Beyond the hype: Practical AI for smarter payment exception handling

Hello. Welcome. I'm Teresa and I'm moderating today's webinar. Thank you for joining us for this webinar with Finextra in association with Pega. Today we're going to be exploring a subject that's really at the heart of payments but is sometimes overlooked, if I'm honest with you, and it's beyond the hype. Practical AI for smarter payment exception handling. And before we start, just a few words, if I may, on this subject, payment exception handling is one of those areas that doesn't always get the spotlight that I think it deserves, but it sits right at the heart and the intersection between operational risk, client experience, working capital and regulatory scrutiny. And then as payments. And this is a context of say, as payments volumes grow and they become faster and more real time, even small inefficiencies can scale really quickly into material problems.

Problems like trapped liquidity for clients. Operational costs. And of course, customer dissatisfaction. So today is very deliberately beyond the hype. We're not here to talk about AI or generative AI as a silver bullet, and we won't be giving futuristic visions detached from operational reality. What we'll be focusing on, or the panel of experts will be focusing on, is ways in which institutions are reducing exceptions where AI and other tools are already delivering practical value. And also what's coming up on the horizon in that respect, and how innovation can be deployed safely in our highly regulated environment. In addition to me asking our experts questions today, you can too, either using the Q&A function on the screen here in front of you, and I'll do my best to blend as many of them into the conversation and put them to our panelists as possible. They will always.

There will also be some polls coming up throughout the session. Please take a moment to respond, because it helps ground our discussion into what's important to you as, say, in addition to insights from our experts, information is available for download as well. On the screen you'll find the download for two documents the next great leap leaders haven't made quite yet, and also a genetic automation for payment exceptions and investigations. So without further ado, I'm delighted to be graced by a really strong panel. We're in for a treat today, so let's meet our panel. And Sean, if I could ask you please, to start the panel introductions. Thanks very much, sir. Three very good afternoon everyone. Sean Noonan is my name.

I'm the global head of investigations, uh, based for based in Dublin, Ireland, working for Citibank. And very delighted to be part of today's panel discussion. Thank you. Thank you. Tim, can I come to you, please? Yes. Thank you. So my name is Tim Long. I'm a product owner in the payments and savings domain of Rabobank.

Um, so my main attention, um, point is payments or sorry, payment, uh, exceptions. Um, so I got like 14 years experience, uh, within the payments domain, um, mainly on processing and exceptions. Um, and it's, um, really fun to see that, uh, with my background in sales in tech and in logistics from before, that is really, um, where you really touch business and it, uh, together and making that crossing that border. Um, yeah, I see a lot of that in my, my current job, and it's perfect. Thank you. And Kelly. Hello, I'm Kelly Wilson, I'm an industry principal at Pegasystems where I focus on commercial, corporate and investment banking use cases and commercial payments. I've been at Pega for six years. I have a background in global banking and markets where I worked in sales and trading, relationship management and a CEO role at Pega.

I spend about 50% of my time working on our smart product, which manages global payment exceptions. Swift estimates that actually about 90% of their global E and I are touched by a Pega instance because the majority of the largest transaction banks in the world use Pega today. So this is a great topic that we're very interested in. We've been using, you know, workflow and intent led automation, parsing, etc., other types of AI for years in the tool. And many of our clients have achieved high rates of straight through processing. And we're really excited to see what new developments in AI can bring to even improve those high rates of automation. Thank you. We can see why you're on this panel. And also 90% no pressure and five panelist says, we've had a question in a multi-part question in already of which we will win.

I feel that in our discussion, we will we will tap upon, uh, the questions are what are the most innovative payment trends? And we're going to concentrate here on exceptions. We're going to look at traditional tools as well as sort of current AI and future AI and other tools. Um, we say, what should the first strategic step be for banks as well? And we're going to touch upon all of those. So to start, let's start by discussing how financial institutions can leverage upfront tooling to reduce exceptions, exceptions. And Tim, coming to you first, if you if we look at the bigger picture, when you zoom out at industry level, what tools are there then to reduce exceptions? Yeah. Yeah.

What I see a big part is, is standards. Um, so what you see in payments, we got the Sepa and the Swift standards. Um, specifically on Sepa we see fairly limited exceptions, um, because the standards are really strict and the data is structured. Um, and you see also that in Swift um is pulling also the lack uh by that's in English sounds different than I mean it. But um, by also introducing the ISO standards and CBR plus and we already see some benefits of that. Um, I guess if you talk about preventing or reducing exceptions, it starts with data quality, um, structured data field validations. As soon as somebody touches or brings in an exception, um, the data should basically be locked and, and secured. Yeah. That's it.

And there's something there, isn't there, about sort of free format fields and people being able to put in a whole myriad of different responses. And I know that we were speaking earlier about how labor intensive all of that is as well for, for operational teams. And when I think about that sort of the labor demand, are there other areas as well that kind of where we're for exceptions. It demands more labor. Yeah. So I think within and outside the bank. So we just discussed the point. You brought the point up about about the three fields three, four months. So like typically the older Swift formats have that opportunity.

It's a lot of people see it as an opportunity. But it's basically in the end the limiting factor I guess. Um, um, so you have outside of the bank, you have that, but also internally, um, you see somebody from the first line support putting in a ticket, it gets transferred over. Um, so it's if that is not directly translated in, in structured data, it's already taking a lot of time to do analysis on that. Um, it is also comparing, um, the incoming payment to all kinds of, um, payment messages that were used in the payment factory itself during the process, Comparing these two analysis. What went wrong? Um, that is taking really a lot of time. Um, again, it's it's the multiple sources and it's unstructured to structured. Uh, the longer that you delay that or the less you do that in the process, the more labor you you need.

Yeah. I love the way that you describe those those three format fields as an opportunity, but also for, you know, for the client, I guess it might be an opportunity. But when it comes to processing and actually then delivering upon what that client has asked to happen, it's it's far more problematic. Um, Sean, I'm going to come to you next and I want to, to sort of to, to zone zone back a little from an industry perspective and talk about from an institutional perspective, um, what tools can reduce exceptions and where have you seen success? So maybe just to add on to Tim's point, first of all, obviously Tim referred to, you know, data quality there in relation to it. So obviously with the migration in the payments world from MT to WM format, that's giving more richer data within the actual packs away Paxil online payments trees so that more richer content of data. Hopefully we'll see from an industry perspective bringing additional benefits, for example, in the reduction of maybe funds transfer regulatory queries when it comes to, you know, full name and address and stuff like that. So certainly from our perspective, when it comes to the investigation side of things, we we break it down into two opportunities. So one, from a technology point of view, but also then in relation to potential process engineering and client education.

So maybe I'll touch on first of all in relation to the the process engineering and the client education side of things. So maybe the process engineering, we had some good wins in in relation to some maybe legacy offerings that we had for clients. So for example, uh, in some scenarios where returned payments were coming in, we would give the client an option to send an amendment. And basically using data, we were able to to see that in many cases, the client didn't utilize the the amendment offering. They got the funds returned to them. So we made some process changes in relation to that, that reduced risk on our side, reduced aging cases and definitely increased some productivity. In addition to that, based on, you know, deep dive of particular clients where we saw high volume, we saw opportunities to work with the clients and educate them. So, for example, if there were payments that they were making to a destination country where a payment code was included and they weren't doing it, and it was generating a query from the downstream bank, working closely with those clients to ensure that they were making the necessary amendments within their own payments to to drive more systemic FTP and query reduction. So I'd say those two lenses being very important.

Knowing your data, knowing what's driving the trends, which of your clients are contributing it, how we can work with our clients with potential opportunities to remediate the queries in the first place, working with clients then obviously for for self-service opportunities. So maybe one example being in relation to the self-service side of things. We still know today that a large number of our clients were sending us empty 1993 format messages for debit authority. So this is where they'd received the credit. They wanted to return to the market and instead of doing a direct return themselves, they were coming in issuing a 199 requesting us. And that obviously is a manual touch point. It's a it's a case file. Of course it's done on the maker checker. But as with all payments there is the potential risk element.

So you know, we've been driving many of our clients and have an active work stream in relation to pushing clients to instruct them through Paxil for. So using trees of the industry best practices to our benefit as well in educating clients to, you know, move away from the legacy thought process when it comes to investigations, that everything can be generated through a 199 so as we migrate both from the payment side of things and from an investigation side of things to more structured messages, structured data, working closely with the clients in relation to that point, so that angle trees in relation to what I call the non-tech workstream i.e the client education, pushing self-service And then in addition to that, you know, from our perspective, when you're talking about, you know, upfront looking to remediate, reduce payments, what's critical from my side is in relation to a work stream within city that we're calling in relation to payment validation. So based on very detailed analysis from our side of over a million returns, we were able to see that 40 to 40 to 50% of return refunds queries relate to some form of issue or problem with beneficiary details i.e. beneficiary account number, invalid penny account, closed account name, number mismatch, etc.. So if we're able to validate the accuracy and validity of beneficiary details before we process payments, that will be a big opportunity for us to reduce timely returns and be able to notify the customer straight away that there's a potential problem. That program of work is in its infancy. We are making good inroads and progress, and certainly over the next year or so, we really hope to be able to roll that out further to our in-flight payments, where we can look to identify potential problems real time. Thank you. Teresa.

You're welcome. That's a big number, isn't it? Of for 1,000,040 to 50% as well. So cracking something like that is going to bring, you know, back to Tim's. Tim's point about sort of the labor intensity required for exceptions that that's that's going to sort of delight customers more but but also help support the bottom line. Kelly, first of all, thank you for your patience because because so so whilst we're we're still on this sort of this up front tooling type piece, um, if we get data and standards right, okay. How far then can AI actually go in preventing exceptions in real time? In the short answer, it can go very far. Um, to get into some detail, we could talk about the different types of AI that are available.

So for instance, today a lot of machine learning and NLP used to parse and scrape and read swift messages and emails that come in asking what's happening with this message and then defining what it how it needs to be fixed based on the data and the questions in those swift messages and or emails. As Sean and Tim both mentioned, a lot of those are empty. 199 so the parsing can get you so far, but it can't get to as high accuracy as hopefully we'll be able to use when we're looking at ISO structured data so that the system can very clearly look for account numbers, look for the reason for the issue, etc. and more accurately define the process and then resolve the issue. There's also a lot of potential for using AI with both LMS and Agentic AI agents. So for instance, an LLM is where a client could be asking a tool what should I do here? What's what's the problem, what's the question? Etc.. And the LLM can be directed at internal processes and procedures to answer that question.

We see a lot of clients using that today already to take advantage of AI. But that is guiding the person. When you then can add in Agentic AI, you have a whole other set of capabilities that are going to be able to look at what's happening in that case, in that exception, and define the problem and quickly, autonomously resolve it. Obviously, it can be reviewed by humans afterwards, but really a ton of capabilities that are coming in that space. Um, like Shannon mentioned, if 50% of the returns have incorrect beneficiary details, a lot of those are going to be repeat payments. So if the bank is able to save the data of a a previous payment and then look at that payment and say, we received these many details for this payment, but it looks like we completed this payment a month ago with different details. Then they could choose to automatically fix the details and send the payment on its way, or even have the benefit of sending that, um, sending what we believe are the corrected details back to the sending bank and ask them, you know, this seems incorrect. Can we use these details instead? So either way, depending on the comfort level of the bank, there's capabilities using that historical data, analytical and predictive AI to resolve something that potentially had been already in error that had been resolved.

So it requires saving data, tagging data, Etc. but the ability is definitely there. Um, what we see is that a majority of payments can be resolved within 15 minutes if action is taken using data using technology. So in that time frame, usually the client has no idea that there was even a problem. So then you're improving customer service. You're using the various AI so that we believe the technology is there. It's getting to the comfort level of using it, um, using the various forms of AI and technology together. That's perfect. Thank you.

And you repeat it. And you have responded actually to a question from the audience which said, can we repeat the stats, actually, of the percentages of payments which ended up as investigation payments? Um, although there is a nuance on this actually. And Sean, it might be based upon your example. So the question is on the stats shared, which end up on the investigation processes and percentage of patients which end up returned versus resolved. So there's something about how many do get resolved and how many get get returned. I know you said you were early days into the project, but is there an answer you can share on that or not at this time? Yeah. So just just to repeat the question.

Sorry. Sure I will. Yeah. And apologies because Sean's driving. Driving. I hope you're not driving. You're joining us by phone and it's on the statue. You said the payments which end up in investigation cases. What percentage, though, ended up returned versus which ended up resolved?

Perfect. You might not be there yet. I don't know. Yeah. No, but I think in general terms, right. In relation to the return of funds is our biggest case type that we have within city. Obviously there's two particular use cases for return of funds. One, when a we make a point, a payment on behalf of our client to a downstream bank. Um, what tends to happen Trees is when the beneficiary bank receive us, and there is a problem with the beneficiary data.

The account number is invalid, account name, number mismatch or whatever. Today they raised that 199 query. Uh and if we can get additional details from our client you know that may resolve uh, the scenario. So what we basically see is case types can evolve. So what may start for example, as a beneficiary claims Non-receipt query can subsequently turn into a return of funds. The way that we want to be going from an industry point of view is that if the payment is made correctly with all of the, uh, information and there is a problem on the bank side, for example, with the validity of the data, so they can't process it because there's a problem, um, in relation to the account number or something like that. The best practice is to for the bank to return that payment, and then the remitting party can can instruct and send a fresh payment order. So certainly from our side, I think we see the only thing today, um, is that is, uh, helping the sort of return volume, uh, not be really more excessively high is that scenario where, you know, banks are today offering, um, you know, empty one nine, nine in relation to, to changing over information, etc.. So I think to answer, you know, very clearly when it comes to the investigation types, I think more banks are starting to move now towards that.

You know, if details are invalid or incomplete. Exactly. They're just going to send a Paxil for a return message that way. That's the actual way that the industry is certainly going, but certainly from our side when we'll come to the actual, you know, volume of payments that would be, uh, returned, uh, you know, that are queried from, from a downstream bank, you know, it could it could be around the sort of 50 to 60% end up getting returned anyway, and maybe half of them can be resolved through additional details. Okay. Thank you. And I and I guess if somebody wants to follow further on that, they can contact you via the usual channel, which I guess is LinkedIn. But we'll see. Um, before we go to talk further about AI, I want to bring up our first poll question, if I may.

And this this is to help us get a an idea and a barometer of the level of activity. And the question is this at what stage is your institution regarding identity key for payments exceptions. Now, this reminder of Kelly's very good description there, which I wrote down for a TKI. And it's to autonomously look at exceptions, define the problem and resolve. Now that's resolved, whether it's resolved directly and independently or via a human type route. So the question is at what stage is your institution regarding Agentic AI for payments exceptions? The options are not rolled out and not really seriously considering it. Or we're just investigating or planning for a 2026 deployment or a limited implementation so far. Or actually, we're really on it.

Extensive genetic AI already rolled out, or at least a portion of it rolled out. So we'll just have a moment or two longer. And then if we could, we'll have a look at the answers if we if we may. So we'll just wait for those two to come on before we do because I'm going to actually what we might do is we might come back to those in a couple of minutes. So apologies. Um, so you will have longer now to answer that question. Um, just whilst you've still got the floor, I want to change the focus now to risk, um, so obviously banks can't afford risky processes and exceptional protection, handling reconciliation or triage. It's just that the risk appetite is zero. Um, but what proven strategies are there for deploying AI without adding risk to clients working capital?

Because at the end of the day, clients want money going from A to B for whatever reason. So so so what's the operational reality? I think today when it comes to investigation query management, you know, many banks still take that linear approach for payment investigation. So for example if there's four banks in the chain each of the bank will contact the next bank in the chain maybe requesting an update requesting a status update. So I think as an industry there hasn't been too much evolution in relation to investigations, communication, investigations, resolutions. And you know that process then, um, contributes to, you know, delays investigation resolution. It can contribute to client frustration because obviously you have a remitting party and a party and you have your intermediary banks, right? So if you're not resolving queries promptly and quickly, it obviously can can create that frustration. I think one important thing is for the group to to be aware of is probably over the last three years, um, you know, a number of institutions have been working closely with Swift.

Uh, and our colleagues from, from Pega have been at the table right from the outset as well, just in relation to the evolution of what we call case manager, which basically is, you know, moving away from this non-structured empty. 199 to more structured mm ISO messages that will be called camp camp cash management messages, where each of the queries will have its own coding four letter code to differentiate what the actual query type is. The query types will also then be resolved with the assistance of old GP data. So basically trace out that the payments payment queries will be routed to the respective bank in the chain, likely to have the answer based on GP data. Rather than going down the chain. Yeah. Exactly. Today it's gone that linear approach where in the future world for certain case types like beneficiary claims, loan received unable to apply, Swift themselves will systemically route the query to the most likely party to have the answer based on GP data. So, for example, if you have a beneficiary bank that has received a payment and their client is saying, you know, maybe it's missing an invoice number or something like that.

Today, what happens is the 199 gets sent to the corresponding bank chain. The bank need to work with their clients. Well, in this new world, when all banks are on board, and hopefully by the end of 2027, that query type will be systemically routed right down to the underlying remitting bank, who will, uh, have the opportunity to to resolve the query and, and engage with the beneficiary bank directly. So I think that's definitely going to be a key insight. So sorry. That sounds wonderful. And I'm conscious that I think all of you have kind of gone, oh that, that that 199 you know, everything else. So, so from what you've described there with sort of the case manager, um, is it the death knell then for the 199? I mean, when will this come in?

Yeah. So from our side, we're one of what we call the early adopter banks. Right. So working with Kelly and the team. We have two branches that are live from a city perspective city Dublin and City New York. There are a number of other a number of other banks that are also early adopters in relation to it. At this stage, it's more in relation to, you know, going through the capability, making sure that the, you know, the messages can be sent and received. So some, you know, pilot production exchanges between ourselves and some other banks. Um, so from an industry perspective, for the initial case types, which are better claims on receipt, unable to apply requests for information and unable to execute those for initial case types, all banks will have to be in a position to receive the message by the end of November of 2026, and to be able to send and receive these new letters by the end of 2027.

So I think the focus has obviously been on the payment landscape change obviously since November 2025. Now you're seeing a lot more momentum and shift towards investigations, exceptions and investigations. So those dates are dates that all banks will obviously need to be working towards to ensure that they're ready. But the opportunity is that it's going to bring Citi, it's going to bring Rabobank, that it's going to bring all institutions globally. I really think is going to be a game changer when it comes to actions and investigation space. And we often use Game Changer quite liberally, don't we. But but I but I agree, I think that that's kind of you know I'm over ten. So I use the word literally correctly. Um, so you know, so, so, so I think literally that that could be a game changer.

I think we should look at the poll results. So now if we may. So, um, here we go. And as a reminder, the question was, um, where are you using AI at the moment? Genetic AI for payments exception so not rolled out and not seriously investigating 14%. Just investigating half 5,050%. Planning a 2026 deployment 21.4%. A limited implementation so far 14%. And nobody yet in the camp that says extensive genetic AI already we're there.

We're sorted. It's all happening. Kelly, what do you make of those results? Those make sense to me because that's about 65% not ruled out or investigating. Um, people are, you know, taking this looking at gen AI with kid gloves, making sure firms can trust the models, etc. so that makes sense. I'm hearing that 15% have a limited, 14% have a limited implementation. That's great. Early adopters moving forward.

Um, I know both City and Are early adopters for specific use cases. But it's also changing quickly. I mean, I was actually with a client last night who's a new client, and they recently said, no, Jen, I to begin with, we don't want any of it. And then they've already started building the product. And last night they're like, no, we want to add it. We want to improve customer service. So it's changing quickly. Um, firms are at different stages. Model risk management teams are getting up to speed and becoming more comfortable with the various agents.

So these numbers make sense. But I bet in 3 or 6 months those 65% that aren't rolled out are not just investigating, will have converted to planning maybe a 2027 deployment. That would be so interesting, actually, to repeat it in sort of six and 12 months that that would to see the trend. And Tim, does that resonate with you? All Tim we can't hear you at the moment. I'm sorry. I had some background noise. I put myself on mute. Sorry for that.

So, um. Yeah. Within Rabobank, um, there is a lot of AI initiatives ongoing. Um, so the just as a governance, the bank has set up, uh, AI board and basically everything is run by a very strict regulations, um, just to secure, um, the way we do it. Um, so basically, if you want to try out, do a proof of concept or whatever, you always you go via the board and follow those rules. Um, I see quite a lot of implementations already. Um, for example, within our financial and economical crime department, there is a lot of use of AI. Um, there are, um, AIS that help you to, to basically ask questions, um, about stuff that you would normally search in your manual. Um, they're working on code body, uh, Uh, functionality.

Um, so it's a lot of AI, uh, for employees. Um, so what I see in the short term for my domains or the payment exceptions, um, I guess at the base, again, it's optimize the integration between systems and, and get as much as possible structured data. Um, I think it's very easy to start using process AI. Um, that is also a way to, to let not have it uh, an it driven, um, uh, it will not be an I.T driven action. We're going to go on to talk about that. Why wise words. Absolutely. Yeah. Um, yeah.

And Sean, the poll results, um, do they resonate with you? Yeah. I think it's fairly reflective of what we see right across, uh, other forums and industry discussions that is going on. Certainly, you know, from a from a city perspective. And, you know, within our own case management platform, we, we, we use Pega. You know, we have activities that are live today. Um, but there are some which are I would call maybe some of the more foundational additional, uh, elements of bringing AI into our use cases. But we know that there's a lot more that we want to to do as the months evolve. So certainly resonates with ourselves as well.

Lovely. Thank you. And, Kelly, I need you. If you were to pick up the thread from. From what what Tim was saying because, you know, from a solution lens and a solution sort of implementation and development lens as well. You know, Tim said there, you know, it's not just an IT thing, is it? And and rightly so, but who needs to be involved then to avoid unintended risk or indeed any risk that particularly unintended risk. We see the best projects are most successful when business and IT are both closely involved for the development process, for the process optimization plan, and they need to stay in lock step and stay in agreement. A lot of times when we start a project, one or the other is leading it and sometimes dragging the other one along, but it's definitely much more, much better when it's collaborative.

When you're designing a solution and optimizing a solution together, also, that makes sure that the business team is getting what they want from it. And change is hard. A lot of times we hear, well, you need to build it this way because this is how we've always done it. Or compliance says we have to do it this way. And because I work across many global clients, I'll say, and I was in banking myself, I was like, well, you don't actually have to do it that way. Your compliance department is telling you, you have to do that that way. But we know you don't. And then that sometimes requires pulling in compliance and trying to get changes approved, which is a very challenging process and not one that people want to take on in their in the spare time they don't have. But it often leads to better processes and more efficiency, better user experience, you know, reducing redundant steps that maybe were needed in an analog world but not needed in a digital world.

So we see really keeping those two IT and business involved and working closely together. And then transformation teams, strong managers who will who will push back when people are just trying to do it the way it's always been done. That's really needed too, because sometimes the implementation partner or the IT team, they can't push back themselves. So they need the managers or the transformation teams from the business to really be forward thinking, embracing change, etc.. Um, from a development standpoint, there's a lot of new gen AI tools. Pega has one called Pega Blueprint that is a great tool for collaboration at design time so you can feed in existing processes, Visio diagrams, even media files that talk to the process and then that will spit out what the optimized, what a new process would look like in stages and steps. And that is a great tool because it accelerates design time. It increases collaboration. It's easy for the business to look at and understand easier than Visio diagram or other other ways that it might present a solution to the business.

And then they can edit it together and collaboratively. So tools like that are really helping those two. Those two teams work together in a way that maybe wasn't as easy for the business to engage and stay engaged in the past. And when you're describing that, though, the word that was in my mind was holistic. You know, you put everything in and it gives you a holistic type answer that everybody could, can, can look at concurrently. But, Kelly, where where do you start then. Do do do you start from from your existing rules and processes? Or do you start from somewhere completely different? You can do either or.

You can use the existing things and optimize it. Or you can say, let's start fresh and then make sure we're designing to adhere to compliance requirements or or business requirements, etc.. Um, probably depends on on where you are today, right? If it's a very old antiquated process, maybe just start fresh. If if it's something that's been, you know, kept up to date fairly well, then then start with the existing process. Hard to answer unless you see the specific. For the specifics. But but, but but in general you know you can make it fit and and Tim I'm I'm sort of keeping keeping on on on the dark side here with risk. Um, what has to be true from a government perspective before AI can safely operate?

Yeah. So if I, if I look through, um, basically the journey that a that a payment exception takes. Um, there's a lot of decisions, uh, a lot of steps taken. And that has to be auditable. Um, so I think one of the approaches that we are now considering ourselves is just get your basic workflow in order and then implement pieces of AI which help the, the, the user or the agent to, um, to, to to bring a faster solution. Um, that will, uh, do two things. Uh, first you have your, um, your business involved. Um, they will stay, uh, accountable, um, take the responsibility that will also help, uh, to to to get their, um, get their help, basically. Um, and so, yeah, I think, I think that the AI should sort of be be discreet.

It will do a part of the job, it will offer a solution. And then the decision is being made. If you have your workflow there, you will also have your standard logging. So it is way less hard to after the fact, explain what happened. We are using Sox compliance applications. So we are doing all these audits, um, which can have very heavy implications. So that that is at least how I envision for the short term how we will work this. And, well, maybe at a later time things will change. Yes.

I don't know where that will go. I guess nobody knows. But for now, that would be my approach. And, um, maybe one thing to add there is when you start with process AI that is just a little bit outside of the payments domain. Yes. It's not in the processing part. So you have a bit more freedom. Um, code is maybe a bit more like um, or sort of risky for, for outsiders, but that is again, uh, giving already a lot of, uh, gains, uh, without directly touching the payments, processing itself. So, so, so it can be discreet, but it clearly needs to be supervised and not, not not allowed to to go over itself.

I'm going to come on now to to the second poll question, which which really, really looks at about automation. And the question is this what percentage of payment exceptions are exceptions are automated at your institution. And this can be automation either by sort of what we might call traditional workflow okay AI or any combination of those tools and or others. And we'll give a few moments to respond. And then we will we will come back to to that. And I want to switch now though to, to the lessons because people are pioneering in this space. Right. So what lessons can we learn from these pioneers? And Sean, you know, beyond the hype and we we said at this stage, this is not hype.

This is in the real world, what's actually working in practice. What can we learn from these pioneers? I think so where I see and I fully, you know, echo the comments from, from Tim and Kelly just in relation to the related controls and oversight. First of all, that's really, really key. But some, some opportunities that we have that are live uh, today, uh, a key example would be, you know, taking key supporting documentation and summarizing that for our service representatives so they can support a client query more timely. So what I mean by that is if a particular query comes in from a client into our service team, rather than the service representative having to go through, you know, certain manuals or process documents or, you know, uh, other materials to help them navigate the query or to get the answer, you know, utilization of the AI tools to basically summarize that where the where the representative put in the search what exactly they're looking for and that extraction to come. So that's one benefit that we see that that is alive today. That obviously contributes to a lot of quicker resolution times from the queries. And it also then opens the door to having service teams support more broadly.

So maybe even branches that they may not be as familiar with, with and having access to this materials and the opportunity to have things summarized, it broadens the sort of landscape of potential queries that our team members can can answer. So that will be where we see ourselves as sort of live today for those initial pieces. That's lovely. And and Kelly sort of staying with, learning from from pioneers. And you've got a lot of experience in this space. What is what is sometimes I want to say often what is sometimes underestimated at the start. What what what what can you share that people can avoid those pitfalls? We think about it as crawl, walk, run. So if you're at a stage where you're completely manual, no automation.

You want to think about crawling, then walking, then running. So as Tim stated, crawling is, you know, looking at a technology to standardize, create workflow around a process, that workflow can be deterministic, it can be intent led, it can be governed. Make sure that the experience is consistent in treating clients equally. Then once you have that in place, look at adding in various types of AI and then that AI can start to add more intelligence, more automation using some of the tools I mentioned earlier, like parsing technology or predictive analytics, uh, to mention process AI, which is looking at historical data to resolve issues. And then as you're starting to get to higher rates of automation, then thinking how you can add in gen AI to a specific process. So a lot of our clients have gotten to 80, 70, 80, 95% automation just with that intent led workflow and some in the parsing and NLP etc.. So those are great results. Then obviously Agentic AI and data mining etc. can improve those results even higher to get to 95 100% automation.

Um, I saw there was a question in the chat about data privacy and can an agent, can you use an agent and how would you track it to resolve a payment autonomously? And our belief is, yes, and we are developing a lot of different types of payment agents that can say, look at a payment exception, scan it for improper um identifiers, scan a swift message, figure out the question and the problem if it's a complicated 199 or 110, etc. and then look directly at the data in that we call it a case or that exception to identify the problem, surface the problem, and then autonomously resolve it. So but if when that is done in a workflow, the entire the actions of the agent are tracked and recorded. So it will stand up to regulatory scrutiny. And the agent is working within the confines of intent based workflow and being told specific places to get the data, um, looking at staying compliant with Swift message types, etc., etc.. So it's it's governed, guided, predictable agents. They're working autonomously, but within the confines of a specific exception type, not, you know, just deciding what we're going to do today with that. So that's perfect.

Thank you for responding to those questions. That's that's great. Thank you. Um, we also we also talked about resistance as well. So I'm still on lessons learned. There's yeah some of this affects operational teams. Do you sometimes feel did you see resistance from operational teams or are they fully embracive. It's a mix. We definitely see resistance.

Um, there's always resistance to change. It's human nature. So, um, we definitely see resistance. But then when people see that, uh, AI can write an email for them and they don't have to go search for that incorrect account number, or search through the whole case for all the information and write the email. They now have to become a reviewer of an email that's been written for them, or there's guided steps in what emails should be sent or what swift message should be created. Then people embrace the change because they realize they're using their intelligence to review the manual work that's being put together and compiled for them. So there's still a still human in the loop in many cases. And but the mundane work of putting it together is taken out. So then they see the benefits.

And I think it's a bit like our personal lives as well, isn't it? You know, if you pop something into your eye and you think for the first time especially, you think, my goodness me, that that saved me a good however long 20 minutes, an hour or whatever it is. I think, Tim, I do I do want your thoughts on this, but I think we should just have a quick look before we do that at the at the results from the poll question. So and this was all about the automation of the automation, the percentage automation of the workflow. So we have um, I where are we. I have got um for 0%. So nobody's automated workflows, 18% between 1 and 25% of automation of workflow, 27%. So nearly a third, uh, 25 to 50%, 18 zero for half to three quarters of a percent of the workflow is automated 75% or more 18 and unsure 18%. So so quite, quite a spread of results there, actually.

Um, Sean, what are your thoughts on those results? Yeah, I think it's very, very interesting. And I think obviously, you know, within the workflow you're going to have potential elements that are going to be automated. But, you know, if we take the example, what Kelly talked about there, where, you know, a draft response to a client email can be completed, but you're going to have a checker, you know, from a service representative, you know, giving the okay before that maybe goes out the door. So I think, you know, within the actual end to end workflow, as we see in the results here, that there is a good number of processes and elements of processes that, um, that can be automated, uh, with the workflow. I certainly think maybe there are certain elements when I think we'll touch on it in a while. Um, trees, when it comes to actual payment adjustments and movement of funds, is maybe a little bit different. But now I think that the results are very promising as well. And before you tell us about your thoughts on kind of the lessons from pioneers, um, your your thoughts on the results.

Yeah. Yeah, I, I sort of recognize it. Um, I think in Rabobank we have quite a lot of flows automated, but within the flows there is still a lot of manual steps. It really also depends on the product for the the amount of exceptions is really low. Um, and most of that is fairly straightforward. So a lot of these exceptions are already automatically, uh, fixed. Um but yeah, I recognize that there is a lot of work to be done. Every time when we look at something or a connection, there is so much that we can improve. And, um, sometimes I, I even well, I'm not saying ashamed, but I pity that when I talk to my users that I have to pick because there are so many things that, that I can do for them, but there is only also so much that you can do in a year, basically.

So yeah. Absolutely. Yeah. And um, Kelly, your what do the results strike you? It strikes me that we have a good diverse audience in the group. Um, yeah. Yeah. Yeah. Not completely surprised.

I'm probably most surprised no one picked 50 to 75%, because I do know a fair amount of banks that are at that level. But, um, as Tim said, there's there's work to be done. And you also sometimes need to define, um, define the question maybe a little more closely as what what is considered automated? Is it all the way through? Is it certain steps? ET cetera. ET cetera. Good. Good call.

And Tim, thank you for your patience. So, so when it comes to comes to the pioneers and sort of lessons learned, um, what would you share with us there? Because some things, you know, they, they fly, but some things they stall, sort of, you know what, what do you think are the ingredients or some of the ingredients for success? I guess two things. It's one, uh, a bit, um, in the length of what? Uh, Kelly stated, like, you need to fix your stuff first, so don't throw AI, uh, against everything. And two is there will be a lot of initiatives within your company with AI. So we are approached by other teams that have an AI platform and want to automate, uh, stuff. And often that is automating stuff that is not really needing 88.

It's just not, uh, nicely automated yet. and before you know it, you get into a situation and I. I don't know if this is a general thing for other companies too, but, um, we had if there is some issues or capacity to make or automate things properly, um, people tend to put something else in it, which can be quick. So a quick fix with robotics or a quick fix with an AI platform. Um, and I think that is really important that you, you go for the long term and sustainable solution. Uh, instead of putting layer over layer over layer, which can cause in the end, uh, an unsustainable, um, platform or environment. Yeah, I get that. And you make a good point there about, you know, there are there are tools, other tools other than AI available in this space to prevent and reduce exceptions and to help the investigation and management, management of them. But when we think about AI, Tim Agentic AI can supercharge human behavior.

How though do do we do that without it overstepping limits? Yeah, that is the basically the name of the game at the moment. Yes. How do you keep control? Um, um, also, how do you keep your people on board? You still need the people having the right knowledge. Uh, you need, um, a traceability. Um, yeah. For us, the focus is really, um, keeping the responsibility at at the, at the employee.

Um, so they make the decisions and the, the AI is doing the the nasty work, the repetitive. Yeah. Doing an analysis. Yeah. 15 minutes going through all kinds of types of messages. That is just, um, not fun. Yes. Well, maybe some people do like it, but, uh. Um, yeah.

So most of us. But. Yeah. Do you ever envisage a time though, Tim. Biotin, where AI would authorize a payment and do the whole thing. Well, I guess yes, I expect yes. Um, but, um, I think it all depends on the risk profile. So the risk profile of the task, and if it could be a payment or it could be a task. Um, so low risk, low value.

Um, yeah, I can see that happening. Uh, I was even thinking about we have a lot of processes that have four I, six I approval, um, uh, routines. Um, so there could be a situation where maybe your AI would be the first two eyes. I don't know, that is really something. We are not there yet, but, um. Yeah. But it's something that we're thinking is we're not saying it's on the horizon, but we're saying it's something that that we're thinking about, so that's fine. And Sean, a similar question to you, can you envisage a time where AI will automatically authorize Payments. Even so, for just a certain client, can you see that world at all?

Well, as we evolve through 2026 from our own utilization of the Pega platform, my own objective is, you know, for example, some lengthy email, um, from clients that can be summarized, you know, that the application will propose the next best action, propose the response back to the customer, uh, the case classification, ensuring that the case is obviously tagged correctly, etc., from our own side and similar to Tim and all institutions. Right. You have very strict controls. There's a difference between, you know, a client query and a payment adjustment. Right. Because obviously a payment adjustment, um, you know, if there's a mistake made in relation to it, it can have big consequences for our customers. It can have big consequences for the organization. So certainly in the more shorter term might my focus is on in relation to, you know, the actual client queries, handling the client queries, etc. in, in relation to it.

So I think that's a that will be my take on things. Maybe next time. Teresa, when we sort of speak in relation to this, um, you know, for me the payment processing element side of things is really sensitive and something that we will need obviously, to be have with our with utmost care. Thank you. That's perfect. Thank you. And Kelly in a couple of minutes we have have remaining sort of a double a double question for you if I may. Um, so firstly where do you see. Because, you know, we talked about this is not the hype.

So so where do you because you know, you're you're as I said before, you're very experienced in this space. You're talking to a lot of banks. Where do you genuinely see AI delivering benefits and or. And what is stopping financial institutions? Maybe, you know, looking at some of those earlier poll results, you know, we're not really thinking about it or we're not doing it. And maybe that's caution, I don't know. So where's it where's it delivering results. And what's stopping people. What's stopping firms.

I will answer that in three minutes as best I can. We can spend an hour on that just alone. Um, so we definitely already seeing benefits where clients are using, um, internal tools for guiding employees. Um, a lot of clients start to do that. So you can direct an LLM at internal processes and procedures and or sites like, say, a Swift website. So if someone's asking, how do I resolve this? Then those directions are surfaced. It's still a human potentially doing the work, but they're getting clear guidance on what to do, which is when we do operational walkthroughs with our clients, we see a lot of times operators don't know what to do. So that's that's very helpful in a lot of firms are using it.

It's an internal internal tool. getting case summaries. Shawn mentioned reading emails and scraping data out of those emails and then creating a the process or the case of what needs to be done. So those are definitely use cases where the AI is doing the first part of the job, and then the human is taking over to mention same thing taking creating data and then having the human be the the final approver. So you can take one approver out of the chain or have the AI do the work. Um, there's definitely a ton of benefits. You know, compensation can be calculated by agents. And maybe that's a lower risk, one where compensation payments and requests could be calculated and sent out. Um, so there's a lot of different ones and a lot of benefits.

What's stopping firms is the trust factor, right? The AI has to be not just a prompt that could give you ten results for ten questions. It has to be put in that guided workflow that we've all mentioned that the AI has specific tasks, and then it has to be monitored and tested and QA to make sure it is doing those tasks. Um, I had a funny experience where Amtrak is now. Amtrak is the train system in the United States. They are now using agents to book trains for you when your train gets canceled. And I had them book three trains for me on the same day at overlapping times to go to the same place. So kudos to them that they're forward thinking and trying to get an agent involved. But I don't want to pay for three trains at the same time that I can't be on.

So, you know, it's a balance, right? And so these things need to be controlled and tested before put in production. And that's and that's perfect. Thank you. And we've come to the end of the session. We've kind of we've we've we've eaten into our hour and there's still lots more we could say. But I do need to to wrap things up. And I've been making notes as you've seen my pen, um, as we've been going along. And I think I have four key takeaways from this.

And the first one is better upfront design and standardization can materially reduce Downstream manual effort and and liquidity and liquidity of clients. Secondly, AI can deliver real measurable value. And Kelly, you just said about, you know, just just a moment ago about, you know, you can quantify that, which I think is really important for business cases, by the way, I think, you know, that that that really that really helps. And thirdly, human accountability remains essential. You all spoke very, very well to that. And for me, finally, it's about collaboration. Um, it's either at industry level when you talk about standards and data and, you know, stuff like that, but also internally as well, collaboration between human and and AI and artificial type agents and bringing in technology like LMS. So it's that. So thank you.

Thank you as well to everybody that responded to the poll results. And for those that asked the poll questions, I think we got to to all, if not most of them. So so thank you very much. But the biggest thank you to our panelists. Thank you so much, Kelly, Tim and Sean, for sharing. So it was beyond hype. And you've given us real life practical insights and examples, and I thank you for that. Um, if today's discussion has sparked further discussion, please do pick up with our panelists in the usual fashion, which is probably LinkedIn. Um, but thank you for joining us and enjoy the rest of your day.

Thank you Teresa. Thanks, everyone. Thanks, everyone. Thank you. Thank you. Great to be here.