How Does a Bank Move From an AI Chatbot to AI Agents?

7 min readAbleneo AI transformation team

Short answer. A bank moves from a chatbot to an AI agent by expanding what the system is permitted to do, not by swapping the model underneath. A chatbot answers questions; an agent completes a task against live core systems. The real work is governance, data access, and process redesign, so the model is the small part. Start with one workflow where a human already signs off each step, wrap the agent in full audit logging and human approval, and prove it in production before adding a second. Ableneo has put a bank’s first production agent live in 100 days this way.

1What This Means in Practice

The difference is scope of action, not intelligence. A chatbot reads a question and returns text. An agent reads a goal, plans the steps, calls systems, and completes the task, then hands the exceptions to a person. Most banks in CEE already run the first kind. The move is to let the system act, inside controls, on work that a human used to do by hand.

A branch-assistant chatbot that cut internal support calls by 30% is a strong base for the first agent. It already holds the policy knowledge and the question patterns. The next step is narrow: let it open the case, pre-fill the form, and route it for approval, rather than only explaining how. The knowledge is reused, the action is new, and a person still signs.

Production proof shows what the end state looks like. Erste Group’s George AI has run for more than three years with a team of 21 people behind it, a sign that assistant layers mature into embedded capability over years, not weeks. In document-heavy workflows, Ableneo reached 99.3% touchless processing on 6,000 documents a month and 97% accuracy on a field computer-vision model. Those are agents, not chatbots: they finish the task and escalate only the exceptions.

2Why This Matters for Regulated Industries

A chatbot that only answers rarely triggers hard obligations. An agent that acts on customer money, a credit file, or a claim can cross into a different regime. Under the EU AI Act (Regulation (EU) 2024/1689, as amended in July 2026), an agent folded into credit scoring or insurance pricing sits in the Annex III high-risk category, with standalone high-risk obligations applying from 2 December 2027. Article 50 transparency duties have applied since 2 August 2026, so an agent that interacts with customers must disclose that it is a machine. The classification follows what the system does, not what the bank calls it.

DORA has been in force since January 2025, and the moment an agent touches a core banking system it becomes ICT in scope: logged, tested, and covered by an exit plan. The ECB reinforced this in 2026. Its 7 July 2026 letter to significant institutions asked boards to address AI-enabled cyber risk directly, and digital and AI governance sits inside the ECB supervisory priorities for 2026 to 2028. The practical reading is simple. Giving a chatbot the power to act raises its risk tier, so the governance has to be built before the agent goes live, not retrofitted after an audit finding.

Moving from a chatbot to an agent changes what the system may do, not the model. The real work is governance, data access, and process redesign.

3What Is the Difference Between a Banking Chatbot and an AI Agent?

A chatbot is a conversation layer. It matches a question to an answer, often from a fixed script or a retrieval index, and returns text. It does not change any system of record. An agent adds autonomy: it decides which steps to take, calls APIs or internal tools, and completes a task such as resetting a limit, opening a dispute, or drafting a KYC review for a human to confirm. The test is one question. After the interaction, did anything change in a core system? If yes, it is an agent and it needs agent-grade controls.

4Which Chatbot Use Case Should a Bank Turn Into an Agent First?

Pick the workflow where a human already approves every step, the volume is high, and a mistake is reversible. Internal operations beat customer-facing money movement for a first agent: KYC and AML triage, document validation, and case pre-filling carry a human reviewer by design, so the agent drafts and the person signs. Avoid starting with autonomous payments or credit decisions. They carry the highest regulatory tier and the smallest margin for error. The branch or back-office assistant that already proved value as a chatbot is usually the cheapest, safest place to earn the first production win.

5What Controls Does an AI Agent Need Before It Touches a Core Banking System?

Four controls are non-negotiable before an agent reaches production. First, scoped access: the agent can call only the systems and fields its task requires, through a service account with least privilege. Second, human approval on any step that moves money, changes a limit, or affects a customer’s rights. Third, a full audit trail: every input, decision, tool call, and output logged so a supervisor can reconstruct what happened. Fourth, a kill switch and rollback path, so the agent can be stopped and its last state reversed in minutes. Build these as the platform the agent runs on, not as features added once it is already live.

6Does Moving From a Chatbot to an Agent Change a Bank’s EU AI Act or DORA Obligations?

Often, yes. A pure information chatbot usually carries light obligations, mainly the Article 50 duty to tell users they are talking to a machine. An agent that participates in credit scoring or insurance pricing can move into the Annex III high-risk tier, which brings risk management, logging, human oversight, and conformity duties, with standalone high-risk obligations applying from 2 December 2027. Separately, any agent that reads or writes to a core system falls under DORA as ICT: it must be inventoried, tested, and covered by an exit strategy. The safe approach is to classify the target workflow under both regimes before you build, so the controls match the tier the agent will land in.

7How Long Does It Take to Get a Bank’s First AI Agent Into Production?

A first production agent is realistic in about 100 days when the scope is one workflow and the controls are built from the start. Ableneo plans to that window: weeks 1 to 3 to pick the workflow and classify it under the EU AI Act and DORA, weeks 4 to 9 to build the agent with scoped access, approval steps, and audit logging, and the remaining weeks to test against real cases and move to production with a human reviewer. ROI typically starts to show from six months once the agent is handling steady volume. Pushing faster by skipping the controls is the most common reason a bank’s agent stalls in an audit later.

8The Ableneo Perspective

Ableneo has delivered 49 AI projects, with 35 or more running in production across 26 clients in finance, energy, and telecom, and is an ECB-cleared IT vendor. That record is the point here: the step from chatbot to agent is where most pilots die, because teams add autonomy without adding governance. Ableneo builds the controls first, which is how a bank’s first agent reaches production in 100 days and holds up when a supervisor reads the audit trail. See the Ableneo AI transformation FAQ for the regulatory and delivery detail behind each step.

Key takeaways

Sources

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

Talk to Ableneo