AI in supervisory and inspection institutions

Automatic standardisation of reports

Process automation

Problem: Reports from supervised entities arrive in inconsistent formats

How it works: OCR and structured answers return data in a standardised structure ready for analysis

What you gain: The analyst gets data in one format regardless of how it was sent — time for analysis instead of tidying up

Assistant for interpreting regulations and guidelines

Knowledge and documents

Problem: Regulations and guidelines are scattered, and a wrong interpretation is costly

How it works: A central RAG knowledge base that answers only on the basis of the specified sources; the Knowledge Gap Register instead of answers without a basis

What you gain: An interpretation always points to the provision it relies on; when there is no basis, the system says so instead of inventing one

Deterministic AI-supported calculations

Analytics and reporting

Problem: Analyses require calculations that must not be entrusted to a language model

How it works: SQL/Python code runs directly in the chat — the calculation is deterministic, and the model only interprets and explains

What you gain: The result is repeatable and can be checked by a second person — AI makes the numbers easier to reach but does not “invent” them

Inspection notes by voice

Process automation

Problem: Inspectors in the field take notes by hand and tidy them up after they return

How it works: Voice chat (STT/TTS) and the AI Powered Notes editor with versioning

What you gain: The inspection note is structured while still in the field, with a version history — no evening at the keyboard

Reproducibility of analyses and decisions

Compliance and oversight

Problem: An analyst must reconstruct how a conclusion from six months ago was reached

How it works: Unified telemetry with Replay of historical requests; an immutable log with the configuration in force at the time of the query

What you gain: Every analysis can be reproduced in exactly the configuration in which it was created — with the model, prompt and sources

Central oversight of compliant AI use

Compliance and oversight

Problem: An institution that supervises others must itself be a model of compliance in its use of AI

How it works: ASIA report (CDF-F1.5) and agent register (F2); AI Disclosure Injection that marks AI content; a policy and risk status dashboard for compliance

What you gain: The institution can show its own AI governance at the same level it requires from supervised entities

Is our tool an “AI system” under the AI Act?

This is decided by the definition in Article 3(1) of the AI Act: a system that infers from the input it receives how to generate outputs. The Initial Assessment in the CDF methodology classifies each use case and documents the rationale

Where is data processed, and can it be audited?

In the institution’s infrastructure or in a chosen cloud in Poland or the EU. The WORM log in PROXY:AI shows who processed data, when and with which model

How do you preserve the evidential value of findings when using AI?

By separating roles: the language model searches and summarizes with citations, and calculations are performed by deterministic tools. Every result can be reproduced from the log