AI-Powered Decision-Making: Transparent and Traceable

Companies make decisions with far-reaching implications every day – from investments and product roadmaps to compliance issues. AI tools are already being used in decision-making processes but often remain opaque and difficult to substantiate. This is precisely where the technology project of the Steinbeis Research Center Management Automation comes in: its goal is to develop a transparent AI agent system for internal organizational decision support. The system integrates LLMs, a decision aggregator for weighting expert opinions, and structured recommendations outputs. It is designed to be traceable and extensible.

The methodological approach covers system development – including agent roles, aggregation logic, and interfaces – as well as use-case simulation through business scenario testing and evaluation through user studies and comparisons with existing tools.

The Idea: Roles Instead of a Black Box

At the heart of the system is a multi-agent approach: several AI agents – such as CEO, Legal, Controlling, Marketing and Strategy – assess a decision-making issue from their respective perspectives. Each response follows a standardized JSON schema (recommendation, rationale, risks, opportunities, confidence). A decision aggregator consolidates the individual assessments using weighted criteria and generates an overall recommendation – including conditions (‘approve subject to conditions’) and a concise management summary. The aggregator translates the specialist logic of real-world roles into specialised AI agents, generates structured and traceable outputs, and creates transparency.

The approach is particularly compelling due to three key advantages:

• Explainable: Justifications, risks and sources are visible.
• Comparable: Each role follows a uniform format..
• Auditable: Results can be versioned, exported and audited.

Research, SMEs, and Consulting in Collaboration

The project is designed as a collaborative technology initiative. Academic partners investigate acceptance, explainability and process impact (decision support systems, explainable AI, organizational psychology). Companies contribute real-world use cases as pilot partners (investments, supplier selection, location decisions, budget issues). Consulting partners test the system in workshops and develop templates and transfer formats.

The open nature of the core enables participation and rapid iteration; proprietary extensions such as industry-specific roles or CRM connectors, remain optional.

Results, Timeline and Participation

“We are looking for pilot companies from industry, services, and administration, as well as research partners from business informatics, psychology, and law, and transfer partners for workshops and training,” explains Steinbeis entrepreneur Dr. Helmut Döring. The project timeline spans two phases:

• Spring 2026 (project objectives): MVP with at least five agent roles, JSON schema, aggregator, export; piloting with one to two partners; evaluation studies on acceptance, time-to-decision and quality of rationale.
• Summer 2026 (further development): API integration (ERP/CRM), RAG enhancement, multi-user operation; governance functions (audit trail, roles/permissions concept); optional Docker/Kubernetes deployment.

Outlook: From Decision Log to Learning Organization
In the long term, the agent system can evolve into a learning decision archive: Which justifications led to which outcomes in the past? Which conditions proved effective? This creates an organization-wide memory – data-driven, explainable and scalable.