A custom model: a private LLM that drafts, classifies and extracts the way your team does

The problem

RAG search answers the question "what do our documents say". A lot of work in a company is not search. Support writes fifty similar replies a day that have to sound like the company, not like a robot. The office sorts incoming e-mail into ten categories by hand. Someone copies the same fields from every purchase order into the ERP.

General models do these tasks reasonably well, but not in your style, not by your categories and not always in the same format. Long instructions and examples in the prompt get you a long way. When the model has to repeat a task a thousand times a day, exactly the same way, and the content may not go to a foreign service, further training of your own model is a sensible step.

What the private AI does

Fine-tuning means further training an open language model on your examples. We use LoRA, a method that trains a small additional part of the model, so training is faster and cheaper and the base model stays unchanged. The result is a model that runs on your server and does the task the way your team used to do it.

The three tasks where fine-tuning most often pays off:

• Draft replies in Slovenian (or English) in the style of your support team or office, which a person checks and sends.
• Classification of tickets, e-mail or documents into your categories, with a confidence score.
• Data extraction from purchase orders, invoices or contracts into a fixed structure for the ERP or a spreadsheet.

A sample task and result (a made-up ticket, shown for format):

Input: "Hello, we received the parcel yesterday but the adapter is missing. Order number 48213. Regards, Mojca"

Model output: category Complaint: missing part, confidence high; extracted: order 48213; draft reply: "Hello Mojca, thank you for letting us know. We are sending the missing adapter for order 48213 today, and it should reach you by post within 1 to 2 working days. We apologise for the inconvenience."

Source: model trained on 2,000 past support replies (Support/export_2024-2025.csv); categories from Support/categories_v4.xlsx

With fine-tuning, the source is not a paragraph in a document, as it is with RAG, but the training set. That is why we document clearly, for every model, which data it was trained on.

Data sources

Fine-tuning needs pairs of "input and expected output", which companies often already have:

• Support tickets and replies from the ticketing system or an e-mail export.
• Classified e-mail or documents, where a person once assigned the category.
• Letters and quotes that show the company's style and structure.
• Manually transcribed data, such as purchase orders and ERP entries, where we have both the document and the correct result.

How much data is enough depends on the task. Classification into a few categories needs fewer examples than writing in a house style; we set the exact number in the pilot. Consistency matters more than volume: if two colleagues classified the same ticket differently, the model will learn both. That is why we review part of the examples by hand before training and resolve the inconsistencies with you.

We always hold back part of the data for evaluation, so the model is tested on examples it did not see during training. Before training we clean the data: we remove personal data the model does not need and examples that are not a good model to follow.

Architecture

1. Data preparation: export, cleaning, removal of personal data, split into training and evaluation sets.
2. Baseline: we first measure how well the base model does the task with good instructions. If that is enough, we do not fine-tune.
3. LoRA training: on your server, or, if it is not powerful enough for training, on rented hardware under your control, by agreement.
4. Evaluation: comparison with the baseline on the held-out examples, using a measure agreed in advance (classification accuracy, the share of drafts the team sends without major edits).
5. Deployment: the trained model runs on your server behind an OpenAI-compatible API, so you can connect it to the ticketing system or ERP.

The model can be combined with RAG search: fine-tuning takes care of style and format, RAG of facts from your documents. We compare the two approaches on the private AI page.

Security and regulation

Training data is often more sensitive than documents for RAG, because it is real communication with customers. So the rule is: data is prepared and the model is trained in an environment you control, and the trained model stays on your server.

A model can memorise parts of its training data and repeat them later. We reduce that risk by cleaning the data before training and by checking outputs before deployment. The General Data Protection Regulation (GDPR) applies to customers' personal data also when you use it to train a model.

Timeline and price

• Pilot (from 2,900 €, 2–3 weeks): a review of the data, a baseline measurement with the base model and a decision on whether fine-tuning is needed at all. Often good instructions plus RAG turn out to be enough.
• Fine-tuning (4,000–12,000 €): data preparation, LoRA training, evaluation and deployment. The range depends on the number of tasks, the volume of data and how much cleaning it needs.
• Support (290–790 € per month): quality monitoring and retraining once new examples accumulate.

A fine-tuning pilot is above all a measurement. In the first week we export a sample of the data, check its quality and agree with you how success will be measured: for classification, the share of correctly classified examples; for drafts, the share the team sends without major edits. In the second week we measure how the base model handles the task with good instructions and, where it makes sense, with RAG search. In the third week you get a report with the results and a clear recommendation: what we already have is enough, or fine-tuning is worth the investment, together with an estimate of how much data it would need.

Prices are final; the provider is not VAT-registered. What each tier covers is on the private AI for business page. An example where fine-tuning complements document search is private AI for accounting firms, where the model drafts client replies in the firm's style.

What we do not promise

• Fine-tuning when it is not needed. If good instructions and RAG solve the task, we tell you after the pilot and do not sell extra training.
• New knowledge from training. Fine-tuning improves style, format and classification. For facts from documents, RAG is the better tool.
• Flawless output. A trained model still makes mistakes. That is why a person reviews drafts, and classification has a confidence threshold below which a case goes to manual review.
• The quality of the largest cloud models on every task. A trained open model is often very good at a narrow task and weaker than the largest cloud models at general writing and reasoning.