Local LLMs for process auto­ma­tion – how lean AI models work more effi­ci­ently and what SLIMDOC has to do with it.

Every day, millions of documents end up in the systems of banks, insurance companies, and govern­ment agencies: invoices, contracts, damage reports, forms. Many of them are still reviewed manually – an enormous factor in terms of time and cost.

Arti­fi­cial intel­li­gence can change that. But the more powerful AI systems become, the larger, more expensive, and more resource-intensive they also become. This raises a crucial question: Do we really need an AI system that can do ever­y­thing – or is one that can handle our requests really well enough?

This is precisely the task that the SLIMDOC research project – a col­la­bo­ra­tion between Insiders and RheinMain Uni­ver­sity of Applied Sciences – is dedicated to.

AI is becoming more powerful – and more expensive

Modern AI language models such as GPT and similar systems are impres­si­vely versatile. They write texts, answer complex questions, translate languages, and solve pro­gramming tasks. But this ver­sa­ti­lity comes at a price: the models are getting bigger every month, and operating them requires enormous amounts of computing power and energy – and therefore money.

For companies that want to use AI for clearly defined tasks – such as auto­ma­ti­cally reading an invoice – an obvious question arises: Why should a model that processes documents also know how to cook pasta?

It doesn’t need to. But that’s precisely the problem: today, large general-purpose models are used that are hardly eco­no­mical to operate locally, or small spe­cia­lized models that are resource-efficient but also require time-consuming training with manually annotated data.

What is SLIMDOC – and why is it special?

SLIMDOC stands for “Syn­er­getic LIght­weight Mul­ti­modal DOCument Analysis” and is a research project at RheinMain Uni­ver­sity of Applied Sciences with a clear goal: AI models that analyze documents reliably – in a more stream­lined, faster, and more sus­tainable way than before.

The approach: “Knowledge distil­la­tion” is used to transfer the knowledge of large language models to small, task-specific models. The result is compact language models that can only do what they need to do – without cloud depen­dency, without unneces­sary resource con­sump­tion, and with full data pro­tec­tion.

Since documents rarely consist of text alone, SLIMDOC relies on mul­ti­modal analysis: text, images, and layout are evaluated together. Spe­ci­fi­cally, two use cases are planned—analysis of annual reports and plau­si­bi­lity checks of insurance claims—in col­la­bo­ra­tion with R+V Ver­si­che­rung and Doxis.

Why we are involved

At Insiders, we have been auto­ma­ting document-centric business processes for more than 25 years. Whether hand­written forms, complex tables, scans, or digital PDFs, our software reads, under­stands, and processes all types of documents.

We have long relied on a com­bi­na­tion of our own AI tech­no­logy and large language models. And we are incre­asingly reco­gni­zing that the direction in which AI models are deve­lo­ping will sooner or later become a serious problem for practical use.

That is why we are actively involved in research and are a key practical partner in the SLIMDOC project.