You have an AI agent in your budgeting process. It analyzes plan variances, writes comments and initiates correcting entries.
As long as everything is right: perfect. But what happens when it makes a mistake? Who is liable? Who approved it?
This is exactly the problem that became a regulatory obligation for finance teams in 2026: AI governance. Without defined rules, control mechanisms and escalation paths, every AI agent in controlling is an uncontrolled risk.
What does AI governance mean in a controlling context?
AI governance refers to the structured framework of policies, processes and technical control mechanisms that companies use to ensure that AI systems act reliably, traceably and in compliance with the rules.
In a controlling context, this means: Every decision an AI agent makes, whether it flags a variance, adjusts a forecast or triggers a posting, must be traceable. It must be approved, or at least approvable. And if in doubt, it must be reversible.
AI governance is not an abstract IT concept. It is the answer to the question: Who is responsible when the AI agent gets it wrong?
Why are guardrails for finance AI agents becoming a regulatory requirement?
Two sets of rules created the decisive pressure in 2026: the EU AI Act and DORA.
The EU AI Act classifies AI systems that influence financial decisions as high-risk. This means: documentation requirements, a risk assessment before deployment and ongoing monitoring during operation. Anyone who uses an AI agent for forecasting or reporting without meeting these requirements risks substantial fines.
DORA, the Digital Operational Resilience Act, directly affects financial institutions. It requires all AI-supported processes to be reviewed for resilience and controllability. Companies outside the financial sector feel the pressure too, because banks and insurers increasingly pass these requirements on to their suppliers.
Beyond regulation, there is a practical reason: AI agents make mistakes. Not often, but when they do, it happens fast and at scale. A poorly calibrated forecast agent can deliver the wrong basis for decisions worth millions within minutes.
Which regulatory requirements do finance teams need to know in 2026?
Three sets of rules are decisive for finance teams in the DACH region:
EU AI Act
AI systems in finance typically fall into the “high-risk” category. This obliges you to:
- Risk analysis before going live
- Documentation of all training data and model decisions
- Setting up a human oversight mechanism
- Logging of all AI decisions for at least five years
DORA (Digital Operational Resilience Act)
Applies directly to financial institutions and indirectly to their service providers. Key points:
- ICT risk management must include AI systems
- Critical AI processes need defined backup strategies
- Mandatory incident reporting for AI-related outages or wrong decisions
Local requirements (HGB, IDW, BaFin)
The German Commercial Code (HGB) and IDW standards require traceable and auditable bookkeeping. AI-supported posting processes must meet the same audit-proof requirements as manual processes.
Which requirements apply to your use of AI?
We help you define the right framework. No blanket verdict, no off-the-shelf consulting solution.
How do automated guardrails work in practice?
A guardrail is a technical or process-based barrier that prevents an AI agent from acting outside defined limits.
For finance AI agents, this specifically means:
Rule-based barriers
The agent may only act within defined parameters. Example: forecast adjustments of more than 15 percent compared to the previous month require manual approval. Postings above a certain amount are blocked until a controller has confirmed the output.
Audit trail and logging
Every action of the AI agent is fully logged: What did it analyze? Which decision did it make? Based on which data? This audit trail is the basis for every later audit.
Escalation paths
In the event of anomalies or uncertainty, the agent escalates to a human controller. No AI agent should act autonomously in areas where the outcome is not clearly compliant.

Monitoring and drift detection
Models age. When market conditions or internal structures change, an AI model can systematically produce wrong predictions unless it is adjusted. Automated monitoring detects when a model deviates from its historical hit rates.
Platforms such as IBM watsonx.governance offer these functions as an integrated tool chain. But your own governance frameworks based on logging standards and process mining tools can also be sufficient. What matters is not the tool, but consistent implementation.
What does a finance AI need to meet in terms of governance?
Here is a field-tested checklist for the governance readiness of your AI agents:
| Governance checklist for finance AI agents | |
|---|---|
| ☐ | Is a risk classification under the EU AI Act documented for every AI agent? |
| ☐ | Is there a complete audit trail of all AI decisions and data sources? |
| ☐ | Are escalation paths defined for edge cases and wrong decisions? |
| ☐ | Are there technical guardrails for critical actions (postings, approvals, forecasts)? |
| ☐ | Is a human oversight mechanism in place for high-risk decisions? |
| ☐ | Are AI models regularly monitored for drift and loss of quality? |
| ☐ | Is the governance framework documented in an auditable way for the next audit? |
| ☐ | Are employees trained to critically question AI results? |
How can BI2run support you with AI governance?
AI governance is not a one-off project. It is an ongoing process that combines technical know-how, process understanding and regulatory knowledge.
Many finance teams we work with already use AI. Yet the governance documentation is missing. Not because they forgot about it, but because they didn’t know exactly what needs to be documented and in what form.
We help you assess the status of your AI usage, identify governance gaps and develop a pragmatic plan that holds up from a regulatory perspective. If the check shows that you don’t need to take action, we’ll tell you that too.
AI governance is often done faster than you think
Drop us a quick note on which AI systems you use in finance. We’ll give you an initial assessment.
Book a meeting →Facts & sources
| 1 | EU AI Act: AI systems for creditworthiness assessment, accounting and financial planning are considered high-risk (Annex III). Source: European Parliament, 2024. |
| 2 | DORA: Requires EU financial institutions to fully document all ICT risks, including AI, from January 2025. Source: EU Regulation 2022/2554. |
| 3 | 67% of finance leaders see a lack of governance as the biggest obstacle to using AI in controlling. Source: Gartner Finance Survey 2025. |
| 4 | IBM watsonx.governance enables automated monitoring, audit trails and bias detection for AI models in enterprise use. Source: IBM TechXchange 2026. |
| 5 | Fines under the EU AI Act for violations involving high-risk AI systems: up to EUR 30 million or 6% of global annual turnover. Source: EU AI Act, Art. 99. |
| 6 | Companies with structured AI governance reduce unplanned AI-related outages by an average of 43%. Source: McKinsey AI Risk Report 2025. |
Glossary
| Term | Definition |
|---|---|
| AI governance | Structured framework of policies, processes and technical controls that ensures AI systems act in a compliant, traceable and responsible way. |
| Guardrail | Technical or process-based barrier that prevents an AI agent from acting outside defined parameters. Example: automatic blocking of postings above a threshold. |
| EU AI Act | EU regulation governing AI systems, tiered by risk class. High-risk AI systems, including many finance applications, are subject to strict transparency and documentation requirements. |
| DORA | Digital Operational Resilience Act: EU regulation on the digital operational resilience of financial institutions. Among other things, it mandates ICT risk management and incident reporting. |
| Audit trail | Complete, tamper-proof logging of all actions and decisions of an AI system, including the underlying data. |
| Model drift | Gradual deterioration in the prediction quality of an AI model when conditions change and the model is not retrained. |
| Human oversight | Requirement of the EU AI Act: For high-risk AI decisions, a human must be able to override or prevent the decision. |
| watsonx.governance | IBM platform for monitoring, documenting and assessing the risk of AI models in enterprise use. Supports compliance with regulatory requirements. |
FAQ: AI governance in controlling
Which AI systems in controlling fall under the EU AI Act?
AI systems that control credit decisions, forecasts for financial reporting or automated posting processes typically fall into the high-risk category of the EU AI Act. Systems that only give recommendations and make no autonomous decisions may fall into a lower risk class. The classification depends on the specific use case.
How much does it cost to set up AI governance?
That depends heavily on your existing setup. If you already have documented processes and logging infrastructure, a few weeks of effort are often enough. If you run AI agents without any documentation, you need a structured inventory as a starting point. In practice, we see that the first prioritized steps can be implemented faster than expected in most cases.
How are AI governance and IT security related?
AI governance addresses content control and regulatory traceability. IT security addresses technical protection against attacks and data loss. The two complement each other and cannot be separated: a manipulated AI agent is both a security problem and a governance problem. Our article on IBM Vault 2.2 shows how agents’ credentials can be secured.
Do I need to create separate governance documentation for every AI agent?
For high-risk systems under the EU AI Act: yes, for every standalone agent. For low-risk systems, a shared framework that only adds agent-specific details can often be used. A central governance framework with modular extensions for each use case is more efficient.
What is the difference between AI governance and AI ethics?
AI ethics is the normative foundation: What should AI be allowed to do, and what not? AI governance is the operational implementation: How do we make sure the AI acts according to these rules? Finance teams need both, but in day-to-day controlling, AI governance is the immediately relevant topic because it directly supports regulatory obligations.
When does DORA apply to our company?
DORA applies directly to regulated financial institutions in the EU: banks, insurers, investment funds and payment service providers. Other companies are indirectly affected if they act as ICT service providers for financial institutions. Even if your company is not directly affected, banks and insurers increasingly pass these requirements on to their suppliers by contract.

























