Banking technology has typically focused on digitizing transactions and speeding up processes. Agentic AI could change the scale of that work: agents can interpret information, plan, take action and coordinate across systems. That opens the door to redesigning whole workflows, while raising questions about data, accountability and where people need to stay involved.
Previous technology waves largely focused on digitizing processes and automating transactions. Agentic AI extends those possibilities into tasks involving interpretation, planning, reasoning and action. Mehmet sees agentic AI as a shift beyond automating individual transactions. AI agents can work across systems and coordinate activities, giving banks the opportunity to rethink entire workflows and operating models. As he explains, “we can now interpret information, generate content and make recommendations.” That ability to take on cognitive tasks is what sets this wave apart.
As those capabilities develop, strategies and operating models need to evolve alongside them. Katharina highlights the importance of clear decision rights, reusable platforms and strong application programming interfaces (APIs) that allow systems to connect securely. The discussion captures a practical message: reliable data and a robust banking operating model are essential for successful AI implementation.
Banks are beginning to concentrate their AI investments on high-volume, repeatable processes that involve substantial amounts of unstructured information. Mehmet identifies financial crime investigations, customer onboarding, know your customer (KYC) processes, lending, underwriting, customer servicing and trade finance as areas attracting attention.
These use cases offer opportunities to improve customer experiences while measuring changes in processing speed, efficiency and conversion rates. Selecting the right model or platform starts with understanding the business problem and the value the bank expects to achieve.
Andy, Mehmet and Katharina also discuss matching model size to the task. For specific applications, a smaller language model may be more economical, making operating costs part of the business case.
Scaling requires more than model selection: trusted data, clear ownership, access controls and model risk management provide the foundation. Banks also need business teams equipped to monitor performance, manage incidents and escalate problems to a person when necessary.
The discussion explores a spectrum between people-led, AI-assisted work and fully automated processes that require no human intervention. The appropriate level of automation depends on the task, the risks involved and the bank’s ability to maintain control.
Andy points to payment repair as a process that could support extensive automation when both the rule violation and the required correction are understood. Decisions that require more judgment or affect customers more significantly need stronger human involvement.
Preventing errors and bias at scale means treating AI as part of the complete business process. Katharina emphasizes representative data, clear decision rules, testing across user groups and reliable ways to investigate unusual outcomes.
She emphasizes that governance should continue after the go-live since models, data, and customer behavior change over time. Ongoing monitoring, human escalation, and the ability to stop an AI-driven process help banks maintain oversight as they expand their use of agentic AI.
Regional differences influence how banks approach automation and customer trust. Mehmet describes the UK approach as emphasizing principles and outcomes, including accountability, fairness, transparency and robustness.
He expects banks to expand automation where risks are more controlled while maintaining human oversight for sensitive areas such as lending, fraud outcomes and dispute resolution. From a German perspective, Katharina highlights expectations around privacy, explainability, reliability and control.
Meeting those expectations can also create opportunities for banks to differentiate their services. Trust becomes a feature of the customer experience, supported by how the bank governs and operates its AI capabilities. Katharina also points to the pace of change:
“AI is advancing much faster than in previous years. Many banking strategies are already outdated, and operating models are struggling to keep pace.”
That pace makes it essential for banks to keep their strategies and technology assets current, continue learning as AI capabilities evolve and identify where they can differentiate.
Listen to Andy Schmidt, Mehmet Eryilmaz and Katharina Dubberke discuss how to scale agentic AI while maintaining trust. Part 2 explores AI-driven customer experiences and interactions.
- Chapter 1: How agentic AI can reshape complete banking workflows
-
Andy Schmidt:
Hello and welcome to the latest episode of Transactions to Trust, the financial services podcast from CGI. My name is Andy Schmidt. I'm the Global Industry Lead for banking at CGI. And today's topic is going to be agentic AI. Many of us just came back from Sibos in Miami. And of course, agentic is one of the hot topics in banking. And I'm joined by my colleagues Mehmet and Katerina. And this is the first of two episodes we're going to have on this topic, talking about the moving AI frontier. We'll also be talking about moving from POCs to scale. And of course, a very important part of AI overall: trust and governance. So, why don't we start off with the first question? And Mehmet, if you could lead us off and introduce yourself as well. What makes this AI transformation different from previous technology waves?
Mehmet Eryilmaz:
Thank you, Andy. I'm a partner with CGI's banking and finance team based in London. I predominantly work with clients on business strategy, helping them bridge that gap between vision and technology and execution. So, we have been really working heavily on the AI front. And what sets apart, I think, the AI transformation from other waves of technology change that we have seen is that, in the past, it was more about often automating transactions, digitizing processes, trying to move faster, better with access to information. But I think AI is more profound because we now have a richer and more malleable toolset. We can now build skills, reuse the tooling, push that boundary more.
And probably what also sets it apart is we used to have tools that we had to decide on what they did. But now, increasingly, we can perform things which we used to call uniquely human, which are more the cognitive tasks, as well. We can now interpret information, generate content and make recommendations. So, really moving from simply prompting to using AI agents that do the planning, reasoning, executing actions, and collaborating even with other systems, as well. In terms of a technology wave, it has opened the possibility of not just attacking one part of the enterprise but actually redesigning the entire workflow and the operating model itself.
Andy Schmidt:
Okay. And Katarina, we'd love your thoughts on the next question. Why does the possibilities frontier keep moving?
Katharina Dubberke:
Hi, Andy. Thank you for having me here. I think the frontier keeps moving because the technology, as we just see it and already feel it also, within CGI as well as with our clients, is that the development is going much faster than in the last few years. Most banking strategies are already old. Operating models cannot keep up with AI. And that's the reason why we have at the moment so many questions and topics around AI.
Andy Schmidt:
As you point out, the opportunities to rethink business processes, rethink how we organize, how we manage the bank is also changing thanks to developments in AI. What I've seen is the banks that take advantage of that are the ones that are going to get the best impact, the greatest results from their AI investments.
Katharina Dubberke:
Exactly. The idea is that when you have a clear strategy around your data, your clear decision rights, reusable platform, especially strong APIs, then it will work pretty well for you to handle the AI topics. But as we know, many of our customers are still struggling with old legacy stuff. We need to check, okay, is the data the right one? Is it doubled? Do we have duplicates and stuff like this? So, if the data, which is key, is not ready, then we cannot discuss AI. To put it in other words, to bring it back to the Back to the Future topic again: the DeLorean is optional, but the robust operating model for banking is key.
Andy Schmidt:
Excellent. And what new capabilities should executives be watching or watching for before they place their bets?
Katharina Dubberke:
My expectation is that there are multiple steps to take. First, yes, the AI agents can do a lot of work. They can plan, they can complete tasks, but the models that need to work with the structured banking data, the documents and rules, they need also to be in place. And the third capability is that we need secure orchestration across the several systems within the banking chain. And it's also a definition of where this chain stops. Normally, we have our APIs in place, but as we already learned within the block, when agents connect directly to the APIs and select the bank with the best data, this process stops.
- Chapter 2: How high-volume, repeatable processes can move AI toward scale
-
Andy Schmidt:
We talked about the moving AI frontier. Let's pivot now to moving from POC user proof of concept to scale, moving from experimentation to measurable business outcomes. In your opinion, where should banks focus first to create a clear business impact?
Mehmet Eryilmaz:
Andy, tying this to your earlier question on the frontier, this is a really interesting time because technology change swings like a pendulum. When we introduced a new technology, it swings all the way to the right. We try a lot of things, and often part of that fails. So, the pendulum swings back and we find that area where we have a lot of business impacts, the right use cases, the right improvements that we bring. I think with AI, the pendulum is swinging farther each time.
So the question that you just asked in terms of, well, how do we scale this? How do we take POCs into scale becomes even more difficult because my view today is, I'm sure, going to change slightly. But a lot of banks that we're working with are looking at this question and I would say, on the onset, had a broader approach. And they asked their organizations and came up with tens, sometimes hundreds of cases that they wanted to try. So, slowly we see what I would call a convergence across the high-volume, repeatable processes with more unstructured information that needs to be analyzed, which seems to be the kind of prime candidates from that POC to scale.
Initially, I think all my clients looked at the software development cycle; AI bringing engineering improvements with significant gains, and that has probably been the beachhead. But for an executive, there are considerable business functions, especially in the UK: fraud detection, financial crime investigations, perhaps on the customer onboarding KYC, definitely on the lending operations, underwriting, customer servicing. At CGI, we work a lot on trade, especially on the trade side, and on document optimization in the back office. They seem to be kind of the hotbed where a lot of attention is being focused. So, from POC to scale is going to go more. What we're seeing is more happening on those specific business verticals.
Andy Schmidt:
Those high-value customer experiences, you know, onboarding, underwriting, et cetera, and making those as smooth as possible, make a ton of sense. And the customers will sense or be able to understand the improvement to their experience working with the bank, but also the bank will be able to measure how much more quickly these processes operate and what the conversion rates are and how those improve, as well. So yeah, absolutely great advice.
Katerina, you know, just taking that a little bit further, when we're thinking about this topic, how do leaders choose between targeted models and broader platforms?
Katharina Dubberke:
That's a tricky question, Andy, because you cannot just say yes or no to each of them. I think the choice of our customers should follow their business problem, not the technology fashion. Yes, no matter who you ask, everybody does AI. But if you dig a bit deeper and you're asking, okay, what exactly are you doing there? You recognize quite quickly that most of them are struggling and combining the topics. As I already said, data is key. That means data has to be well understood, and the outcome should be measured clearly.
In the end, you should handle it like a normal project with a dedicated business case, because we all know tokens are raised and not paid for by themselves. It needs to be clearly checked if the business case is really the one that the bank wants to go with, and that it is really valuable for the bank. And also, the other buzzwords, you need to have the monitoring, the access control, the model governance that also needs to be in place. So, maybe generally speaking, the AI guardrails have to be in place to make sure that either way, it works out for the banks.
- Chapter 3: How trusted data, governance and operational readiness enable scale
-
Andy Schmidt:
Along with that, choosing between large language models, small language models, if it is a very specific use case, picking the right tool for the job is vitally important. But if it's a more targeted use case, having a smaller language model that uses fewer tokens, for example, can be much more economical, much more efficient in terms of giving you the measurable outcomes that you're looking for.
And taking that a little bit further, in terms of data, in terms of foundations, et cetera, what data and operating foundations must be in place in order to scale your AI operations, your agentic bank?
Mehmet Eryilmaz:
Definitely AI scaling challenge is ultimately that data operating model challenge rather than the core technology. I think that goes hand in hand. But it's very, very difficult to really get to that perfect data set. And competitively, when you look at this, you need to move fast. Can you really afford to wait to get all the data aligned? Often that's very challenging. And on the other side, you can't really make it work without some baseline data. So, you need several things to work hand in hand: data, your ability to have some foundation, and the guardrails.
So, a couple of things that I think must have are the trusted and governed data foundation, the clear data ownership and who is accountable, how do I manage that in the organization, the governance framework and the model risk management processes. I think that's the essential basis.
Katharina Dubberke:
We need to also take into account that AI is, most of the time, seen as a bit technical. So, we need to have the LLM in place, we need to make sure that the system's working properly. The data has to be in place, yes, but if it comes to a problem, for example, later on and in the go-live process, et cetera, then we need to make sure that also the operations part need to be in place for the banks. So that the responsible business functions can recognize, okay, there is something wrong. We can step in and, for example, testing, monitoring, having the incident management in place. And, for example, if we have an AI agent here in our use case, that's some kind of human escalation is also in place in case something goes wrong.
- Chapter 4: How risk and customer impact determine where human oversight matters
-
Andy Schmidt:
Absolutely. The off-switch, the combination of responsible AI and an off-switch is going to be key for sure. Having those foundations in place does tie in nicely to our next topic, trust and governance and focusing on deciding what to automate, what to keep human-led and how to govern the operations overall. So, where should banks use a human?
Mehmet Eryilmaz:
We have been working with some of our clients, and this question comes up often. And I think it's a really interesting future that we're living in right now. If you look at this and think of two ends of that spectrum. On one end of the spectrum, there's no human intervention at all. And that's what we call a dark factory, a place where you don't even need to put light bulbs because the factory will never have human intervention on the process. The other hand, you have the no-AI automation. Most of the functions are using an AI as an assistant, but they're human-led. Now, what organizations are exploring is kind of how far along this journey should we or could we move to being a fully automated dark factory in our processes? And the question becomes also: is that even desirable? Is that risky? Do I trust the technology to move in that direction? And also, in a regulated industry like financial services, am I even allowed to do this?
So, imagine for a second, as a financial services institution, that means that you have an agent that decides on your overall strategy, builds the requirements, builds the code, regression tests it, deploys it, does the acceptance, rolls it out, and communicates with the customer. Now, not many are comfortable with such a scale. So, we're seeing that evolution and small migration of work across requirements, build code, test deployment, but again, with many guardrails and checks with a human in the loop. So, we anticipate that there are multiple paths and we advise it as well. Some processes where full automation could be explored, and others where there's going to be, perhaps, for a while strong intervention.
Andy Schmidt:
Right. So, for example, payment repair could be heavily automated because you know what the rule break is and you know what the repair is, and so you can make that a regular action, whereas things where human decision is made, or you have to have more aggressive, shall we say, human intervention or human involvement. Thinking about that. What safeguards prevent banks from automating bias or errors at scale?
Katharina Dubberke:
I think the first safeguard would be to treat AI as part of the complete business process, not as some kind of magic component that is placed within the middle of one process. So, by doing so, the banks need to have representative data, clear decision rules, testing across different kinds of user groups, ongoing monitoring, and have then a reliable way to investigate unusual outcomes. And what is really, really important is that it should not stop when you go live. Also, the governance should continue after the go live. Because models change, data change, and also customer behavior change. You need to make sure that also after a go-live, the agent is still resilient.
- Chapter 5: How regional expectations shape trust in AI adoption
-
Andy Schmidt:
This has been a great conversation. And now, I want to switch gears a little bit and look at regional differences. How do regional differences, including the UK market, given that it interacts with the EU but is outside of the EU, change the approach?
Mehmet Eryilmaz:
Good question. I think in the UK, the regulators have generally adopted the principles-led approach. What that means is you need to focus on the outcome, such as accountability, fairness, transparency, robustness, the consumer product. That's still there. That means that organizations retain responsibility for the decisions even if AI is involved.
Banks are absolutely very cautious in the UK because customer trust is still a very, very critical factor. While customers may be comfortable with AI and supporting operational efficiencies, they're less comfortable with fully autonomous decisions, especially around lending, financial well-being, fraud outcomes, and dispute resolution. So, I think we are going to continue seeing a balanced model in the UK, expanding automation, perhaps a bit more aggressively, where the risks are more controlled and low, but definitely maintaining human oversight, because fundamentally the principle is that organizations retain that responsibility.
Katharina Dubberke:
In regard to Germany, I expect that, as well as our customers, the regulators tend to place their emphasis on privacy, explainability, reliability and control. So, that's a really high bar. We are aware of that, but it can also become a competitive advantage for the banks. And it's a really huge power that they can turn into trust. So, for example, into a product feature. And that makes the whole topic quite interesting.
Andy Schmidt:
Absolutely. I mean, trust as a feature is definitely going to be key as we advance with Agentic and as we advance with banking and financial services in general.
Thank you both, Katarina and Mehmet, for talking with me about some of the strategic choices that banks need to make in terms of where to place bets, what to scale, and what to keep human-led.
And thank you all for joining us for the first episode in our discussion on the Agentic Bank here at from Transactions to Trust a Financial Services podcast. I'm your host, Andy Schmidt, and we will see you in episode two, where we want to focus on the customer experience and how AI becomes a channel where clients and agents interact.
Thank you very much.
