Artificial intelligence is increasingly becoming part of operational finance processes. AI agents analyse data, create cash forecasts and support decisions faster and at greater scale than ever before.

But in finance that is not enough. Here, it is not about whether an answer sounds plausible - but whether every figure is reliable, precisely justified and auditable. A cash forecast is not a recommendation, but the basis for operational decisions with direct financial impact. Plausibility is not a quality criterion here, but a risk.

This fundamentally shifts the focus: not how eloquent a system is, but how controllable. Engineering in this context means:

As agentic as needed, as deterministic as possible.

The problem: AI without accountability

Many AI systems today are optimised for speed and versatility. At first glance, they deliver convincing results – but they are not built to take responsibility for correctness at the individual data level.

Even when outputs are correct, the ability to trace or reproduce them in detail is often lacking. This is precisely what becomes a problem in finance.

Typical weaknesses of generic AI in finance are:

• Lack of traceability back to specific data sources

• No stable decision logic over time

• Insufficient auditability at process level

• Lack of embedding in governance structures

The consequence: AI does not become an enabler, but a potential risk.

What makes AI trustworthy in finance

Trustworthy AI is not distinguished by better wording, but by its structural properties. A system is trustworthy when it not only makes decisions, but can trace every figure back to specific data and produce exact repeatable results under the same conditions.

A key component of this architecture is the Agentic Ledger.

Every action – from data access to decision logic and automated processes - is documented comprehensively, durably and in an audit-proof manner. This ledger operates on an append-only principle: information is only added, while the ground truth data from the systems-of-record remain untouched.

At the same time, Trustworthiness requires consistent technical safeguards. These include dedicated, isolated instances in private cloud environments on AWS that communicate via encrypted connections. A semantic layer within the context architecture ensures that the agent operates only within defined boundaries. This is complemented by the principle of minimal privileges and fully European data residency.

Why a cognitive architecture is necessary

These requirements cannot be solved with better models, but only with a different system architecture. A cognitive architecture clearly separates core functions such as memory, reasoning, perception, execution and governance.

We use a two-pass architecture: in the first phase, the system learns rules that are applied deterministically in the second phase. The aim is to execute as much of the process as possible via a deterministic and thus highly effective path – away from statistical probability, towards exact logic. To keep context stable over time, we use a multi-layer memory architecture that combines different storage approaches to map both short-term tasks and long-term financial logic without logic drift.

Within our Multi-Agent System (MAS) we integrate LLM-as-judge steps. These validate the results of each individual agent node before they reach the user. Every recommendation is documented including the applied logic and possible alternatives.

Trust requires control over data

A frequently underestimated aspect of Trustworthiness is the role of humans. We consistently follow the human-on-the-loop approach. There is no fully autonomous execution of critical financial transactions without oversight. The system acts proactively, identifies opportunities, names root causes and shows the exact data points – but final approval always remains with the human.

Trust is only created when companies retain full control:

• No uncontrolled sharing of sensitive data through private LLM deployments

• Seamless integration into existing systems such as ERP, banking or TMS

• No complex IT transformation projects

• Compliance with the strictest security and compliance requirements

The future of AI in finance

As AI becomes more widespread, the performance of many systems will converge. Speed and model size will no longer be the decisive factor. The difference will lie in which systems companies can trust.

The decisive questions will be:

• Does the system remain stable over time?

• Are decisions auditable in the ledger down to the level of agent nodes?

• Can results be explained at any time through deterministic paths?

Organisations that can answer these questions build a structural advantage.

Conclusion

Trustworthiness is not a feature that can be added later. It is the result of a well-designed architecture.

Finance does not need generic AI that produces plausible answers. Finance needs systems that think in context, act within clear boundaries and make every decision traceable. The future does not belong to the systems that formulate best.

It belongs to those that decide reliably.

Experience in a short demo how a digital treasury employee works:

Foto von Bayu Fajariyanto auf Unsplash