Agent Types
Document Agent
A specialized AI agent that reads, understands, and processes documents. No template recognition, no rigid OCR - contextual language understanding. Document Agents process invoices, sick leave certificates, contracts, certificates, and receipts. Every decision is documented via the Decision Layer.
Workflow Agent
An AI agent that orchestrates business processes across systems. From incoming mail through approval to booking. Every step is logged. When queries arise or information is missing, the workflow pauses - it does not abort.
Knowledge Agent
An AI agent that provides contextual answers from enterprise knowledge. Company agreements, policies, collective bargaining agreements. Every answer includes source reference and rule version. Without a source, no answer is given.
Governance Terms
Decision Layer
The Decision Layer decomposes every business process into individual decision steps and defines upfront for each: Does a human decide, does a rule set apply, or does the AI decide autonomously?
Where discretion, discrimination risk, or employee representation requirements are involved, the architecture enforces human review. Where a decision is deterministic - collective agreement terms, deadline checks, booking logic - the agent applies the rule set consistently. And where the agent is confident enough and has permission: it decides independently. It interprets documents, classifies situations, evaluates context - demonstrably more consistent and legally sound than manual processing. Confidence Routing controls when the agent acts autonomously and when it escalates.
Every decision is documented - who decided what, when, on what basis, with what outcome. Auditable for external auditors, employee representation bodies, and internal compliance.
Technically, the Decision Layer sits between AI Agent and target system (e.g., SAP, DATEV, Sage, Workday). It evaluates every agent decision by confidence and risk category and routes automatically. Every decision generates a complete, immutable decision record: input, model, applied rule set with version, confidence score, decision path, outcome, timestamp.
Human-in-the-Loop
An architectural principle where final decision authority remains with humans. Not added retroactively, but integrated into the agent architecture from the start. At low confidence or exceptions, the agent automatically escalates to the responsible specialist - with complete context and recommendation.
Governance by Design
Governance requirements are not added after the fact but are an integral part of the agent architecture. Audit trail, role concept, Decision Layer, and Human-in-the-Loop are present from day 1. The opposite of Shadow IT.
Audit Trail
The complete documentation of every agent decision: input data, applied rule, rule version, confidence score, result, timestamp. Not just for internal traceability - but for auditors, works councils, and certification audits.
Cert-Ready by Design
An architectural approach where all controls and documentation mechanisms are built to meet the requirements of common audit standards (ISA, PS 951, IDW, GoB/GoBD) during normal operations. No retroactive audit preparation.
Co-determination (Mitbestimmung)
The German right to workplace co-determination through the works council (BetrVG). For AI agents that process employee data or prepare personnel decisions, involving the works council is legally required. Gosign agents are designed to be works-council-ready from the start.
→ Co-determination & AI Agents
Micro-Decision
A micro-decision is a single, defined decision step within a business process. The Decision Layer decomposes every process into micro-decisions and defines for each one: does a human decide, does a rule set apply, or does the AI decide autonomously? Where discretion, discrimination risk, or employee representation requirements are involved, the architecture enforces human review. Where the decision is deterministic, the agent applies the rule set. Where the agent is confident enough and has permission, it decides independently - interpreting documents, classifying situations, evaluating context. This is not if-then-else - this is judgment within defined guardrails. Each micro-decision generates its own audit trail entry.
Confidence Routing
Confidence Routing is the mechanism in the Decision Layer that automatically evaluates every agent decision by confidence and risk category and controls the decision path. High confidence with low risk leads to autonomous processing. Low confidence or high risk leads to escalation to a human. Thresholds are configurable and tenant-specific.
Works Council Agreement as System Constraint (Betriebsvereinbarung als System-Constraint)
In Gosign’s architecture, works council agreements (Betriebsvereinbarungen) are implemented as technical constraints in the Decision Layer. Each works council agreement is translated into a versioned rule set. If the works council agreement specifies, for example, that performance evaluations may not be fully automated, this is technically enforced as a Human-in-the-Loop rule - the agent cannot bypass it. The works council can verify in the Auditor Portal whether its requirements are technically implemented.
Shadow IT
IT systems operated without the knowledge or approval of the IT department. Particularly critical for AI agents, as uncontrolled LLM usage creates data privacy, compliance, and security risks. Gosign agents are integrated into existing IT governance - not as a parallel system.
Infrastructure Terms
Retrieval Augmented Generation (RAG)
An architecture where an LLM doesn’t just respond from its training corpus but additionally retrieves relevant documents from a knowledge base. In enterprise RAG: semantic chunking, metadata enrichment, hybrid search, source citation, and regular re-indexing.
LLM Hosting
Operating Large Language Models. Cloud (Azure OpenAI, Google Vertex AI - EU regions with DPAs), Self-Hosted (Llama, Mistral, DeepSeek - open source, own hardware), or Hybrid (sensitive data self-hosted, non-critical loads in the cloud).
Model-Agnostic
The property of an agent architecture not being bound to a specific LLM. Gosign agents work with Claude, ChatGPT, Gemini, Llama, Mistral, DeepSeek, and gpt-oss. Switching models does not change business logic, rule sets, or the audit trail.
Co-Build
Gosign’s model for client enablement. The agent is developed jointly, the client’s team is built up in parallel. After 12-18 months, the client operates their agents independently. No vendor lock-in, no permanent consulting model.
Regulatory
EU AI Act
The EU Artificial Intelligence Act (Regulation (EU) 2024/1689). Classifies AI systems by risk level and defines requirements for transparency, documentation, human oversight, and risk management. Particularly relevant for AI agents in HR and finance.
PII (Personally Identifiable Information)
Personal data that can directly or indirectly identify a natural person. In the context of Document Intelligence: name, address, date of birth, social security number, salary, bank details, biometric data. PII detection is the prerequisite for pseudonymization and redaction.
Roundtrip Pseudonymization
Three-stage process for GDPR-compliant document processing with language models. Step 1: Personal data is replaced with consistent pseudonyms. Step 2: The pseudonymized document is processed by the language model. Step 3: In the result, pseudonyms are replaced with real data. The model never sees personal data at any point. The mapping table never leaves the pre-processing layer.
Document Intelligence
The Document Agent’s capability to not only read and process documents, but also protect them: PII anonymization for LLM input, rule-based contract redaction for different recipients, and automated signature detection. All three capabilities are governed by the Decision Layer - for each decision step, it defines: human, ruleset, or AI.
Data Residency
Control over where data is physically stored and processed. In enterprise AI infrastructure: EU regions (Azure West Europe, GCP europe-west3), DPAs with all sub-processors, no data transfer outside the defined scope.