Experts free to do expert work.

Context

A grant-awarding institution faced large intake volumes against tight deadlines. Each application meant dozens of pages of documentation read by a small team of experts.

Problem

Manual verification was slow and error-prone. Experts spent most of their time on formal checks: does the budget add up, is the application signed by an authorised person, do the cost categories stay inside the regulatory limits. The substantive review, which is what the institution actually employed them for, kept getting squeezed. GDPR and the data processing agreement made standard cloud-based automation a non-starter.

Solution

We designed and deployed an on-premise hybrid system. The system handles the rules that are unambiguous and quantitative. An AI expert layer handles the cases that need contextual interpretation. Application batches are processed offline against anonymised personal data. The output is a structured report that flags formal status, surfaces inconsistencies, and routes only the cases that need human judgement to the expert team.

Implementation

The pipeline runs in four stages:

• Read and check. Extract project data and budget tables. Verify the basic rules: implementation window, cost-category caps, minimum own contribution, page numbering.
• Eligibility. An AI expert layer assesses applicant entitlement, signature validity and category exclusions.
• Completeness. Budget detail, attachments and cost-category eligibility checked against regulations.
• Substantive scoring. Model-assisted analysis of alignment with statutory goals. Final decision stays with the human expert.

An optional module verifies entities against the National Court Register (KRS).

Outcome

Experts got a tool that handles the formal grunt work. Verification time compressed dramatically. The risk of missing a formal error dropped close to zero. The process became transparent and cost-effective. The architecture stays aligned with GDPR and the other regulations the institution operates under.