AGENTIC BRAIN PLATFORM — COMPLIANCE DOCUMENTATION
EU AI Act — Annex III High-Risk
Prepared for the European Parliament ITEC Directorate — April 2026

An Agentic Brain Platform
Built for European Compliance

This document presents a European-built agentic AI platform — designed from the ground up with human oversight, data sovereignty, and EU AI Act compliance as architectural properties, not afterthoughts.

Brussels
Company HQ & server location
Annex III
High-risk classification (employment AI)
100%
Human-in-the-loop on all external actions
Aug 2026
Compliance deadline — ready now
9
Active brain deployments across 3 orgs

01 What Is the Brain?

An agentic AI platform for workforce empowerment

Not a Chatbot. A Structured Knowledge Workspace.

The brain is a persistent, agentic AI platform where human operators command specialized AI agents through natural language. Each brain is a structured workspace — with its own identity, knowledge base, tools, and memory — that empowers a professional to operate at a level that would otherwise require a team of specialists.

The AI model (currently Claude API by Anthropic) is the engine. The brain — the accumulated knowledge, the operating rules, the audit trail — belongs to the user. The model is replaceable. The brain is not.

02 Architecture

Five layers that make a brain

Identity boot/
Operating rules and ethical boundaries. Defines the agent's role, expertise, behavioral constraints, and what it explicitly cannot do. Every agent has a strict mandate.
Knowledge wiki/
Structured domain knowledge. Companies, candidates (pseudonymized), projects, legal frameworks. YAML frontmatter, searchable, linkable, versioned via Git.
Tools tools/
Actionable capabilities. Email (O365), calendar, document generation, web publishing. Every external action requires explicit human approval.
Agents agents/
Specialized modes. Recruiter, Finance, Legal, Social, Background Check — each with defined scope and boundaries. Activated by the human, not autonomously.
Memory diary/
Full audit trail. Every decision, action, and interaction timestamped. Institutional memory that compounds. Complete reconstructibility for regulators.

03 Data Flow

How information moves through the system

Human OperatorNatural language
Brain WorkspaceOn-premise, Belgium
AI Model APIClaude (Anthropic)
External ActionsEmail, calendar, publish

CONSTRAINT — every external action requires explicit human approval

04 Why High-Risk Under the EU AI Act

Annex III, Point 4 — Employment & Recruitment AI

When used for candidate screening, scoring, or assessment support, this platform falls under Annex III, point 4(a) of Regulation (EU) 2024/1689: AI systems intended for recruitment or selection of natural persons. Full compliance with Articles 6-49 required by 2 August 2026.

Our position: compliance by architecture, not by retrofit. Human oversight, transparency, and bias mitigation are structural properties — not features that can be toggled off.

05 Live Deployments

This is not a concept — it runs daily

Primary Brain

David — CEO Brain

8+ specialized agents. Manages recruitment operations, business development, finance, legal, social media, client communications. Non-coder operating at specialist level across domains.

Second Brain

Yoni — Director Brain

Same architecture, different content. Client delivery, pipeline management, operational workflows. Proves the multi-user model works independently.

Third-Party Validation

Emisfera — 6 Agency Brains

Six brains deployed at an Italian web agency. Independent organization, independent users. Third-party proof that the model works beyond the founders.

"Europe doesn't need to create unicorns. Europe's strength is specialized SMEs — PME sartoriali — that solve vertical problems with governance, compliance, and quality."
— ITEC Directorate, European Parliament

01 Infrastructure & Data Sovereignty

Belgian-hosted, EU-sovereign, no cloud dependency

Data Residency

Belgian On-Premise Server

Dedicated server physically located in Belgium. No cloud database, no US provider dependency for data storage. Full EU data sovereignty. Data at rest never leaves Belgian jurisdiction.

Architecture

File-System Based

The brain uses structured files (markdown + YAML), not a cloud database. Full portability, no vendor lock-in, version control via Git, standard formats. The user can export their entire brain at any time.

02 Workspace Isolation

Each brain is sandboxed

Multi-Tenant Security Model

  • Every brain workspace has its own isolated credentials (.env with 600 permissions)
  • Workspaces cannot see, access, or interact with other workspaces' data
  • File-system permissions enforce isolation at the OS level
  • Shared tools verify credentials before executing — no credentials = no action, clean error
  • Each brain has separate API keys, email accounts, and calendar access

03 GDPR — Native by Design

Candidate Data Protection
  • Pseudonymization is mandatory. Candidates always referenced by initials (M.R., A.B.) — enforced at the system level
  • No full names in logs or diary entries
  • Salary data protected in context
  • Client data separation enforced (no cross-client leakage)
  • Right to deletion fully supported (file-based = easy to locate and delete)
Data Retention Policies
  • Active recruitment: Duration + 6 months
  • Talent pool (with consent): Maximum 2 years
  • Unsuccessful candidates: 3-6 months post-process
  • Background checks: 6 months post-completion
  • Deletion certificates generated for audit trail

04 Security Measures

MeasureImplementationStatus
Encryption at restAES-256 for stored dataActive
Encryption in transitTLS 1.3 for all API communicationsActive
Access controlsRole-based access, SSH key authenticationActive
Credential isolationSeparate .env per workspace, 600 permissionsActive
Audit loggingGit version control — full change historyActive
Breach response72-hour notification protocol per GDPR Art. 33Documented
Data Processing AgreementsDPA with clients (controller-processor)Active
Incident response playbookSeverity classification, response proceduresPublished

05 AI Model Communication

What goes to the API — and what doesn't

Anthropic API — Data Safeguards

When the brain processes a request, relevant context is sent to Claude's API for inference. Safeguards:

  • Anthropic does NOT train on API customer data (contractual commitment)
  • API inputs are not stored beyond the request lifecycle
  • The brain minimizes data sent — only relevant context, never full database exports
  • No biometric data, no photos, no social media scraping is ever performed
  • Candidate data is pseudonymized before AI processing where possible

01 Human Oversight

Architectural enforcement, not an optional feature

Email Protocol (Mandatory)
  • Step 1: AI generates a draft
  • Step 2: Draft shown to human in chat
  • Step 3: Human reviews, modifies if needed
  • Step 4: Explicit confirmation required
  • No email is ever sent without human approval
All External Action Gates
  • Publishing content → human approval required
  • Calendar events → human approval required
  • Client communications → human approval required
  • Data deletion → human approval required
  • The AI proposes; the human disposes. Always.

02 Audit Trail

Every significant action logged and reconstructible

Complete Decision Traceability

  • diary/ captures every decision, action, and interaction with timestamps
  • Git version control: complete change history for every file modification
  • Agent activations, mode switches, and command executions are traceable
  • Candidate screening decisions include reviewer attribution and timestamp
  • Logs retained indefinitely via Git; minimum 6 months guaranteed per Art. 26(5)

03 Agent Boundaries

Each mode has defined scope — agents cannot exceed it

AgentScopeCan DoCannot Do
RecruiterRecruitment opsDraft outreach, screen CVs, prepare presentationsMake hiring decisions, reject candidates autonomously
FinanceFinancial analysisP&L modeling, projections, invoice trackingExecute payments, sign contracts, commit obligations
LegalLegal advisoryDraft contracts, review compliance, researchProvide binding advice, sign agreements
SocialLinkedIn contentDraft posts, editorial calendar, translationPublish without approval, comment as the user
Background CheckVerificationStructure process, generate templatesMake pass/fail decisions, contact refs without approval
Client OpsProject managementScore applications, generate notices, prep assessmentsReject candidates without human confirmation

04 Bias Monitoring

Current Measures

Built-In Bias Mitigation

Blind screening (initials only, no photos/age/gender). Nationality excluded from scoring. Culturally-adapted psychometrics. Adverse impact analysis after each scoring run. Interviews 60% / tests 40% — human judgment dominates.

Academic Partnership

ULB-KTO Validation

Partnership with Université Libre de Bruxelles for algorithmic auditing. Focus: bias detection in screening, fairness constraints. Third-party, academically rigorous validation of AI-assisted decisions.

05 Fundamental Rights Impact Assessment

FRIA completed March 2026 per Article 27

Five Fundamental Rights Analyzed

A comprehensive FRIA has been completed for the CEFTA Secretariat recruitment project (6 positions, Western Balkans candidate pool). It covers:

  • Right to non-discrimination (EU Charter, Art. 21; ECHR, Art. 14)
  • Right to privacy and data protection (EU Charter, Art. 7-8; GDPR)
  • Right to human dignity (EU Charter, Art. 1)
  • Right to effective remedy and fair trial (EU Charter, Art. 47)
  • Right to good administration (EU Charter, Art. 41)

This FRIA methodology is reusable across all client engagements and serves as a compliance template for AI deployment in recruitment.

06 Self-Assessment

Status against key EU AI Act articles

ArticleRequirementStatusEvidence
Art. 9Risk Management SystemGreenFRIA completed, risk register maintained
Art. 10Data GovernanceGreenGDPR framework, pseudonymization, retention schedules
Art. 12Record-KeepingGreendiary/ logs, Git version control, audit trails
Art. 13TransparencyGreenFull documentation of capabilities and limitations
Art. 14Human OversightGreenArchitecturally enforced — all external actions gated
Art. 15Accuracy & RobustnessAmberOperational validation; formal metrics in progress
Art. 26(2)Human Oversight (Deployer)GreenDesignated persons with competence and authority
Art. 26(5)Log RetentionGreenRetained indefinitely via Git; 6+ months guaranteed
Art. 26(11)Inform IndividualsAmberAI disclosure notice on forms; update in progress
Art. 27FRIAGreenCompleted March 2026
Art. 4AI LiteracyGreenOperator deeply trained in AI capabilities/limitations
Art. 5Prohibited PracticesGreenNo emotion recognition, biometric categorization, social scoring

01 Empowerment, Not Replacement

The brain amplifies human capability — it doesn't automate away roles

What the Brain Does
  • Gives professionals access to domain expertise, playbooks, and tools that would require years of experience or a specialist team
  • A non-technical CEO manages finance, legal, and client comms through specialized agents
  • A junior recruiter operates at senior level from day one
  • Knowledge stays in the organization when people leave
  • Capability compounds over time
What the Brain Does NOT Do
  • Does NOT make autonomous hiring or firing decisions
  • Does NOT replace human judgment in critical matters
  • Does NOT operate without human supervision
  • Does NOT create dependency — deliberate "struggle zones" where humans must apply judgment
  • Does NOT eliminate jobs — elevates the scope of what each person can accomplish

02 The "Broken Rung" Problem

Why this matters right now

-46%
UK tech grad roles (2024-26)
-67%
US junior postings (2024-26)
37%
Orgs replacing juniors with AI
Day 1
Competency via brain, not months of training

The Paradox

Companies are eliminating junior roles (AI handles grunt work that was also the learning path) while struggling to find capable people (nobody's getting trained anymore). The brain bridges this gap: juniors arrive productive immediately while still developing real skills — because the brain coaches rather than dictates.

03 Living Proof

Non-coders empowered, not replaced

CEO Brain

David — IT Proficiency: 2/10

Manages complex RPO operations, builds client portals, handles P&L modeling, drafts legal documents, creates compliance frameworks, runs trilingual content strategy. The brain gave him capabilities that would otherwise require 5-6 specialists.

Director Brain

Yoni — Not a Coder Either

Independent client delivery, pipeline management, operational workflows, personal style adaptation. Same architecture, different person, different content — proves the model works universally.

Third-Party

Emisfera — 6 Independent Users

Web agency in Italy. Six brains deployed for their own operations. Not our company, not our employees. Strongest proof the architecture works beyond the founders.

04 Anti-Shallow-Competence Design

The Brain Coaches — It Doesn't Dictate

Stanford research shows 28% skill degradation when AI is removed after 6 months. The brain addresses this deliberately:

  • Struggle zones: Areas where the user must apply their own judgment — the brain provides framework, not answers
  • Escalation, not automation: Complex decisions go to the human, not resolved by AI
  • Transparent reasoning: The brain explains its logic so the user learns, not just follows
  • Progressive autonomy: As the user grows, they handle more complexity with less scaffolding

The goal: capability that compounds WITH the person, not instead of them.

05 Impact on Employment

Net positive across every dimension

DimensionTraditional AI RiskBrain-Powered Approach
Job displacementAutomates tasks, eliminates rolesAmplifies roles, expands scope per person
Skill developmentCreates dependency, atrophies skillsCoaches through struggle zones, builds competence
Knowledge retentionTrapped in models, not portableOwned by user, stays in org, fully portable
Access to expertiseGatekept by expensive tools/trainingDemocratized through brain templates
Junior workforce"Broken rung" — juniors disappearBridge: productive from day 1, still learning
Human oversightOptional, often bypassedArchitectural requirement, cannot be bypassed

Kill Switch

Any AI agent can be disabled at any time. Every process can complete entirely without AI — slower but fully functional. No AI system is a single point of failure. The human can always revert to manual operation.