Skip to content

Beyonetix OS module

AI agents in Beyonetix OS: start from templates, with approval and a log

Agents take over routine work in your modules. They start from a template, work within fixed limits and show their work wherever a person should decide.

In short

An AI agent in Beyonetix OS pursues one clearly defined goal, uses only the tools it is allowed to use, stays within fixed limits, shows its work to a person at approval points and logs every step. Agentic means here that the system carries out several steps itself, while people decide at the points where it matters.

Updated: November 2026

Definitions

Chatbot, automation, agent

The three terms often get mixed up. The table shows how they differ in everyday work.

Comparison of chatbot, automation and AI agent
CriterionChatbotAutomationAI agent in Beyonetix OS
What it does Answers questions in a conversation. Carries out fixed steps exactly as they are defined. Works through a goal in several steps and reads documents to do so.
What starts it A question in the chat. A schedule or an event. A schedule, an event, an upload, an arrived document, a deadline, an email in the mailbox or a start by hand.
What it may write Usually nothing in the system. Everything in the flow, even without asking. Proposals and drafts. In ENGIVERITY and Ad Pilot, steps that write wait for approval by default.
Who decides The person who asks. Whoever built the flow. The responsible role at the approval point.

Getting started

How an agent goes live

Four steps from the idea to a running agent. Templates and tools sit with each module.

01 Choose a template You choose the template that fits your process, together with us if you wish. In Documents the templates that ship with it first sit switched off in the tenant and can be switched on once all mandatory details are in place. In ENGIVERITY you adapt them in the flow workshop.
02 Assign tools In ENGIVERITY you decide which tools the agent may use.
03 Set rules and approvals You set the limits, the approval points and the role that decides at each one.
04 Check in a dry run, then start A dry run shows what the agent would do. Then you switch the trigger on. We support you with the setup.

Triggers

How agents start

In ENGIVERITY every start is logged, and a safeguard stops a flow from running twice.

Triggers, examples and modules
TriggerExampleModule
Schedule Every weekday at 07:30 a report goes to management. ENGIVERITY, Documents, ERP and CRM
Event in the system A new version of a bill of materials is created. ENGIVERITY
Document arrived An invoice arrives through the mailbox, a folder or the interface. Documents
Deadline A discount is about to lapse, a contract becomes terminable. Documents
Upload A bill of materials or a delivery note is uploaded. ENGIVERITY, Documents
Email in the mailbox An email arrives in the connected mailbox. Documents, ERP and CRM
By hand A person starts the flow themselves. ENGIVERITY, Ad Pilot

Tools

Tools and knowledge from your modules

Each module has a fixed list of tools. Where an agent reads documents, it names the source.

Agent tools by module
ModuleWhat an agent can readWhat it can create or prepare
Documents Source documents, clarification cases, audit log, answers about the document archive Email drafts, clarification cases, tags, status
ENGIVERITY Bills of materials, review results, data sheet values, drawings, research in the knowledge base Open points, notes and reports, written only after approval
ERP and CRM Emails and attachments, documents, data room Drafts for incoming invoices in the customer-specific setup, replies, tasks from emails
Ad Pilot Figures from ads, search and shop Proposals for campaigns and bids, writing back into the account does not happen today

Control

Control stays with you

An agent gets only as much room as you give it.

Approval points In ENGIVERITY every step can have an approval point. It names the role that decides, and a rejection needs a reason. In Documents you can set up approval chains with deputies and the four-eyes principle. More on approval
Rights and limits An agent works within limits. You set which projects and object types it covers and whether it only reads or also writes.
Writing after approval In ENGIVERITY, steps that write wait in a queue until a person approves them. Nothing is written until a person has decided.
Levels and emergency stop Ad Pilot has levels of autonomy. Every new tenant starts on the strictest one, where every proposal needs approval. An emergency stop halts all changes at once.

Testing

Testing before the start

Before an agent works in daily operation, it runs as a dry run. The dry run shows which steps the agent would take and what results they would produce, without changing anything in the system.

In ENGIVERITY, document processing runs in shadow mode first. The agent reads and checks real documents and shows them to a person, while the handover stays a preview. A test week for each document type compares the result with the actual postings before an administrator switches to daily operation.

Dry run and shadow mode

Traceable

Log and source references

Where an agent reads documents, its answer carries a source reference. It leads back to the position in the document, the bill of materials or the data sheet that the statement comes from.

In ENGIVERITY, a check classifies every statement of an agent as supported, contradicted or without a source. The log records every start, every step and every tool call. It only grows, and nothing is changed afterwards.

Example

An incoming invoice, step by step

This is how the incoming invoice template runs, from the first reading to the handover.

01 Arrival The invoice arrives through the mailbox, a folder, the interface or an upload.
02 Reading A first reading and a second reading capture the fields independently of each other.
03 Checking Totals, tax, IBAN, VAT ID and possible duplicates are checked.
04 Presenting Uncertain fields go to a person together with the scan excerpt.
05 Approving The posting waits for approval, by two people if you wish.
06 Handing over The document moves on as a posting batch or file, and the log stays.

Operation

Models and operation

AI calls in the ERP run through a central layer, and changing the provider there is a setting.

The log records every call, meaning who triggered it, when, for which task and how long it took. The question and the answer are not stored.

Each company entity sets whether data may go to external providers. Where the other modules run their models and where declared external services are used is described on the <a href="/en/souveraene-infrastruktur">Sovereign infrastructure</a> page.

To the sovereign infrastructure

Limits

What agents do here and what they leave alone

One goal per agent Every agent in ENGIVERITY has one goal, a list of permitted tools and a step limit. It stops when the goal is reached, the limit applies or someone cancels.
Writing with a stop point An agent does not trigger a payment. You explicitly allow individual steps without asking, agent by agent.
People decide about people Where people are concerned, an agent at most prepares. The decision stays with you.
No model training on your confirmations In Documents, no model is trained on confirmations. Learning happens in a table per tenant, and every entry can be revoked.

Questions and answers

Questions about AI agents in Beyonetix OS

What is an AI agent, and how does it differ from a chatbot?

A chatbot answers questions. An agent works through a goal in several steps, uses the tools it is allowed to use for that, and can start on its own, for instance on a schedule or when a document arrives. At approval points it waits for a person.

Do I need programming skills?

No. In ENGIVERITY and in Documents you can put flows together without code. Anyone who wants to connect other systems uses a REST interface for that.

Can an agent change data or send emails on its own?

By default, no. In ENGIVERITY, steps that write are only queued until a person approves them, and the templates that ship with Documents prepare emails to third parties as drafts. Two exceptions are named: the daily summary goes out by email without approval, and one template hands a checked document over by webhook. In ENGIVERITY you decide per agent which steps it may carry out without asking.

Can I trace what an agent has done?

Yes. The log records every start, every step and every tool call, and answers point to their source in the original.

Which models run, and where is my data?

In the ERP, AI calls run through a central layer, and the log records every call without storing the question or the answer. The Sovereign infrastructure page describes where each function runs and when external services step in.

How can I keep costs under control?

In ENGIVERITY every agent has a step limit and a budget. In Documents a page allowance per tenant applies, in Ad Pilot a quota per tenant across all changes, counted by number and amount per hour and per day.

Which routine should an agent take over first?

We look at your process and find the template that makes a sensible start. It usually works best to begin with one narrowly defined routine.