This article describes three components of automated document processing: the large language model, the highly specialised model and the validation and automation layer. The large language model and the specialised model read content from documents. The validation and automation layer checks the values read, routes them and logs the processing. The final section shows how Parashift covers the three components with its own products, including the measurements from September 2026.
Definitions
A large language model is a general purpose model that processes text and images and is accessed through a provider’s programming interface. It is not specialised in documents but is frequently used to extract fields from documents because it can be deployed without preparation.
A highly specialised model is a model trained exclusively on business documents. It extracts fields, classifies documents, triages them and checks plausibility. Such models are considerably smaller than general purpose models and are typically operated on own hardware or in a private cloud.
The validation and automation layer is not a model. It manages specifications per document type, evaluates a model’s output, checks it against rules, routes it and logs the processing. It is model independent and works with both model types.
Comparison of the three components
| Large language model | Highly specialised model | Validation and automation layer | |
|---|---|---|---|
| Category | general purpose model | model limited to business documents | orchestration layer, model independent |
| Task breadth | very broad | narrow | not applicable |
| Accuracy on document tasks | high, depending on document type | higher than general purpose models in measurements | not applicable |
| Throughput | depends on provider and load | predictable, depends on hardware | not applicable |
| Context budget | large | small, sized for the task | not applicable |
| Place of operation | provider infrastructure | own data centre or private cloud | cloud or own data centre, depending on vendor |
| Cost driver | tokens and page count | hardware | documents processed |
| Version control | provider | operator | operator |
| Confidence per field | usually not exposed | usually available | evaluated against thresholds |
| Data protection | processing at the provider, third country transfer depending on location | no transfer to third parties | accountability evidence through audit trail |
| Evidence under the EU AI Act | not included | not included | logging, human oversight, documentation |
Measured values for a specific specialised model and a specific validation and automation layer are given in the section on implementation at Parashift.
Limits of the models
The large language model and the specialised model return a value for every field. Neither model checks whether that value is consistent within the business case. A correctly read amount can deviate from the sum of line items, a correctly read invoice can already have been booked, and a high confidence score describes the model’s certainty in reading, not correctness in context.
These checks, the routing of uncertain cases and the logging of processing are not part of the models. They are required in every production pipeline, regardless of which model is used.
Data protection and the EU AI Act
Business documents usually contain personal data. Under the GDPR and the Swiss FADP, purpose limitation, data processing agreements and the accountability principle apply, which requires processing to be demonstrable. With a large language model accessed through the interface of a provider outside the EU or Switzerland, third country transfer is added, which has to be secured contractually and technically. A specialised model in the own data centre processes documents without transfer to third parties.
The EU AI Act classifies AI systems as high risk under Annex III when they feed into decisions on creditworthiness, life and health insurance or access to public services, among others. Document pipelines in banking, insurance and public administration can fall under this. For high risk systems the regulation requires technical documentation (Art. 11), automatic logging (Art. 12), human oversight (Art. 14) and accuracy and robustness (Art. 15).
Logging, human oversight and documentation are not properties of a model. They are implemented in the validation and automation layer: audit trail per field, thresholds with routing into manual review, versioned specifications. The accuracy requirement concerns the model and is evidenced through measurements on defined datasets.
Functions of the validation and automation layer
The layer performs the following functions:
- Management of the specification per document type, with versioning
- Evaluation of confidence per field against thresholds set per field and per tenant
- Validation rules for totals, mandatory fields, duplicates, vendors and matching against purchase order, goods receipt and master data
- Routing into automatic further processing or into a validation interface in which only the flagged fields are handled
- Approval workflows with agent in the loop or human in the loop
- Feedback of corrections into the model
- Audit trail with value, confidence, location, specification version, rule applied, editor and time of release
Because the layer is model independent, the model can be replaced without changing rules, thresholds or audit trail. Two models are then compared on the same metrics.
Use of large language models
Large language models are suited to tasks outside document processing, to document types without a specification and to proofs of concept. They can be used without infrastructure, and their performance improves with each model generation without any action by the user.
In production three properties take effect: costs grow with page count and context length, the model version changes at the provider’s discretion, and documents are processed on the provider’s infrastructure. Data residency and evidence obligations have to be clarified separately in regulated environments.
Within a validation and automation layer, large language models can be integrated as a second opinion when a threshold is breached or as a fallback for exceptions. They then work against the same thresholds and write to the same audit trail as a specialised model.
Implementation at Parashift
Dokumentli is Parashift’s specialised model. The Parashift platform is Parashift’s validation and automation layer. Large language models are integrated into the platform through «bring your own model», provided they conform to the OpenAI API specification.
Dokumentli is a vision language model for business documents with a context budget of 8,192 tokens. It runs in the own data centre on a GPU with 24 GB VRAM or in a private cloud. In the Parashift measurement of September 2026 across 3,365 documents in five benchmarks, metric cANLS at threshold 0.8, Dokumentli 5.3 reaches a mean of 89.1 percent against 85.3 percent for the best reference model and leads in each of the five datasets. Throughput is 1.14 documents per second against 0.60 for the most accurate reference model. Domain fine tuning is available as self service.
| Dokumentli 5.3 | best reference model | |
|---|---|---|
| Accuracy cANLS@0.8, mean over five benchmarks | 89.1 % | 85.3 % |
| Throughput | 1.14 doc/s | 0.60 doc/s |
| Context budget | 8,192 tokens | up to 32,768 tokens |
| Place of operation | own data centre or private cloud | provider infrastructure |
Source: Document AI Benchmark, Parashift, September 2026.
The Parashift platform implements the functions of the validation and automation layer described above. Confidence scores are evaluated per field, thresholds are set per field and per tenant, validation rules cover format, mandatory fields, totals, duplicates, vendors and matching against connected systems. Flagged cases appear in an embedded validation interface, operated by the customer or by Parashift under an agreed turnaround time. OneTouchLearning® feeds corrections back into the model, and the audit trail logs every processing step. The platform is API first, multi tenant and offers more than 30 connectors. It runs as a cloud service, while the Dokumentli node is operated in the own data centre.
Summary
The validation and automation layer is required in every production document pipeline. Consistency checks, routing of uncertain cases and evidence are not part of the models, and the GDPR, the FADP and the EU AI Act require exactly this evidence. The layer remains in place for years, while the model underneath is replaced several times.
For document tasks a highly specialised model exceeds large language models in accuracy and throughput, with operation in the own data centre and costs independent of volume. Large language models remain usable for other tasks and as a complementary instance within the same layer. At Parashift these roles are filled by Dokumentli and the Parashift platform.
