Local AI for document batches

In the morning the employee needs a ready list

The documentation team receives a set of reports and prepares short descriptions for an internal catalogue. The material stays in the company's environment. The owner picks the files to process, and the next day opens the list of descriptions. Next to one report they see that an attachment could not be read, and they deal with that exception without going through every file.

A description is a proposal for approval. The employee checks it against the original, corrects a missed topic and only then passes it to the catalogue. An unfinished document stays on the list with a status: the file arrived and is waiting to be processed. If a document may have changed in the evening, the app shows which version it processed.

What local batch processing costs and when it pays back

The count covers the time spent searching, reading and compiling the material. What stays is the time to check the output against the source, 25% of today's work in the scenario. We measure both times in a trial on 20 to 100 cases the team has already handled. In our measurement Qwen3.8-27B on an RTX 3090 processed 230,085 tokens in 563.4 seconds. In batch work that time does not get in the way, because nobody waits for the answer at the screen.

What runs locally and what needs agreeing

vLLM describes processing without a separate serving endpoint and queuing jobs for later collection. Those are capabilities of the engine. The batch view, linking descriptions to files and handling missing results belong to the app you order. The term „offline inference” in that documentation does not mean the whole flow is cut off from the network. Repeating interrupted work leaves the employee a clear list of results to assess.

When your current document flow is enough

If the description comes from form fields, you can build it from a template without a model. If the reports already contain a summary, it is enough to fetch it. A model is worth assessing when an employee reads varied texts and writes each description from scratch. The trial also shows the reviewer's work: many proposals at once create a queue for checking that processing files one by one used to hide.

Where to start

Before you expand, look at one real batch of approved material, including a file whose processing failed. On that basis we set the collection time and the reviewer's workload. We roll out local models as part of private LLMs, and we build the view for working with sources as an AI app. For the free process review, a description of one case and the places where the material sits is enough. Confidential files stay in the company, and you get a written plan within 2 working days.