How Do You Govern an Autonomous AI Agent in a Bank?

7 min readAbleneo AI transformation team

Short answer. A bank governs an autonomous AI agent with four controls: a written charter that fixes the agent’s objective and its hard limits, delegated authority capped below any action that needs a human sign-off, a named human owner who can stop or reverse it, and a complete audit trail of every action it takes. Under DORA, in force since January 2025, the bank stays accountable for the agent even when a vendor operates it. An agent that scores credit or prices a product is high-risk under the EU AI Act, with obligations for standalone systems starting 2 December 2027.

1What This Means in Practice

An autonomous AI agent is not a chatbot. A chatbot answers. An agent acts: it opens cases, moves money between ledgers, updates customer records, and calls other systems to finish a task without a human pressing a button at each step. That shift from generating text to taking actions is what changes the governance problem. A wrong sentence is embarrassing. A wrong action overwrites a record, approves a change, or sends a payment.

The workable model is to treat the agent like a junior employee who holds a mandate. A new hire gets a job description, a spending limit, a manager who signs off above that limit, and a record of what they did. An autonomous agent needs the same four things, written down before it goes live, not after an incident. Banks that deploy agents well set them on a ladder of autonomy and move up one rung at a time:

Production discipline is the deciding factor, not model choice. Ableneo ships roughly 4 of 5 of its AI projects into production, and the ones that stick are the ones with clear owners and limits from day one. In one live deployment at T-Mobile Czech Republic, an agent cut document validation from 7 hours to 30 minutes, 14 times faster, at over 90% accuracy, because the task, the authority, and the fallback were defined before launch.

2Why This Matters for Regulated Industries

For a bank in Central Europe, two regulations decide what “governed” means. DORA, in force since January 2025, treats an agent that reads data, takes actions, and talks to other systems as an ICT system, so the operational-resilience obligations that cover a core banking platform cover the agent too. DORA also keeps the financial entity accountable even when the ICT runs on a third party, so accountability cannot sit only in a vendor’s console.

The EU AI Act adds the decision layer. An agent used for credit scoring or insurance risk assessment and pricing is high-risk under Annex III, with obligations for standalone systems applying from 2 December 2027 and for systems embedded in regulated products from 2 August 2028. Article 14 requires that a person can interpret the agent’s output, override or reverse it, and stop it. Before those dates, the ECB has already told banks to submit AI-cyber action plans by 31 October 2026. The trigger is live now, and the agents being built today are the ones that will be audited.

Govern an autonomous agent with four controls: a written charter, capped authority, a named owner, and a full audit trail.

3Who Is Accountable When an Autonomous Agent Acts?

The bank is, and inside the bank a named person is. DORA removes the option of pointing at a vendor: the financial entity owns the outcome of any ICT system it runs, agent included. In practice this means every material agent has a single accountable owner in the first line of business, risk and compliance in the second line reviewing its charter and limits, and internal audit in the third line testing whether the controls held. An agent without a named owner is an unowned risk, and in an audit that gap is the first thing a supervisor finds.

4What Does the EU AI Act Require for an Agent That Makes Decisions?

Article 14 of the EU AI Act requires that high-risk systems are built so a person can oversee them while they run. For an agent, that means three concrete capabilities: a person must be able to correctly interpret what the agent decided, disregard or reverse that decision, and interrupt the agent through a stop control. If the agent scores credit, prices insurance, or decides access to a financial service, it falls under Annex III and these obligations bind from 2 December 2027. The design work, logging, human-oversight interfaces, and documentation, takes 9 to 14 months, so the window to start is now, not in 2027.

5How Much Autonomy Should You Give an Agent?

Start at the lowest rung that still delivers value and earn each step up with evidence. The OECD, in its 2026 study of agentic AI in organisations, found that organisations prefer rule-based control for low-risk routine tasks and reserve full autonomy for narrow cases, because the named risk is an unintended action: an agent deleting data, overwriting a record with a hallucinated value, or approving a change without consent. A good rule is to cap delegated authority below the first threshold that would need a human sign-off if a person did the same job. A payment agent with a per-transaction limit and a daily cap is governable. One with no ceiling is not.

6What Goes in an Agent’s Audit Trail?

Every action, with enough context to reconstruct why. The OECD study is blunt that explainability is a prerequisite for accountability, not an optional feature, and it separates two needs: explainability at the moment a decision is made, and traceability to reconstruct decisions afterward. For a bank that means logging the agent’s identity, the authority it acted under, the inputs it saw, the action it took, and the reasoning behind it, not just a score. DORA then requires that a major ICT-related incident, including an agent acting outside its mandate, is detected, classified, and reported. An audit trail that records only outcomes cannot answer the supervisor’s question, which is always why.

7What Should a Bank Do First?

Inventory the agents you already have, including the ones a team spun up without telling risk. Then take the single highest-value agent and give it the four controls: a one-page charter, a capped authority, a named owner, and a full audit trail. Add a kill switch that a human can reach without the vendor. That one governed agent becomes the template for the next ten, which is cheaper than retrofitting governance across a fleet after the first incident.

8The Ableneo Perspective

Ableneo builds the governance into the agent, not around it. Across 34 production AI projects in 2025 for 20 clients in regulated industries, including banks and insurers in Central Europe, the pattern that separates the agents that reach production from the ones that stall is the one described above: a mandate, a limit, an owner, and a trail, defined before launch. That is why roughly 4 of 5 of those projects run in production rather than sitting as pilots. For the fuller set of decisions a bank faces when it puts AI into production, see Ableneo’s AI transformation FAQ.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo