AI agents for mid-sized companies Use and training
An AI agent handles a defined task on its own. It reads an enquiry, fetches data from your ERP or CRM and puts a draft forward for approval. We build such agents for mid-sized companies and train your teams to work with them.
On site or online · For managers, specialist departments and IT
What is an AI agent?
An AI agent is software that works with a language model and, within a defined task, decides for itself which step to take next. To do so it is given tools: a search in company knowledge, a query in the ERP, the creation of a draft. You decide which tools these are and what the agent may do with them.
A chatbot answers questions; an agent completes work steps. Of the companies that use, plan or discuss AI, 11 per cent already use such agents. A further 29 per cent are planning to. That is the result of a survey of 603 companies by Bitkom Research in July and August 2026.
Automation, AI workflow or AI agent
Before we build an agent, we check whether a simpler process will do. It costs less to run and is easier to check.
Classic automation
Works through fixed rules, for example: invoices from this supplier go to that cost centre. Reliable as long as nothing unexpected comes in.
AI workflow
A fixed process in which AI takes over individual steps, such as reading a PDF or sorting an email. The order of steps is set. For most office processes, that is enough.
AI agent
Chooses for itself, within its task, which tool to use and in what order. Worth it when cases differ a lot from one another and have many exceptions.
Multi-agent system
Several agents with separate tasks work together. More agents do not make a system more reliable, so we only use this when a single agent is not enough.
Where AI agents help mid-sized companies
Five processes in which agents work well. In each one the agent puts forward a draft, and a person decides on anything with consequences.
From enquiry to draft quotation
The agent reads customer enquiries from email and PDF, recognises items and quantities, looks up prices and delivery times in the ERP and flags missing details. If an enquiry is unclear, it drafts a follow-up question to the customer.
We measure minutes per enquiry, time to draft and rework. Prices, delivery dates and sending the quotation are decided by a person. Item master data and prices must be maintained in the ERP and available through an interface.
Incoming documents with many exceptions
Delivery notes, order confirmations and invoices arrive in many formats. The agent recognises the document, reads the fields and checks them against the purchase order and goods receipt. Where they differ, it gathers the relevant documents for clarification.
The benefit shows in errors per field, processing time and the share of documents without rework. Posting, payment release and any change to bank details stay with a person. Purchase orders and goods receipts must be digital. Structured e-invoices are processed directly, without text recognition.
Service and technical knowledge
Service technicians and the back office ask in plain language. The agent searches manuals, maintenance reports and earlier tickets, names the source of its answer and says so when it finds nothing reliable.
Metrics are search time, first-contact resolution and the accuracy of answers. The reply to the customer and any intervention on a machine remain the responsibility of a person. Documents must be digital, and access rights also apply in the search.
Purchasing: compare offers, monitor dates
The agent puts supplier offers side by side, checks terms and delivery times, prepares follow-up questions and reports when a promised delivery date is at risk.
We measure the time for the comparison, completeness and how early delays are spotted. Ordering and awarding are decided by a person; the agent does not place orders itself. Suppliers and purchase orders must be recorded in the ERP.
IT service
The agent pre-sorts incoming tickets, searches runbooks for known fixes and prepares the diagnosis so that IT can start right away.
Metrics are handling time and the share of tickets solved at first contact. Any action with administrator rights is approved by a person. It requires a ticket system with an interface.
When an AI agent is not worth it
If a process always runs the same way, an agent is more expensive and harder to check than a fixed workflow. This applies to many standard processes in accounting and order processing.
Where a wrong decision could harm people or machines, we advise against agents. This concerns the selection of applicants, medical questions and the control of equipment.
We leave calculations, mandatory field checks and posting rules to fixed program code. The language model reads, classifies and writes.
How we safeguard AI agents
An agent that acts on its own needs fixed rules about what it may do. We build these rules into the technology, because a written instruction alone does not stop an agent.
- Each agent has only the rights its task needs, with its own account instead of an administrator account
- Sending, ordering, posting and payment are always approved by a person
- Every step is logged: what the agent read, queried and proposed
- Run time, number of steps and cost per case are capped
- Before go-live and before every model change, we test with reference cases
Emails and documents can contain hidden instructions meant to make an agent do something else. Germany’s federal cyber security agency BSI describes this risk as indirect prompt injection. That is why the agent treats incoming content as data and never as an instruction.
If you wish, agents process sensitive data with language models on your own hardware in the AI Cube. Documents and inputs then never leave your network. When an agent talks to customers, the EU AI Act requires that it can be recognised as AI.
The tools we build agents with
First we check what your existing software can already do. If you use SAP, Joule has its own agent functions; if you work with Microsoft 365, there is Copilot Studio. Often it is enough to set up these functions and connect them to your data.
For processes across several systems we use n8n, a platform for automation and agents that can run on your own servers. With language models in the AI Cube, agents run entirely in-house. We choose the tool for each process.
AI agent training for your teams
Your teams learn on their own tasks, with the tools you have already licensed. According to Bitkom Research, 66 per cent of companies rate the AI literacy of their employees as low.
Understanding AI agents
Half a day · management, team leaders, head of IT
What agents can do today, where they fail, what they cost to run and which rules apply. Together we go through your processes and assess where an agent fits and where a simple workflow will do.
Result: a list of up to three processes suitable for a pilot, with their prerequisites.
Building your own AI assistants
1 day · specialist departments and key users, no programming skills needed
Each participant builds an assistant for a task from their own working day, such as draft quotations, meeting minutes or searching manuals. We work with your tool, for example Microsoft Copilot, ChatGPT Enterprise, Langdock or a local AI.
Result: one working assistant per person and instructions on how to maintain it.
Agent workshop for IT and key users
2 days · IT, process owners, technically skilled key users
We build an agent with n8n or Copilot Studio that queries data from a test ERP, creates a draft and waits for approval. This includes permissions, logging, error handling and tests with reference cases.
Result: a documented prototype with test data that your IT can develop further.
How the training works
- At your premises or online, usually with 6 to 12 participants
- Exercises with your processes and documents; we replace confidential data with test data
- If you wish, we help write the AI policy, i.e. what employees may do with AI
- Certificate of attendance listing the topics covered
The formats build on one another, but you can book each one on its own. You receive the content, date and price in writing in advance.
Under Article 4 of the EU AI Act, companies that use AI must take measures to ensure the AI literacy of their employees. The law does not prescribe a specific certificate. You can use our certificate of attendance to document the training.
From the first conversation to a running agent
After each step there is a result, and you decide whether to continue.
Initial consultation
About 30 minutes, free of charge. You describe a process, and we assess whether an agent, a simple workflow or no AI at all is the right fit.
AI potential check
Up to three processes, with data, interfaces, expected benefit and risks. You receive the result in writing.
Pilot
An agent for a clearly defined process, tested on real cases. This often takes four to six weeks if data, interfaces and a contact person are available on your side.
Rollout
Connection to production systems, rights and approvals. Your teams learn to work with the agent.
Operation
We monitor the agent, install updates and test new models before they go live.
Frequently asked questions about AI agents
What is an AI agent?
Software with a language model that handles a defined task on its own. It decides which tool to use next, for example a search in documents or a query in the ERP, and submits the result for review.
What is the difference between a chatbot and an AI agent?
A chatbot answers questions in a conversation. An AI agent completes work steps: it retrieves data from systems, creates drafts and triggers processes, as far as its rights allow.
Is an AI workflow enough, or do we need an agent?
For most processes in mid-sized companies, a fixed process in which AI takes over individual steps is enough. It is cheaper and easier to check. An agent is worthwhile when cases differ widely and have many exceptions.
Can AI agents send emails or place orders themselves?
Technically they can. We set agents up so that a person approves everything that goes outside the company or costs money, that is dispatch, ordering, posting and payment. Every step the agent takes is logged.
Which systems can an AI agent use?
All systems with a documented interface, for example ERP, CRM, document management, ticket systems and Microsoft 365. More on the AI process automation page.
Can AI agents work without data leaving the company?
Yes. With language models on your own hardware in the AI Cube, documents, inputs and answers stay in your network.
What does an AI agent cost?
That depends on the process, the interfaces and how it is run. The initial consultation is free; afterwards you receive a written proposal with scope and price.
Do participants need programming skills for the training?
Not for “Understanding AI agents” and “Building your own AI assistants”. For the two-day agent workshop, participants should come from IT or be confident with interfaces and data formats such as JSON.
Does the training count as proof of AI literacy under the EU AI Act?
Article 4 obliges companies to take measures to ensure the AI literacy of their employees. The law does not prescribe a specific certificate. Our certificate of attendance lists the topics covered, so you can document the training.
Do you come to the company for training?
Yes. We train at your premises, and all formats are also available online.
Choosing a process for an AI agent
Describe a process with many exceptions. In the free initial consultation we tell you whether an agent fits and what training your team needs for it.
