Beyonetix OS module
Assistant in the ERP: questions, commands and files in your own words
You write what you want to know or get done. The assistant searches, calculates and drafts, and asks before anything is changed.
In short
The AI assistant is the chat sidebar in the ERP of Beyonetix OS (BEYOS). It answers questions about your data, runs commands once you confirm them and creates files. It is meant for everyone who wants quicker answers in daily work.
Updated: November 2026
What it is for
An assistant that knows your data and respects your rights
Someone in the office asks which quotes have been open for three weeks, or who last called this customer. Nobody has to build filters or export lists for that. The question goes into the sidebar, and the answer comes from your records and documents.
The assistant works inside the session of the person asking. You set per model which data it may draw on. Files are created under the rights of the person who requests them.
Flows that work through several steps in a row and put them in front of a person at approval points are on the agent platform page. This page covers the conversation with the ERP and the layer beneath it that connects the models.
- For case handling, sales and accounting in the ERP
- For managers and IT who control models and data paths
In the ERP
Questions, commands, search and files
The sidebar is available in the ERP views. Conversations are kept, and the assistant knows the context so far.
How it works
From question to answer in six steps
Model layer
The layer beneath the assistant
Many ERP modules talk to the AI through one central layer. A few functions call declared external services directly.
Technical details
The model layer at a glance
| Property | Detail |
|---|---|
| Model connection | Replaceable, with a model server of your own, interface standards and external services. |
| Task per model | Router by task or calling module, otherwise the default model applies. |
| Record per transfer | Trigger, time, provider, model, task, record, volume, duration, result. No wording of question or answer. The entry is written even when a call fails. |
| Labelling | Email drafts and replies carry a note, and the log keeps count. |
| Transfers | Only with permission per company. Without it nothing goes to an external provider through the model layer, and an entry goes onto the work list. |
| Failure | Short time limit per call so that work does not stall. |
| Setup | Per company, in the ERP settings. The interface and the instructions to the models are in German. |
AI suggestion with approval
Suggestions that wait for your yes
This building block is developed and comes into your setup on request.
Further
Where the agents are
Agents that process documents, use templates and work with a dry run and approval have a page of their own. It also explains how you set up and supervise an agent.
Agents that run in the ERP use the same model layer as the assistant, and their transfers appear in the same log. Agents in other modules keep their module's own log.
- See the agent platform
- Document processing with approval
Questions and answers
Questions about the AI assistant
Does the AI decide on its own?
No. A command runs only after you confirm it, and an AI suggestion never writes anything by itself. Purely read-only evaluations, such as a question about your data, need no approval.
Can the model behind the assistant be swapped?
Yes. Modules that run through the central model layer follow the provider you switch to in the setup, without any change to the modules.
What does the model layer's log record?
It records who triggered a transfer and when, to which provider, with which model and for which task. The log does not store the wording of question and answer.
Where does my data go?
That depends on your setup. Without the company's permission nothing goes to external providers through the model layer, and the case lands on a work list without its content. This does not replace legal advice.
What can the assistant produce?
On request it delivers files in Excel, Word and PDF. The data in them comes under the rights of the person asking, so whatever that person may not see stays out of the file.
Can generated texts be recognised as machine-made?
Email drafts and email replies carry a note that a machine produced them, and the log keeps count. Further tasks can be registered during setup. This supports your labelling duties under the EU AI Act but does not replace legal advice.