AI integration

The jobs we are asked to solve most

None of this is hypothetical. These are the requests that retail, manufacturing and service businesses bring to us most often:

• Reading invoices, consignment notes and delivery reports: supplier, VAT number, date, line items, totals and VAT are extracted and filled in, and a person just checks them.
• Classifying documents and email: incoming mail is sorted by type — order, complaint, enquiry, invoice — and lands with the right person.
• Search by meaning in a catalogue or knowledge base: “rubber seal for a pump, about 40 mm” finds the right part without the exact code.
• Product descriptions for an online store, written from the supplier's technical data in Bulgarian and other languages, reviewed before they go live.
• Summaries of long documents — contracts, tender files, meeting minutes — with references back to the original.
• A voice assistant for situations where people's hands are busy: in the warehouse, on the production line, in the van.

How it fits in technically

The AI part almost always sits behind a separate service rather than inside your core system. That service takes a document or a question, calls the model — Gemini on Vertex AI, a dedicated document-recognition model or something else, depending on the job — and returns a structured result. Your core system receives clean data over an API, exactly as if a person had keyed it in. If you later want to change model or provider, the change happens in one place.

For documents we add checks that do not depend on the model at all. Line items must add up to the total. The VAT number is checked against the register. The date cannot be in the future. Anything that fails goes to manual review instead of slipping quietly into the books. Each field also carries a confidence score, so the reviewer looks only at the doubtful spots rather than reading the whole document.

How the work goes

We start with a trial on your real data. If the job is invoice reading, we ask for a hundred or so invoices from different suppliers, including the badly scanned ones and the odd formats. We run them through and show you the result field by field: where it is right, where it slips and why. You have something solid to decide on before paying for a full rollout.

After the trial we build the feature into your system, add a screen for review and approval, and train the people who will use it. For the first few weeks we watch every correction they make and tune the solution accordingly. If the system is one we built, this goes faster. If it is someone else's, we work through its APIs or an import it already supports.

Data, privacy and running costs

Invoices and contracts carry sensitive information. We use services where your data is not used to train the model, processing can stay in an EU region and documents are not kept longer than needed. If your requirements are stricter, we can discuss a model that runs on your own server.

Model costs are charged per document or request, and are usually small next to the time saved. We estimate them up front from your actual volumes, and the admin panel shows what has been used each month.

After rollout

A new supplier with an unfamiliar invoice layout, a new kind of email, a new product range — it all happens. Support covers watching the share of manual corrections, adding new formats and checking quality whenever the model changes. If corrections start creeping up, we notice before you do.

Once one feature runs reliably, the next one usually costs less, because the connection to your system, the review screen and the checks are already in place. That is why many clients start with invoices and add email sorting or catalogue search a few months later.