When AI comes up in Controlling, the conversation eventually reaches accountability. Who is liable when the number comes out of an agent?
The objection is fair. A language model produces output that looks plausible, and plausible is not enough when the number steers a business.
I spoke about this at the 3rd Conference on AI in Finance at TU Dresden, under the title "Harness Design for Agentic Decision Support on Driver-Based Models". What the harness does:
- The agent operates defined tools. The calculation logic is fixed.
- The meaning of the drivers is given. The agent does not have to infer it.
- Every step stays inspectable on its own, weeks later too.
Abstract
In Controlling, analytical questions are often dropped because building a model for a one-off question does not pay off: the binding constraint is the cost of construction (acquiring and preparing the data, building the model, and getting the organization aligned on it), not the value of the answer. Agentic AI lowers that cost, so its primary effect is a shift in which questions become worth asking. We argue the resulting engineering problem is harness design: the scaffolding that lets an agent operate safely on driver-based decision models. We present a working harness in which an LLM agent drives Valsight, a production planning and simulation platform, through the Model Context Protocol (MCP), distil four design principles, and close with what already works and what stays human.
Across the other contributions, from credit scoring to liability design, one constant held. People decide one way or the other. Either they accept the decision, or they put in place the algorithm that makes it.
Many thanks to Lars Hornuf and the Faculty of Business and Economics at TU Dresden for organizing the conference - and the funny hats ;)