TL;DR: Greek banks are using three different forms of AI that are often placed under the same agent label, although they do not have the same authority: conversational assistants answer questions or route customers, AI-assisted workflows classify, analyse and draft work for employees, and action-taking agents connect to operational systems and complete bounded tasks.
National Bank of Greece has published the clearest examples of the third category, including services that recover banking credentials, manage lost-card incidents and prepare internal lending documents, while Alpha Bank's current voice deployment is mainly a conversational routing system and Eurobank is building an agentic AI platform while its publicly described applications still look mostly like assistants and supported workflows.
Across all three banks, and across the Greek companies using similar technology, sensitive decisions such as fraud handling, loan approval, money transfers and final regulatory sign-off remain with people or are being introduced only in controlled stages.
What counts as an AI agent
The term AI agent is being used very broadly by banks and technology suppliers, so it helps to separate the systems according to what they are actually allowed to do, rather than according to the name used in a press release. A chatbot can sound natural without having permission to act, while a system that looks less impressive on the surface may be connected to several internal tools and may complete a substantial part of a business process.
• Conversational assistants understand questions, retrieve information, explain a process or route a customer to the correct person, but they do not normally change an account or complete an operational task on their own.
• AI-assisted workflows classify cases, summarise documents, analyse information or draft material that an employee reviews, and they can remove much of the manual work even though the final decision remains human.
• Action-taking agents connect to operational systems and carry out bounded steps, such as booking an appointment, resetting a service, submitting a prepared filing or executing an approved transaction, while operating within permissions and escalation rules.
These categories can overlap, because one assistant may answer a question in one conversation and execute an authorised action in another, and the same product may move from conversational support to operational execution as new integrations are introduced. The distinction used here therefore follows the capabilities that the banks and their suppliers have described publicly, and it does not assume that every product marketed as an agent already performs autonomous work.
European banking gives the Greek examples their context
The European Banking Authority's monitoring shows that AI is already common in banking, because 92% of the EU banks in its sample were deploying AI in some form while the remaining 8% were piloting it or discussing use cases, and 55% reported using general-purpose or agentic AI in processes that face consumers directly. The most common applications include fraud detection, support for call-centre staff, automated guidance for digital services, financial-education tools and customer assistants, including systems that communicate by voice.
The EBA also identifies the conditions that determine how far these systems can go, including dependence on third-party providers, the quality of the data available to the model, customer consent, human oversight, explainability and the legal or reputational consequences of an incorrect answer. Those concerns matter when the Greek examples are compared, because the important difference is not whether a bank uses AI, but whether the system only communicates, prepares work for an employee or has permission to act inside a banking process.
National Bank of Greece has the clearest action taking systems
National Bank of Greece is the most developed public case because it has described several layers of Sofia and has also listed internal uses across the front, middle and back office. Its July 2026 results presentation calls Sofia a customer-facing agent with more than 14 services in production and about 200,000 requests a month, while the same presentation describes a voice agent handling more than 130,000 requests a month and says that the bank expects it to absorb more than half of contact-centre traffic within six to nine months.
The public Sofia is mainly an informational assistant
The version available on the public nbg.gr website should not be treated as evidence that every Sofia interaction is agentic, because the bank's own terms describe it as an AI-enhanced chatbot and informational service that provides general information, helps users navigate the website and answers questions about the bank's products and services.
It can also book a branch appointment, which is a limited operational action, but it does not provide personalised or binding service, users are told not to enter account numbers, card numbers or access credentials, and the bank states that this public service does not make automated decisions or build customer profiles.
The authenticated and voice versions can complete bounded tasks
The position changes when Sofia operates inside authenticated digital channels or the contact centre, because coverage of the bank's July briefing says that the digital version can support account and card management, customer onboarding, branch appointments, payments and subscriptions, while the new voice service can recover or reset Internet and Mobile Banking credentials, manage incidents involving a lost or stolen card and route a call to the correct team.
When a human employee is needed, the system transfers the conversation history as well as the call, so the employee can continue from the point already reached instead of asking the customer to repeat the entire request.
This combination belongs in the action-taking category because the system is not limited to producing an answer and is connected to processes that can change the state of a service, although its authority remains narrow and is expanded in stages.
NBG has said that balance and transaction enquiries and bill payments are being introduced before more complex services such as money transfers, and each capability is released only after reliability, quality and security testing, while the fraud-incident line remains exclusively under human management.
Many of the internal agents prepare work rather than decide
NBG executives have said that more than 40 agents are already operating in production for customers and employees, but the bank's own results presentation shows that the word agent covers several forms of work.
It lists agentic classification and routing of complaints together with response drafting, credit-memo preparation for corporate and small-business underwriting, bond-loan contract drafting, and an end-to-end software-development engine built on Claude Code agents, so some systems execute multi-step technical work while others prepare documents or recommendations that still enter a human decision process.
The distinction is especially important in lending, anti-money-laundering and fraud prevention, because using AI to gather evidence, draft a memo or identify a suspicious pattern is not the same as allowing the model to approve a loan or decide that a customer has committed fraud. NBG's public material supports the conclusion that its agents are doing significant operational work, but it also shows that their permissions depend on the channel, the identity of the user and the consequences of a mistake.
Alpha Bank is using conversational AI before operational execution
Alpha Bank's voice assistant uses ElevenLabs technology for speech and Moveo.AI for conversational understanding, and its published purpose is to let callers describe a request naturally before being directed to the appropriate adviser.
The system replaces a conventional multi-level telephone menu, supports Greek and English, and is intended to extend the same voice into the bank's e-banking and mobile-banking chatbot, but the public description does not say that the voice assistant can reset credentials, block a card, make a payment or complete another banking action.
For that reason, the present Alpha deployment is better classified as a conversational assistant, even though the underlying ElevenAgents platform can support more advanced agents and even though the system may become more capable after it is connected to additional bank services.
Alpha said that the assistant would eventually receive the roughly 4 million calls handled by its contact centre each year, which makes the scale important, but scale does not by itself turn routing into autonomous execution.
Alpha had already used Moveo.AI for its text assistant, and a December 2024 supplier statement said that it was answering more than 5,500 enquiries a week with an 88% resolution rate, although resolution in this context can mean that the customer received an adequate answer without human escalation and does not necessarily mean that the system completed a banking transaction. The distinction should remain explicit until the bank publishes the operational actions that the assistant is authorised to perform.
Eurobank is building the agentic platform before proving the applications
Eurobank announced an AI factory with Fairfax Digital Services, EY and Microsoft in November 2025, and the bank describes the project as an agentic platform designed to unify data, automate processes and support systems that can act with greater autonomy. Its 2025 annual report confirms the establishment of the AI Factory and the launch of the EVA customer assistant, while management said at the Banking Forward event in June 2026 that the wider technology programme includes investment in cloud, data platforms, core banking and cybersecurity through 2028.
The applications that have been described publicly, however, should not all be called agents. EVA serves customers through digital channels and myEVA supports employees, while other AI applications help assess mortgage files, process contract documents, analyse customer feedback and develop new digital services, so the available evidence places most of these uses in either the conversational-assistant or AI-assisted-workflow category.
The sources do not say that EVA independently executes banking actions or that an AI system approves mortgages, and the AI factory should therefore be presented as infrastructure intended to support future agentic systems rather than proof that those systems already control core decisions.
Outside banking the distinction becomes easier to see
The same three categories appear in Greek telecommunications, gaming, energy and commerce, and the differences are sometimes easier to identify because suppliers describe the integrations in more operational detail. These cases also show why a voice interface or a natural conversation is not enough to define an agent, because the decisive question is whether the system is connected to tools that let it complete the requested task.
OTE and Cosmote TV
OTE uses a Wonderful voice agent on part of the Cosmote TV support line, and the supplier says that the system integrates with telephony and back-end APIs across ten skill areas, including subscriber authentication, troubleshooting, satellite resets, package and price explanations, profile management, service barring, programme lookup and post-call surveys. Because it can authenticate a subscriber and change or reset services, this is an action-taking support agent rather than a voice system that only answers questions or routes calls.
Wonderful reports that the agent reached a 50% deflection rate, almost three times the initial level, and reduced average handling time by 30%, while OTE executives quoted in the case study say that it began handling tens of thousands of calls in less than two months.
Those figures come from the supplier, and the same page shows a different use case and a different headline containment figure in its sidebar, so the operational description is useful but the performance numbers should remain qualified.
Kaizen Gaming and Enerwave
Kaizen Gaming uses Moveo AI across 19 markets for player-support conversations involving withdrawal status, verification steps and account problems, and the supplier says that the system connects to more than ten external systems while actions are controlled by market-specific rules.
That description suggests an integrated support agent, although the public case study gives more detail about contextual responses and governance than about the exact account changes the system can make, so it is safer to describe the deployment as a mix of conversational support and governed workflow execution.
Enerwave has given AI an outbound role by using a Moveo AI collections agent to contact former customers who still owe money, and the vendor reports a return of 19 times the amount invested.
Debt collection is a sensitive use because the conversation concerns payment and can affect a vulnerable customer, so this example is important not only because the system initiates contact, but also because it shows that companies are willing to automate a process with reputational consequences when the workflow can be constrained and monitored.
Skroutz
Skroutz operates two AI products that should not be placed in the same category, because Vlassis is a Moveo AI customer-support system that the supplier says connects to back-end services, while Skroutz's own shopping assistant compares product features, prices and reviews and recommends what a customer might buy.
The first may combine conversation with operational support, depending on the actions exposed through its integrations, while the second is primarily a conversational recommendation tool and behaves more like a salesperson than an operations agent.
Payroll shows how an agent can act while a person remains responsible
RSM Greece introduced payroll agents with the technology partner DGTAL in May 2026, and the company says that they process payroll data, run initial calculations, generate and send reports, and check and submit the APD social-insurance declaration filed with EFKA.
This is an action-taking workflow because the system moves beyond analysis and prepares or submits operational output, but RSM also describes a human-in-the-loop model in which people retain final control over decisions, which matters because a payroll filing has legal and financial consequences.
The state has contracted an action taking agent but has not announced its launch
In September 2025 the Ministry of Labour awarded 01 Solutions Hellas a contract worth EUR 5.7 million including VAT to expand ERGANI, and the project includes a generative-AI assistant in the myErgani application and website that is intended to answer questions and carry out actions for employees in areas such as employment relationships, leave and social-insurance contributions.
Helvia.ai is working on the system as a subcontractor, and the planned authority to act inside an employment record would place it in the action-taking category if the delivered service matches the contract description.
ERGANI II itself entered full production on 16 February 2026, but no public launch announcement for the AI assistant was identified during the review of the available sources, so the accurate description remains contracted rather than live.
That is a larger step than mAigov, which was launched on gov.gr in December 2023 to help citizens locate services and was designed not to request sensitive information such as passwords or payment details, because the ERGANI system is supposed to perform actions within a citizen's employment record rather than simply guide the user to information.
The dividing line is operational authority
The Greek banking cases do not all draw exactly the same line, because NBG has already connected AI to bounded customer and internal actions, Alpha is beginning with conversation and routing, and Eurobank is investing in the infrastructure that could support wider autonomy while its current public examples remain mostly assistants and supported workflows.
What they share is a controlled progression from information, to preparation, to execution, with permissions expanding only when the bank can identify the user, restrict the action, test the system and preserve an escalation route to an employee.
The most mature deployments are therefore not simply the systems with the most natural voices, because they are the systems that can use tools, interact with back-end services and complete a defined part of a process without being given unlimited authority.
Conversational assistants are already operating at significant scale, AI-assisted workflows are preparing complaints, credit documents and mortgage files, and action-taking agents are handling credential recovery, card incidents, service resets, appointments and regulated submissions, although the final responsibility for fraud cases, loan decisions, sensitive transfers and payroll approval still remains human.
The next test will be whether NBG's planned expansion into more complex voice transactions, Alpha's movement from routing toward resolution, Eurobank's AI factory and the ERGANI assistant produce systems that can complete more work without weakening security or accountability.
That is the point at which the market will be able to distinguish an agent that genuinely performs banking work from an assistant that makes the existing process easier to navigate.
How Advisable can help: Advisable helps banks and businesses move from conversational assistants to governed, action-taking systems through its AI Solutions, including AI Agents & MCP Development, AI Automations & AI Integrations and Custom AI Solutions for Enterprises. For teams that want to build these skills in-house, the Advisable Academy offers the AI for Business seminar and AI training for business.
Resources
European Banking Authority: Rising application of AI in EU banking and payments sector
National Bank of Greece via Athens Exchange: 2Q26 Financial Results, 30 July 2026
National Bank of Greece: Terms of use for the Sofia digital assistant
Oikonomikos Tachydromos: NBG launches AI voice service in its contact centre
CNN Greece: How agentic AI is becoming part of NBG's strategy
Investing.com: National Bank of Greece 2Q26 earnings call transcript
Oikonomikos Tachydromos: Alpha Bank introduces an ElevenLabs voice assistant
Economix: Alpha Bank plans launch of its new AI voice assistant
Moveo.AI: Alpha Bank AI assistant update, December 2024
Eurobank: AI factory project with Fairfax Digital, EY and Microsoft
Eurobank: Annual Financial Report 2025
CNN Greece: Eurobank's investment in agentic AI
Wonderful: OTE and Cosmote TV voice-agent case study
Moveo.AI: Kaizen Gaming customer-support case study
Moveo.AI: Enerwave collections-agent case study
Moveo.AI: GenAI agent for ecommerce
Fast Forward: Skroutz launches AI shopping assistant
RSM Greece: Agentic AI in payroll outsourcing
ICTplus: AI assistant included in the ERGANI expansion contract
Greek Ministry of Labour: ERGANI II enters production
Insomnia: Launch of the mAigov digital assistant
