Sibos 2026 in Miami gave us the opportunity to hear from Operations and Technology leaders who are leading the way on agentic AI adoption. A demo can win over a room in minutes, but getting an agent through a bank’s governance, cyber and compliance reviews takes much longer.
Yet things are moving, and fast. The industry has gone past the hype, experimented a lot, perhaps been oversold a few solutions. Now all that testing is beginning to pay off, as green shots of productivity start emerging, especially around AI.
Joining our CEO, Christian Nentwich, on stage for a Meet the Experts session were:
- Catherine Thrasher, Chief Operations Officer, CIBC Mellon
- Ken Utsonomiya, Chief Information Officer, Global Head of Operations, Mizuho Securities
- Tom Elliott, Chief Operating Officer, HSBC
The panel talked through where they are on their agentic journeys, how they decide what to tackle first, how they bring risk and compliance teams with them, and how they know it’s working.
Planning how to get your own agentic projects from pilot to production? Here’s what our panellists have learned on their journeys so far.
The state of agentic AI adoption
To set the scene, Nentwich asked each panellist how far they’ve got with agentic AI, and where they’re putting it to work first.
Thrasher explained that CIBC Mellon has been deliberate about its overall transformation programme, and particularly about how it uses agentic AI. Across AI more broadly, the focus has been on empowering people by giving them a new tool to help with “the work no one wants to do”.
That focus makes sense for a firm that brings in a lot of graduates and digital natives. Some of the work they inherit leaves them scratching their heads, not least when it starts with learning what a fax is, she said.
Our first real agentic use case is with reconciliation. We wanted to start with a very well contained, well understood process where we could prove that, overcome any of the risk hurdles that we’ve encountered, and our plan is to be live in the coming month.
Utsunomiya has both Operations and Technology under his remit, so he has a very wide view of this particular topic, Nentwich observed. So what were his thoughts on where he’s deploying agentic applications versus where he’s building them?
“From an IT point, we have been working heavily on AI-driven development,” Utsunomiya said. “Especially in the equity area, we have a lot of competition, so faster delivery and continuous improvement is really key. Also the US is driving a lot of Anthropic Claude code development, together with Copilot.”
However, he also noted that legacy technology is still slowing progress. Like many banks, Mizuho has “tonnes of legacy” software, often based on antiquated programming languages like COBOL, and the firm is now working on a large AI-led transformation to tackle it.
For now, agentic AI is largely limited to the Operations side. Mizuho has just started using Duco’s agentic workspace for reconciliation tasks, although in the retail side of the business the team has also begun working on internal call centre-style services that answer questions automatically.
For Elliott, who has large resources both on the IT and Operation sides under his remit, the operative word in this discussion is ‘outcome’.
The objective is not - it sounds obvious - to use agents. The objective is to get work done with better outcomes: faster speed, higher quality, less error rates, even better accuracy, and ultimately a better client experience.
As an example, he pointed to Smart Checking, which HSBC launched at Sibos for the trade side of its operations. Documentary credits are central to how global trade gets done, but processing them is still a very manual job. In most cases they still arrive as stacks of paper, carrying hundreds of data points that have to be ingested, processed downstream and reviewed by people. Many of those data points are unstructured.
Smart Checking uses technology including agentic AI to automate much of that work, though Elliott was clear that agents aren’t running the whole process. It’s an exceptions-based workflow, involving hundreds of checks.
Many of those checks are built into the system’s architecture. Different types of agents handle tasks such as querying and evaluation, with confidence scoring helping to move the work along. Exceptions then surface in a newly designed interface, where HSBC’s trade experts review them.
“We’ve been working for around 18 months, with a combination of buy and build for a more proprietary internal solution,” Elliot said. “The outcome is reducing unit processing time, being faster for clients, and the performance levels are much better than we can offer without some of this tooling.”
It’s hard to get these things into production environments, but once you get there the potential is huge. We’re already seeing significant benefits that ultimately accrue to our clients in the experience they have with us.
How to decide where to deploy agentic AI
Potential, Nentwich suggested, was one side of the matrix when it comes to prioritisation. Another axis is “pain”, which covers governance, cost, risk and plenty more besides.
“Every COO and CIO has probably 50 projects they could be doing with agentic; some of them are high impact and almost impossible to achieve, and some of them are low impact but easy,” he said. How did our panelists think about that trade-off? For Elliott, the key was to pick something that you cared about.
One important factor in getting from pilot to production is to pick something that’s meaningful, because then you will invest the time.
This is why documentary credits fitted the bill: HSBC processes more than a million of them every year and, by some measures, handles twice as much of this business as any other bank. It’s a fully global operation with a huge amount of client flow, and its manual, operationally intensive nature means it takes significant resources to run. Potential was the starting point, with the team seeing a chance to make “a really big impact”.
“That’s why it’s been an 18-month journey to get to this point, it’s not been something like we took a few LLMs, we had the idea, we dropped it in and a month later we had a perfect piece. And that’s a latticework of everything from the governance, the technology design, the iterative testing, cyber… these are long, long journeys, but I think the efficacy is starting to come through now, if you invest the time in building these things.”
Thrasher agreed.
“We went after the big ticket items first,” she said. “We hit the area where we knew there was a lot of opportunity for real cost reduction. We knew it was going to be hard, but we took the time to put in place the right tools to really measure the process today and where it had breaks, gaps and bottlenecks that we could automate. And some of that automation ended up being with AI, and some of it ended up being with more traditional tooling.”
The last point is key. For Thrasher, the aim is always to find the right solution to the problem.
When I look at our portfolio of transformation opportunities - not necessarily looking for only an agentic solution - we’re looking for the best tool for the job, and AI is one of those tools.
Of course, if you focus on the areas of biggest impact, you are expected to have a big impact. However, at the same time, getting those “big ticket items” makes it easier to prove the value of agentic AI and therefore go further.
“I do think you have to show success with some of those big ticket items, and I almost think of it as permission to continue. I’m not necessarily asking for permission for every agent that we want to build, but if we are continuing to deliver the core services day-to-day, to the near-perfect SLAs that we need to meet - 99.9% accuracy and timeliness - and we’re continuing to show success on reducing cost, driving transformation, driving real change in the business and delivering a positive operating leverage, we just keep going and we keep adding those new tools to that portfolio of solutions.”
How to overcome internal blockers and get deployed
Once you’ve picked a use case, you still have to get it live. Nentwich shared with the audience how, that week, we’d switched on our agentic platform for 120 companies. “But that tends to be less regulated companies where you can just click the switch,” he said. In a regulated institution, every new technology has to clear risk, compliance and governance checks first. Those controls exist for good reason, but they can make the road to production a long one.
“When you go to production with a new technology like this, normally within a bank there’s a lot of people employed to stop you doing this,” Nentwich said. “Who will tell you ‘This is not a good idea; here are 100 reasons why it’s not going to work’. So how do you overcome that? Is it the size of the business case, or is it the leadership that say ‘No, we have to do this - here’s why’?”
“The most important factor for us is ROI, whether how much revenue we can make, or cost reduction, how much we can save,” Utsunomiya explained. However, one of the biggest blockers for getting agentic AI live is compliance. “Even the compliance officer doesn’t have enough knowledge about AI technology. It is very tough for us to ask the compliance officer ‘Can you do this?’.”
This is particularly an issue in Japan, where banks are “super, super conservative” in general, he explained. In the US, Mizuho is solving the knowledge problem with a dedicated AI compliance officer who “fully understands AI”.
“This person is working inside the compliance department on what kind of potential risks we are going to have. From an IT viewpoint, we can then talk in the same manner. That is really good. Thanks to this setup, the US has quite a lot of new technology, a lot of AI, and is quite advanced because the AI compliance officer can manage the compliance people’s expectation and also risk appetite.”
They are now trying the same thing in Japan. IT, Utsunomiya said, is a good proving ground for agentic technology, because the risk is lower when the deployments are not client facing.
Risk mitigation is not 100% perfect, so in this case, as management, can we take a risk or not? Depends on the risk appetite. From our viewpoint, for IT ones, yes we can take more risk. For client facing, it is really challenging. So that is a starting point for us to prove there is no impact, then go to clients.
Elliott echoed the fact that it is easier to get agentic AI into production when it doesn’t interact with clients.
“Our agents aren’t touching our clients in this deployment,” he said. “It’s a client transaction, but how that gets moved through and decisioned and processed has got agentic deployment along the journey. That helps, because if you’ve got agents interacting directly with clients then the risk tolerance is very different to something that isn’t going to touch them directly from an interactive perspective.”
How does Elliott approach breaking out of the endless cycle of governance meetings and getting something live, Nentwich asked. Is it leadership or client demand that is the defining factor?
Like Utsunomiya, Elliott stressed the importance of working closely with compliance and risk teams.
You’ve got to bring your risk and compliance people in from the start and you’ve got to bring them on the journey. I think we’ve had things done in series often when we engage with risk and compliance colleagues: we’ll build it, design it, we’ll get to the end and then we’ll present it. You can’t do that for these things because it’s such a paradigm shift.
“You have to have them involved early on, whether they’re part of your pods as you’re doing agile deployment, or other mechanisms, but they need to be involved early on to see what’s happening.”
Elliott also highlighted the importance of explaining the similarities between the task an agent will be doing and existing tasks.
We already have straight-through-process in some of these areas, we already have machine learning being deployed with certain logic in certain areas, and on the outcome piece, we’re not always here saying we’re going to full agentic autonomy. Sometimes this is augmentation.
“We retain human decisioning and the ultimate outcome is processed by the human; it’s the facilitation of the different agentic and AI capabilities along the way that are making it even more efficient and augmenting the job of that human so they can review the exceptions and the bit that needs real expert decisioning.”
This helps, he said, because “people hear ‘agentic’ and they think ‘Wow, we’re going to complete autonomy’” - whether they view that as a utopian or dystopian outcome. But the reality is that a lot of these tasks are already happening and are already automated. In many cases “we’re just enhancing what our experts do and the ultimate decision rests with them”.
He highlighted that the same checks and balances still exist across the documentary credits process that he outlined earlier, such as quality assurance samples. The team still pays attention to the same things - error rates, throughput, audit logs. “The inherent product risk has not changed, it’s just how the work is moving through is a much better experience for our colleagues and also ultimately for our clients, who benefit from the transaction fulfillment in an even more efficient way.”
Measuring agentic outcomes
As Elliott highlighted earlier, the key with agentic AI is the outcome. “To convince a company to invest in this technology you need to commit to certain metrics and ROI, and you have to know if you’re making progress or not,” Nentwich said. “Sometimes there’s too much of a proliferation of metrics and it helps to think about it and simplify.”
Our panellists obviously expect to see efficiency gains from their agentic efforts, but what are their leading and following indicators that they pay attention to?
“We expect that this will deliver a positive impact to our operating leverage,” Thrasher explained. She highlighted hours saved as a metric, but noted that the issue here is that “those hours saved just tend to disappear”.
We will track hours, but we’ve found that those hours saved just tend to disappear in some of the work that we’ve done so far. So it’s really important for us that we can turn those hours into quantifiable increases to capacity or quantifiable savings to deliver to support new business, for example.
“That’s been a critical thing, starting with the real clear measurement of what that process looks like today and how we could translate that into a shared service or something where we could translate those hours into real scalable savings.”
“It’s still a positive thing - it’s great to have an hour back in your day, but if we can’t turn that hour into productivity then it’s not generating the outcome I need to see.”
The metrics that are harder to track, she said, are things like risk reduction or things that they weren’t “really able to do” before, like anomaly detection capabilities. “It’s just harder to prove in a business case, but we know it’s there and we know that’s some of the ways that we can use these capabilities to change our model, not just to prove the existing model, because we just couldn’t have human eyes overseeing that amount of data on the level of frequency that we need to do.”
But, ultimately, the overall metrics they use to run their business don’t change. “We’re still tracking the same SLAs, the same service delivery metrics.”
Elliott agrees with Thrasher’s points. “We’re both COOs, so we know that the holy grail of Operations is ‘Can you take more work without adding capacity to the platform to process that work?’ So creating operating leverage, or whether you call it zero marginal cost, etc.”
Even before AI became a thing, the question was always “How can you get more bang for your buck through the Operations platform?” he said.
“Our platform - I have thousands of people in my team globally - is a volumetric game. What are the volumes we have coming through? What are the unit processing times for those different bits of volume? And then what’s the workforce platform you need to process that?”
This is particularly important in the current climate, he said, with trade finance revenues climbing sharply as the AI boom drives huge capital expenditure flows across the economy. This is boosting trade growth by 20-30% year-on-year for goods related to the AI industry, from computer chips to precursor chemicals and raw materials.
“That means for many of our businesses in trade finance, we’re seeing a lot of revenue growth, we’re seeing more volumes come through. How can we process more of those volumes as they flow through our business in the most efficient way where we’re not having to add linearly more people to the workforce?”
He said there has been a decoupling of process volume and cost, some of which is driven by AI, while some of it is the “bread-and-butter disciplines” such as process reengineering.
Some of the AI enablement we’re then able to add to that technology stack and data architecture is creating more capacity. We’re seeing benefits, and I think you’re only going to see more and more of this as these things come into production - not even in our industry, above and beyond. And we’re going to see more of that decoupling, so that delivery of operating leverage or lower marginal costing as you add volume.
For Utsunomiya, as someone spending a lot of money on applications and infrastructure, cost is an important metric.
Cost is very important. Mizuho is one of the largest Japanese banks, so our IT investment every year is huge. So, how are we going to reduce some of our cost, which means how we can do some of the development by ourselves, to reduce costs? We have a lot of initiatives, a few of which were not achievable in the past, but now thanks to new AI in the last year, we can do more. We have a lot of confidence now to reduce cost for the future.
Cost reduction doesn’t mean head count reduction, though. Japanese labour laws mean that Mizuho must retain their staff, so Utsunomiya has to balance automation with finding valuable work for the people whose time he is freeing up.
“My purpose is not to think of cost reduction and headcount reduction, but rather that Mizuho is trying to be a top ten global bank. So we have to do so many new things simultaneously.”
On top of the traditional ways of freeing up capacity, such as encouraging his staff to cut down on meetings, Utsunomiya is now asking them to think about how agents can perform tasks on their behalf.
“Eventually we’ll have more capacity, then we take more jobs, we can even start insourcing some operations from other banks to give us more jobs. We have AI, then I have to think how we can utilise our staff.”
Bringing agentic Operations to life
The journey to agentic Operations is not quick, but as AI understanding and usage matures, the hurdles should shrink as those greenshots of productivity serve as solid examples of the potential. HSBC has spent around 18 months getting Smart Checking into production, and between them, the panel has had to work through legacy technology, governance hurdles and questions of risk appetite. But their experience gives you a much clearer route from pilot to production.
Start with something meaningful enough to justify the investment. That might be a well-understood process like reconciliation, or a high-volume, manual one like documentary credits. Success on those meaningful projects earns what Thrasher called “permission to continue”.
Bring risk and compliance colleagues in from the start, and give them the knowledge they need to assess the technology with confidence. Be clear about where agents support your experts and where a human still makes the call. Measure success against the metrics your business already runs on, and make sure the hours you save turn into real capacity for growth.
To wrap up, Nentwich asked our panel what their outlook is for AI in 2027; what will happen, and what do they want to get done.
“I think we’re going to prove out some of our key use cases, in particular our reconciliation use case, over 2027,” Thrasher said. Additionally, CIBC Mellon is in the final year of the three-year Operations strategy they set out, so “it’s really about elevating now”.
We feel like we built the framework that we needed and this is going to be the proofpoint year. So we’re expecting big things.
For Utsunomiya, it’s not just about expanding agentic AI, but also working on the foundations that underpin them. “We have been working on data definition, because securities houses normally have a lot of confidential information. If an agent looked at some information that would be quite a risk for us. So we’re promoting agents more, but at the same time we are having to do a lot of data definition and governance, labelling - that is a challenge in 2027.”
Elliott had three key things on his agenda for 2027: “More in production in the genuine sense of the word. Better, measurable outcomes where we can see the benefits. Very lastly, hoping for much more work with Duco!”
Want to know more about agentic Operations? Click here to read the value 12 firms have already unlocked in production.