Built for the EU AI Act compliance hackathon

Preliminary EU AI Act assessments,
composed by six specialised agents.

Upload a technical spec, vendor whitepaper, or process note. Lexicon's agent pipeline extracts facts, retrieves the right regulatory chunks, drafts the assessment, critiques its own evidence, and composes a six-section report — every claim cited.

6
Specialised agents
1
Built-in regulatory corpus
<2 min
Per assessment
The pipeline

Six subagents, each with one job. One boss deciding.

Explicit handoffs make every step inspectable. No black-box single prompt.

01

Boss Orchestrator

Triages the case, asks for missing info early

Runs first. Reads the uploaded material and decides whether the pipeline has enough to produce a reliable assessment. If critical facts are missing (intended users, affected persons, deployment context, automation level, GPAI usage, etc.), it pauses the pipeline and asks the user a short list of clarifying questions before any data analysis begins. Once answers are provided, it routes the case through the six specialist subagents.

02

Document Intake

Classifies + chunks uploaded material

Decides whether the upload is a technical spec, vendor whitepaper, policy doc, or process note. Tags every chunk with the source type and stores it in a session-scoped namespace, separate from the regulatory corpus.

03

Fact Extraction

Structured extraction with explicit uncertainty

Pulls purpose, users, affected persons, sector, input data, outputs, automation level, human oversight, GPAI usage, and AI-generated content. Flags ambiguities as uncertainties instead of silently resolving them.

04

Regulatory Retrieval

Selects relevant corpus chunks

Given the extracted facts, decides which parts of the AI Act, Commission guidelines, GPAI Code of Practice, and Finnish implementation material are relevant — labelling every chunk with its source type.

05

Assessment

Definition · risk · transparency · GPAI · role

Produces the preliminary AI Act assessment: AI system definition check, risk classification (Annex III), prohibited practice screen, transparency obligations, GPAI obligations, and provider vs deployer role. Every claim is grounded in a cited article.

06

Critique

Catches weak evidence, triggers revisions

Asks whether each conclusion is supported by the uploaded docs or merely assumed. Sends the assessment back with specific revision instructions when evidence is weak, and writes the missing-info list and follow-up questions.

07

Report Composer

Assembles facts and gap.

Assembles output sections, enforces fact/assumption/citation separation, adds the not-legal-advice disclaimer, and produces the export-ready report.

How it works

01

Upload use-case material

Paste your use-case description or upload a doc. Intake classifies and chunks it into a session-scoped namespace — strictly separate from the regulatory corpus.

02

Agents reason and cross-check

Facts → retrieval → assessment → critique. The critique agent can send the assessment back with specific revision instructions when evidence is thin.

03

Get a cited six-section report

Facts, assumptions and citations are kept separate. Every conclusion traces back to an Article or guideline. Export-ready.

Built-in corpus

Pre-indexed regulatory knowledge

The corpus is loaded at startup and lives in a separate namespace from your uploads, so every citation is traceable to its source type — legislation, official guidance, or national implementation.

legislation
Recitals 1–45 — Background and scope rationale
Regulation (EU) 2024/1689, Recitals 1–45
legislation
Recitals 46–100 — Risk framework rationale
Regulation (EU) 2024/1689, Recitals 46–100
legislation
Recitals 101–180 — Governance and enforcement rationale
Regulation (EU) 2024/1689, Recitals 101–180
legislation
Chapter I — General provisions and definitions
Regulation (EU) 2024/1689, Chapter I (Arts 1–4)
legislation
Chapter II — Prohibited AI practices
Regulation (EU) 2024/1689, Chapter II (Art. 5)
legislation
Chapter III §1 — High-risk AI classification
Regulation (EU) 2024/1689, Chapter III §1 (Arts 6–7) + Annex III

Ready to assess your AI system?

Paste a use-case description and watch six agents do their work in under two minutes.

Start an assessment