Access
To your systems – ERP, CRM, document storage, telephony – with the permissions of its role.
Product
AI employees: role-based AI agents inside your systems
One AI agent per role that works in your ERP, CRM and telephony: with memory for your rules, with access to your systems, with permissions like an employee, and with a human who approves.
Your AI team. Your brand. Your data.
Servers in Germany, critical data offline · AI advises, a human decides · Transparency under the EU AI Act
SubjectAI employee
SheetC-01
Date24 September 2026
DrawingW&K Connect
Does this sound familiar?
- You bought an AI tool. It answers questions, but it doesn't do anything: it doesn't create an order, doesn't check an invoice, doesn't remind anyone of a deadline.
- A provider set up a chatbot for you. After three questions it no longer knew anything about your company, and it never learned your rules.
- The knowledge of how to create an order correctly, which customer gets supplied first and when an approval is needed sits with two people.
- New employees need weeks before they can work without asking questions, and until then they ask the boss.
What an AI employee is – and what a chatbot is not
A language model on its own knows nothing about your business. It doesn't know your customers, your prices or your rules. What it says sounds good and is often wrong. An AI employee needs four things that no off-the-shelf tool brings with it:
Memory
For your rules, your corrections and your decisions. It learns what you teach it and does not forget it.
Limits
What it may do, what it has to ask, what it never does.
Logic
That comes from your process, not from a template.
We build these four things. The language model is the most interchangeable part.
Stage 1
Chatbot
answers questions.
Stage 2
AI assistant
supports a person in their work.
Stage 3
AI agent
carries out tasks in systems.
Stage 4
AI employee
an agent with the tasks, permissions and limits of a role in your company.
Role catalogue
One AI employee per role: what it does – and what it never does
Typical tasks, not customer cases – example scenario.
-
01
Management
Does: Overview of open cases, bottlenecks and upcoming decisions; sees what the employees' agents have learned and corrects rules in one place.
Never: Make decisions on money, contracts or staff.
-
02
Purchasing
Does: Prepare orders, compare supplier prices, track delivery dates, match goods receipts against orders.
Never: Place an order above the approval threshold without a human.
-
03
Warehouse
Does: Handle goods receipt with scale and label, check stock levels, trigger order picking, enter batches and best-before dates (MHD) into the system.
Never: See purchase prices or revalue stock.
Food & logistics -
04
Sales
Does: Record enquiries, prepare quotes from pricing logic and history, create orders, remind the team to follow up on open items.
Never: Promise special prices or payment terms without approval.
-
05
Accounting
Does: Read and assign documents, check completeness, prepare payments under the four-eyes principle, compile supporting documents into a report.
Never: Approve a payment or complete a posting without the required supporting document.
Trade & precious metals -
06
Driver
Does: Show the driver's own route, customers and their rules, report deliveries, pass deviations on to dispatch.
Never: See purchase prices, margins or other drivers' routes.
-
07
Phone
Does: On a call, pull up the customer record and their open cases, file the call note on the case, create callbacks as tasks – together with your phone system.
Never: Make commitments on the phone or hold calls without identifying itself as AI.
What the AI employee may do, what it remembers, where it stops
Permissions per role. Each agent has exactly the permissions its role needs, no more. The driver sees no purchase prices, purchasing sees no salaries, accounting sees no route planning. The permissions come from the permissions and roles concept in your ERP, not from a setting in the chat.
Memory per role. Each agent has its own memory: your rules, its corrections, the decisions it has seen. Management sees what their employees' agents have learned and corrects things in one place. The memory belongs to you and stays with you, kept separate for each customer.
Limits and approval step. When the agent is unsure, it does not decide – it asks. When an action touches money, contracts or staff, it puts it to a human for approval. This is not a note in the manual but a step in the system: nothing happens without approval. Payments, contracts and staff are always decided by a human.
What happens if it makes a mistake? It can only do what its role allows, and only up to the approval threshold. Anything beyond that remains a proposal until a human approves it. Corrections go into its memory. Every action is logged; you can see what it did and when.
Every AI employee identifies itself as AI (EU AI Act Art. 50). We do not use AI that evaluates employees, recognises emotions or makes personnel decisions: that is prohibited or high-risk and does not belong in an SME project. → AI transparency
Operation
Your data stays in-house: how we operate AI employees
On your infrastructure
If you have servers in-house or want to, we operate the agents there, with monitoring by us or your IT.
Language models as processors
The models come from established providers that are contractually bound under Art. 28 GDPR. Critical data does not leave your environment.
Which model suits you is decided by your process and your IT, not by our catalogue. Where AI works on our website and in our products: → AI transparency · → Privacy policy
Proof without names
- AI employees with their own memory and their own permissions work today in more than 20 roles, from driver to managing director.
- We operate what we build, and we work with AI employees ourselves in management and accounting, on our own infrastructure.
- Example scenario, one AI employee per role: management, purchasing, warehouse, sales, accounting, driver – each role has its own agent with its own memory and its own permissions in the ERP. Management sees what the agents have learned and corrects rules in one place. The new colleague asks the agent, not the boss.
Implementation
How the AI employee comes into your company
-
01
Role and logic
Which role, which tasks, which systems, which limits? Who approves? We write it down, you sign it off.
1–2 weeks
-
02
Connect
Permissions from your ERP, memory set up, systems connected – in stages, and after each one you see something that works.
In stages
-
03
Test
Before go-live we test, using records outside its role, that the agent only sees and does what its role allows.
Permissions test
-
04
Prove
Before: time per case, queries, errors. After: the same numbers. Then you decide on further roles.
Measurement
-
05
Operate
Updates, new rules, new roles – with us or on your infrastructure.
Operation
More on custom development, of which the AI employee is a building block: → Custom software & AI agents
Frequently asked questions about AI employees
What is the difference between an AI assistant, an AI agent and a chatbot?
A chatbot answers questions. An assistant supports a person. An agent carries out tasks in systems. The AI employee is an agent with the permissions, the memory and the limits of a role in your company.
What happens if the AI employee makes a mistake? Who stops it?
The AI employee is stopped by its role and the approval threshold. It can only do what the role allows; anything beyond that remains a proposal until a human approves it. Corrections go into its memory, and every action is logged.
What access and permissions does the AI employee get?
An AI employee gets exactly the permissions of its role in the ERP or CRM, no more. Before go-live we test against records outside its role that it sees nothing it is not entitled to.
Does this work with Odoo, Bitrix24, DATEV or our phone system?
With Odoo and Bitrix24 in live operation; DATEV via the document export; phone systems via their interface. Anything else you use we check in the first call.
Do we need additional licences?
You don't need any licences for your ERP or CRM beyond the ones you already have. Language models and operation are part of the offer.
Does AI replace our employees?
No. It takes routine work off their hands: looking things up, gathering information, retyping. It also puts decisions in front of them. Your employees go from clerks to decision-makers. And a new colleague is onboarded faster because they ask the agent.
What if an employee leaves? Does the knowledge stay in the system?
Yes. The rules, corrections and decisions are held in the memory of the agent for that role, not in the person's head. The successor keeps working with the same agent.
How do we measure whether the AI employee is worthwhile?
We measure the benefit before and after with your numbers: time per case, queries to management, errors, onboarding time. Before the pilot we define what is measured, and afterwards we present the figure.
Can we start with one role?
Yes, and that is exactly how we start. A pilot begins with one role: we write down its logic, connect the systems and measure before and after. Only then do you decide on further roles.
First call
Describe the role you want to take work off
In the first call we go through one day in this role: tasks, systems, queries, approvals. Afterwards you know what an AI employee can take over there and what a pilot with one role looks like.
30 minutes, one role from your day-to-day work, no presentation.