Take a furniture exporter in the Savinja Valley. Retailers in Germany, Italy and elsewhere email their enquiries: one attaches a spreadsheet of item codes, another a phone snap of a catalogue page with products ringed in pen, a third just writes “like last year, but oak and no drawers”. Two sales administrators open each message, look the products up, check stock and lead times in Pantheon and type the quote by hand. By Monday lunchtime half a day's backlog has built up, and a single misread dimension can lead to a complaint worth more than the order. The company does not need “an AI”. It needs a ready draft quote, complete with catalogue codes, waiting beside the original email so the administrator can compare and fix it. That is what this service delivers. We choose one workflow, put a price on today's time and errors, test models against enquiries you actually received, and let the winner shadow the team for several weeks. Integration with your software comes only once the numbers stand up. Where customer data is processed, and under what contract, is settled in writing on the first day. We meet by video and configure everything over remote access.
Choosing the model is a modest share of the effort. The bulk of the project is made up of the elements below, weighted differently for each workflow.
Sitting with the staff who handle enquiries daily, we follow one from inbox to reply, noting where it stalls, what gets copied by hand and where errors arise. Every stage gets a stopwatch, so the gain can later be stated in hours instead of opinions.
The model reads products, quantities and wishes from the message or its attachment. Your business rules then confirm the code exists, the variant is buildable and the quantity makes sense. Plain checks like these intercept most model mistakes.
Enterprise services from big vendors processing in the EU, and open models able to run inside your own cloud tenancy, are fed an identical batch of cases. We weigh accuracy, cost per case and processing location.
The administrator sees the incoming message and the suggested draft together. Uncertain fields are coloured, fixing one is a single click, and each fix is kept as evidence for quality tracking.
Results appear in tools the team already has open: a draft quote in Pantheon or Metakocka, a prepared reply in Outlook, a labelled request in your helpdesk. Nobody has to learn yet another window.
A data processing agreement with the model vendor, the processing location, any third-country transfer, an entry in your record of processing activities and, if required, a revised privacy notice. The legal judgement belongs to your adviser or DPO; the technical detail comes from us.
A short report each month covering accuracy, the fields most often corrected and cost per case. If retailers start sending enquiries in a new format or the vendor swaps model versions, a dip in quality is visible straight away.
How long it takes depends on the workflow and the condition of your data. We review results with you at the end of each stage, and you choose whether to proceed.
Over a fortnight we count cases, clock the minutes each takes and classify errors. The goal is set from these numbers.
A handful of models and configurations are run against old cases. Accuracy, minutes saved and running cost arrive in one table.
Live cases go through the model while staff work exactly as usual. Outputs are compared, and no record is altered.
Start with one customer segment or one kind of case, then widen. Sign-off by a person remains compulsory until quality is proven.
Prices, discounts and delivery dates never come from the model. A language model is very good at grasping what a customer is asking for, yet it knows nothing about today's price of an oak panel or the date of the next production run. Let it guess and the reply will read confidently and be wrong. So the model only assembles the products and quantities, while prices and dates are fetched straight from your ERP by rules you define.
What we usually see is the role shifting rather than vanishing. People move from retyping to checking, sorting out exceptions and talking to customers. Involve them early, above all when charting the process and shaping the approval view, and adoption becomes much easier.
It depends on the terms of the specific service. Enterprise offerings generally exclude your inputs from training by default; free consumer accounts may not. We go through the terms with you at the outset and give you a written summary.
The integration is designed to be vendor-agnostic. Since you keep the benchmark of past cases, any replacement model can be scored on the very same examples in a few days, and switching becomes a decision based on evidence.
Starting figures and the prototype normally fit into a few weeks, and the prototype table on its own tells you whether to continue. Shadow mode and launch take longer or shorter depending on case volumes and the system being connected.
Either a fixed quoted price or €50 per hour plus VAT against an estimate agreed beforehand. We forecast model usage before starting, and you pay the vendor for it directly. Whether your project fits a digitalisation voucher from the Slovenian Enterprise Fund or another scheme is for you to check with whoever runs it; we neither write applications nor promise grants.
Tell us what the job is, roughly how often it comes up each month and which system it is handled in. You will hear our first view within one working day.
Enquiry received
An engineer will write back no later than the following working day. Faults that leave your staff unable to work are always looked at first.
No town by that name here. Check the letters, or settle for the nearest larger place. Our engineers work only over the internet, so Kočevje and Koper get precisely the same treatment.
