Where AI takes work off your hands — and where it does not
Most AI projects start with the question of where to put the chatbot. That is rarely the place where AI saves money. An assistant on the home page answers questions that a decent search and a proper FAQ would answer too — and produces answers you are liable for. The value almost always sits in the background: in the repetitive donkey work someone does by hand today, because it is too irregular for a script and too dull for a specialist.
In concrete terms: product descriptions, attributes and SEO copy for thousands of items in the PIM, instead of typing them for months. Invoices, delivery notes and orders turned from PDF and email into structured data. Incoming enquiries and tickets classified and routed to the right person. Translations for multilingual shops that an editor then checks instead of writing from scratch. These are unremarkable tasks. That is exactly why they pay off.
The function belongs inside the system, not next to it
An AI function that lives in a separate tool stops being opened after a few weeks. So we build it where the work already happens: as a button in the Shopware back end that suggests missing product copy. As a field in TYPO3 that drafts a meta description. As a step in your order processing that reads an incoming order PDF and hands it to the ERP. Nobody has to learn a new tool or keep a second browser tab open.
Technically this is integration work, not model research. We work with PHP 8, Symfony and Python, address the models over their APIs and put the results where they are needed. For semantic search in the shop — customers search for their problem, not for your article number — we combine vector search with Elasticsearch and your existing range. Internal knowledge search follows the same principle: the model answers from your documents, with the source named, not from memory.
Data protection, the EU AI Act and where the servers stand
As soon as customer data, invoices or personnel records are involved, choosing a model becomes a data protection question. Before implementation we establish which data may leave the building and which may not. Then the decision follows: a provider with a data processing agreement and processing inside the EU, or an open-weight model that runs on our own servers in Germany and sends nothing outside at all. The local model is somewhat weaker than the largest offering on the market. For classification and extraction it is usually good enough.
The EU AI Act places obligations on companies whether they build models themselves or only use them. Our typical cases — product copy, classification, document extraction — do not fall into the high-risk category. What does apply is the AI literacy duty under Article 4: anyone deploying such systems has to make sure the people using them are competent to do so. The transparency duties under Article 50 cover particular systems, for example those that interact directly with people or generate synthetic content. It gets more serious where systems decide about people, in recruitment for instance — that is classified as high risk. We document which model does what with which data, and record where a person intervenes. The legal assessment stays with your data protection officer or your lawyers.
A person checks before anything goes out
Language models make things up. They do it fluently, in complete sentences and with no sign of doubt — an awkward property you plan around rather than argue away. Our rule: anything that goes out to the world or moves money is seen by a person first. Product copy enters the approval process as a draft. Extracted invoice data gets a confidence score; anything below the threshold lands on a review pile. The model does the donkey work, the responsibility stays in the house.
That is also why we start small. We look for a process that runs often, can be described clearly and costs time today — and build a prototype for it on real data from your system, not on samples. Then we measure: how many cases go through without correction, where does someone have to step in, what does a month of operation cost. If the result holds, we roll it out and take on the next process. If it does not, we say so — before a year’s budget depends on it.