Commerzbank put a transacting AI agent in its banking app: 30,000 conversations a month, ~75% resolved without a human

Last Updated: October 6, 2026

The problem: a service channel that could not keep up with the app

Commerzbank serves around 11 million private and small-business customers in Germany and transacts roughly 30% of the country's foreign trade financing. Its retail service model runs across branches, a remote advisory centre, and a mobile app. When the bank first announced the avatar project in November 2023, it framed the mobile channel as the beachhead: the assistant would be developed for mobile devices as a first step, for the target group of 2.2 million active banking app users.

The underlying problem is familiar to anyone who has run a retail bank's contact centre. The high-volume requests — what is my credit card limit, I have lost my card, can you raise my limit for a holiday — are simultaneously trivial in content and non-trivial in consequence. They are identity-sensitive, they are regulated, and they arrive at 11pm. Routing them to human agents is expensive; routing them to a 2018-era intent-classification bot produces the deflection theatre customers have learned to bypass by shouting “agent” at the phone.

Commerzbank's own framing in the Microsoft write-up is blunter than most vendor case studies allow. Gerald Ertl, who runs the bank's Strategic AI Program, describes a gap between expectation and capability:

We were seeing a necessity to build a bridge between what our customers expected and what our systems could deliver. We needed a new way to serve—not just faster, but smarter and more human.

— Gerald Ertl, Managing Director, Head of Strategic AI Program, Commerzbank AG, Microsoft Customer Stories

What was actually built

Ava is a conversational agent that lives inside the Commerzbank banking app. It is rendered as a photorealistic avatar — the bank says Ava is modeled after the likeness of an actress — and it is available 24/7 in natural language. That is the consumer-facing surface. The engineering underneath is a fairly standard enterprise agent stack, assembled from Microsoft components.

From the Microsoft customer story and the parallel Foundry Labs entry, the published architecture is:

Two design decisions are worth isolating. First, Ava was given write authority over the credit card lifecycle from the start. Commerzbank's launch release lists ordering a new credit card, blocking or unblocking a current credit card, and changing limits as in-dialogue actions. Most banks that shipped an LLM assistant in 2024–25 restricted it to read-only Q&A precisely to avoid this. Second, the safety layer is treated as an employee-equivalent control, not a content filter bolted on at the end. Denise Reffelmann, the business product owner, puts the standard explicitly:

Security and trust are foundational. Ava doesn’t just talk to customers—she acts on their behalf. That means she must be as secure and compliant as any human employee, which is why we rely on Content Safety plus customized filters to monitor and help protect every interaction.

— Denise Reffelmann, Business Product Owner, Commerzbank AG, Microsoft Customer Stories

The measured outcome

The 2× development figure deserves a note, because it is the metric integration teams most often ignore. Microsoft attributes it to the combination of GitHub Copilot, Foundry tooling, and agentic frameworks, together with streamlined build-and-release processes and automated testing incorporated into the DevOps pipeline. In other words, part of the reported gain is AI-assisted coding and part is ordinary platform engineering hygiene. The case study does not separate the two.

The governance move most teams skip

The May 2025 appointment is easy to read past, but it is structurally the most transferable part of this deployment. Commerzbank did not leave Ava as a product owned by a digital channels team. It put bank-wide AI responsibility and data management under one executive reporting into the board. Christiane Vorspel, the COO with responsibility for IT, framed it as integration rather than innovation:

By appointing Oliver Dörler as Chief Data & AI Officer, we are ensuring that our central themes of data and AI are strategically integrated. This will enable us to offer our customers improved products and services while optimising our internal processes more quickly and effectively.

— Christiane Vorspel, Member of the Board of Managing Directors and COO, Commerzbank AG, Commerzbank newsroom

Technical product owner Christoph Jakfeld describes the same discipline at the architecture level — modular, transparent and highly secure, so the team can build fast without creating risk. That phrasing, speed through auditability rather than in spite of it, is the pattern worth copying. Teams that treat compliance as a release gate ship late; teams that treat it as an architectural property ship repeatedly. If you are scoping something similar, our notes on production automation and model selection cover the same trade-off.

What the public record does not tell us

This is a well-documented deployment by the standards of European banking, and it is still thin in the places that matter most to a buyer.

No customer satisfaction data. Microsoft notes that Microsoft Fabric is being implemented to persist and analyse conversation flows so the bank can measure outcome quality and continuously improve customer satisfaction. That phrasing implies the satisfaction measurement loop was not yet fully in place when the story was published. No CSAT, NPS, or complaint-rate figure has been published for Ava.

No cost figure. Neither Commerzbank nor Microsoft has published a cost-per-conversation, a headcount effect, or a return-on-investment number for Ava specifically. The Microsoft story says the automation brings considerable relief to call center agents — a qualitative claim, not a measured one.

No accuracy or incident data. There is no published hallucination rate, no wrong-action rate on card blocking or limit changes, and no record of how often escalation happens because the agent failed rather than because the query was genuinely complex.

The metrics are vendor-published. The 30,000 and 75% figures appear in a Microsoft customer story and a Microsoft Foundry Labs entry, with Commerzbank executives quoted by name. We did not find those figures in Commerzbank's own press releases or investor materials. Named-executive attribution raises confidence; it is not the same as a bank-audited disclosure.

Scale context is missing. Commerzbank identified 2.2 million active app users in 2023. 30,000 conversations a month works out to roughly 1.4% of that base per month — our arithmetic, not the bank's. That is a real production system, but it is not yet the primary service channel, and no source states what share of total retail service contacts Ava now carries.

How this compares to the common alternatives

What an integration team should take away

Scope by consequence, not by difficulty. Commerzbank picked the credit card lifecycle — high volume, emotionally urgent, and crucially a domain with clean, well-defined, reversible operations. Blocking a card is a high-stakes action with a low-ambiguity trigger. That is a much better first agent target than open-ended financial advice.

Build the audit path before the conversation path. Content safety, grounded retrieval, persistent context and escalation all appear in the published architecture as first-class components. In a regulated environment these are not optimisations; they are the reason the thing is allowed to transact at all.

Expect a long runway and say so. November 2023 announcement to April 2025 launch is about seventeen months, and the bank released Ava gradually to devices rather than flipping a switch. Anyone promising a transacting banking agent in a quarter is describing a demo.

Treat the first agent as a platform decision. Microsoft reports that Commerzbank now sees Ava's framework as a blueprint for the bank's next wave of AI initiatives, from internal agents that support employees to advisory agents. Ertl's own summary of the shift:

Foundry Agent Service gives us the building blocks for the future and has changed how we think about service. Ava offers expert knowledge 24/7 without waiting time and empowers our teams, because they can now focus on what they do best: building customer relationships, solving complex problems, and driving innovation.

— Gerald Ertl, Managing Director, Head of Strategic AI Program, Commerzbank AG, Microsoft Customer Stories

Note what that quote does and does not claim. It claims availability and focus. It does not claim cost reduction. Executives who have the savings number usually publish the savings number. For more deployments read the same way, see our case study library, or talk to us about scoping your own.