Administration & purchasing
Check invoices against order and goods receipt, prepare quotes from enquiries and bills of quantities, compare supplier offers, answer internal questions from manuals, each with a source.
An AI agent handles a task from start to finish: it reads an invoice, matches it against the order and goods receipt, and proposes the posting. It answers a service enquiry from your documents and hands over to your team as soon as it gets sensitive. We develop such agents, run them on our own infrastructure and build in a sign-off wherever a decision is made.
Contact: Juri Hoffmann, certified AI manager (MMAI®)Updated:
Check invoices against order and goods receipt, prepare quotes from enquiries and bills of quantities, compare supplier offers, answer internal questions from manuals, each with a source.
Answer enquiries by email, chat or phone from approved documents. Classify website enquiries and pass them to the CRM with a summary. Difficult cases go to your team together with the history.
Suggest posts from your stock of topics and dates, generate text and image, publish on schedule after approval. This is how the editorial agent at the IHK-Bildungszentrum works.
AI avatars for advice and reception, explainer videos without a shoot, translation during a live conversation. Always recognisable as AI, as Art. 50 of the EU AI Act requires.
An example workflow in n8n. Every step is visible and can be adjusted. Your team only steps in when something does not match, and approves before anything is booked.
A chatbot answers questions. An agent acts: it pulls data from the mailbox, looks things up in the ERP or CRM, writes a draft and puts it in front of someone for approval. The difference is not the language model but everything built around it: the connection to your systems, rules for what it may do, a log of every step and a point where a person agrees. That part decides whether an agent stays in operation or disappears after the demo. That is exactly where we start.
A chatbot answers, an agent gets things done. It is connected to your systems, works through a task in several steps and puts the result forward for approval or passes it on. Answering a question is only one part of the work.
Only where you allow it. For anything with external effect or money flow, meaning posting, ordering or publishing, a sign-off sits in between. Every step is logged, so you can trace what the agent did and why.
With most that have an interface: email, merchandise management and ERP, CRM, Microsoft 365, shop systems. We build the workflows mostly in n8n, which we run ourselves. Where there is no interface, we check whether an export or a mailbox is enough as a handover.
On servers in Germany or on your own hardware. Where contracts, customer data or calculations are involved, we use local language models. Then the data does not leave your premises. A data processing agreement under Art. 28 GDPR is standard.
With one task and a pilot. Beforehand we measure how long the task takes today, afterwards how long it takes with the agent. Only when that pays off does the next one follow. Each agent starts on its own and can be connected to others later.
Not in our projects so far. It takes over repetitive work nobody was hired for and frees time for what your people are actually there to do. Concept, decision and responsibility stay with people.
Two things affect almost every use: people must be able to tell that they are talking to an AI (Art. 50), and anyone using AI must equip their team to handle it (Art. 4). We label our agents and avatars accordingly and brief your team.
A first conversation takes half an hour and costs nothing. If you would rather have the existing site checked first, the website audit is the smaller way in.