for understanding the question and formulating the answer. Each session has a summary, and Reports → AI contains usage and effectiveness screens.

Rate limit

You can send up to 10 questions per minute — and the limit applies to the entire company, not each user. One person sending questions in quick succession can temporarily block everyone else.

Interrupting a question

You can interrupt a question in progress: the send button changes to a stop button. The answer is discarded.

To be clear: interrupting does not always save the cost. If the model has already calculated the answer, the charge has been incurred; the answer is simply not displayed.

Proactive alerts

Separately from your questions, the system periodically checks several things — negative stock, an excess of warehouse errors, a high AI error rate, and orders unconfirmed for more than two days — and creates alerts. Repeated alerts are muted for a day, and resolved ones are cleared after a month.