21 Jul 2026 · 54m · Oliver Bussmann, Michael Krigsman · 19 quoted facts
Oliver Bussmann, former group CIO of UBS, argues that AI agents are technically ready for banking but adoption is gated by regulation, controls and, above all, customer trust. He sees large productivity gains, from 10–30% with copilots to one to three times with agents, but only if governance, testing, audit trails and human oversight are in place first.
The central tension is speed versus trust. Software companies adopt agents quickly, while regulated banks, especially in Europe, must rework risk frameworks and prove every output is traceable. Bussmann warns that one hallucination-driven incident could destroy a bank's reputation within minutes. He still expects agents across bank functions within a year, with peer pressure driving adoption and new work shifting toward process redesign.
Bussmann says the technology already works, but banks must first adjust risk methods, controls and processes, as they did for cloud and blockchain. That is why he expects many proofs of concept but few large production deployments.
Higher-risk use cases, such as credit decisions affecting consumers, get more cross-model validation, human verification and cross-checks. Bussmann supports the EU's risk-based approach and says he would be careful with high-risk uses.
Banks already track lines of code, code quality and standards compliance for human developers. Applying the same measures to agents shows whether the claimed two-to-four-times gains are real.
Bussmann is waiting for a hallucination-driven industry incident and cites a KPMG research report with wrong data as a warning sign. He notes that bank runs happen when customers lose trust.
Agents make code cheap, so the constraint becomes business analysts and product managers redesigning processes and writing good prompts. Junior engineering job ads are down about 40%, yet he argues banks must keep hiring juniors who will become future seniors.
Bussmann says roughly half of financial institutions are using agents in some way, but most of that is proof of concept. Real production use is slower because banks are regulated and must adapt their risk controls first.
Banks are regulated, so every new technology means adjusting risk methods, controls and processes so regulators are satisfied. Bussmann says software companies in unregulated businesses adopt agentic AI much faster.
It means an agent runs a process almost entirely on its own, and humans step in for exceptions and critical cases and monitor results against agreed KPIs. Bussmann gives accounts payable as an example: AI handles about 95% of invoices and people handle the rest.
Bussmann says job ads for junior software engineers are down about 40%, but companies still hire juniors with AI skills. He argues banks cannot stop hiring them, because they need them as senior supervisors in three to five years.
Bussmann recommends checking important results with several different AI models, then reading the information yourself before sharing it. Using more than one model reduces the risk of misinformation.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Oliver Bussmann estimates that roughly half of financial institutions are using AI agents in some environment, most of them as proofs of concept.
“Roughly 50% of financial institutions are already using uh agents in there in an environment. I think most of them are proof of concept.”
Bussmann describes European AI regulation as restrictive and based on classifying use cases by risk, with significant fines for companies.
“the European regulation is very restrictive, um, and very risk-based classification, um, uh, controls.”
Bussmann says the US takes the opposite approach to the EU, assuming companies will handle safety audits themselves, which lowers administrative overhead.
“we see the opposite trend in the US, for example, that, uh, uh, assume that, you know, uh, companies will take care about the safety audit”
Bussmann says audit firms PwC and Deloitte are already bringing their own AI agents into auditing and reporting.
“audit firms like, uh, PWC already, and and Deloitte, they're bringing their own agent now into auditing the end reporting also”
Michael Krigsman says Rivian uses AI agents in its financial processes, with explicit human review points.
“I know Rivian uses, uh, AI agents as part of their financial processes.”
Bussmann says KPMG published a research report on AI in financial services that contained hallucinated, wrong data, and financial institutions pushed back.
“KPMG a couple of weeks ago was on the pressure, they put a research report out there and about usage of AI in financial services”
Bussmann says JP Morgan announced it is building its own AI infrastructure with a chip provider other than Nvidia, which he sees as an exception among banks.
“It's only one bank, JP Morgan announced that they are also building their own infrastructure with, uh, chip provider, then not the typical Nvidia, et cetera.”
Bussmann says building own AI hardware requires scale such as JP Morgan's, with US banks of that size having IT budgets above $10 billion a year.
“only large bank with, you know, JP Morgan, US banks are usually, after, you know, among above 10 billion, uh, dollars of budget, IT budget per year”
Bussmann says job advertisements for junior software engineers have dropped by about 40% over the last couple of years.
“demand for junior software engineers, uh, uh, dropping, you know, the job advertisement are down by 40%.”
Bussmann says Salesforce announced it is still hiring graduates, even more than the previous year, particularly those with AI backgrounds.
“I think Salesforce announced that they are still hiring even more than last year.”
Bussmann says the legal industry now uses large language models customized by jurisdiction, reducing the work needed to draft and review contracts.
“Legal industry is now using, uh, large language models, totally customized for their jurisdiction.”
Bussmann says consulting firms' valuations are down 40% or more this year as AI changes their business model.
“You see, consulting firms also, their valuation is down this year, forty and more percent.”
Bussmann says agentic payments being developed by Visa and Mastercard are probably the first agent-to-agent use case, and they depend on verifying that the agent is legitimate.
“I think the the, uh, agentic payment topic that Visa and Mastercard working is is the maybe the first one.”
Bussmann says the European Commission is working on digital identities, which he sees as needed before digital twins or agents can transact at scale.
“I know that the European, uh, commission is working on certain governments on on those kind of digital identities”
Bussmann says shared know-your-client utilities have been discussed for ten to fifteen years and mostly failed for lack of critical mass of sponsors.
“most of these, um, utility setups failed because there was not enough critical mass, um, sponsoring this kind of development.”
Bussmann says blockchain can give a tamper-proof record of decisions, adding to auditability beyond ordinary logging.
“you can go to the next, uh, level of maturity, saying, you know, I'm using blockchain, that nobody can even correct, uh, um, uh, transaction”
Bussmann says the European Central Bank, his regulator as a supervisory board member, tells him to follow existing rules such as the European AI regulation and DORA for AI.
“I talk to the to the ECB as my regulator, uh, they're telling me make sure that, you know, for this kind of topic, the existing regulation, like the European AI, the Dora,”
Bussmann says bank board agendas have shifted in recent years toward operational risk, which is largely about technology and operations.
“the board agenda focus has shifted the last couple of years into much more operational risk.”
Bussmann says robo-advisors did not replace human advisors, because clients still prefer to talk to a person before executing financial decisions.
“People like to talk to a bank at the end, say, you know, is it a good decision?”
Bussmann sizes up adoption and explains the shift from copilots that assist people to agents that run tasks on their own.
“Roughly 50% of financial institutions are already using uh agents in there”
He explains that banks must adapt controls and risk frameworks for each new technology, so deployment takes time even though the technology works.
“Always takes time to make sure everything is bulletproof.”
He explains how banks can prove agent gains by measuring developer productivity and code quality before and after.
“baselining the productivity is is absolute necessary.”
He covers what must be in place before agents write to core ledgers, and how audit and control functions will use agents themselves.
“I think there will be also, um, audit agents”
He argues that a single hallucination incident could wipe out reputation, and ties this to bank runs and the KPMG report.
“a bank has to take care is trust.”
He says junior hiring is down but still necessary, and that engineering work moves toward customer projects and domain expertise.
“you can't stop hiring juniors because at the end”
He argues that AI-equipped clients expect faster, cheaper work, which is already changing pricing for consultants and lawyers.
“the customers are not willing to pay anymore $5,000 per per day.”
He says machine-to-machine banking depends on verifiable digital identity, and financial crime work could improve by linking silos.
“without that verification, it will be hard to to do mass volume of transaction”
He agrees that organizational change and process redesign, not software coding, is the real constraint.
“if you had set up your right agent, coding is not the problem.”
He predicts agents in production across bank functions within a year, driven by competitive pressure and larger productivity gains.
“Here we're talking about, uh, one, two, three x in certain use cases.”
I'm waiting for an incident in the industry hallucination is or some other issues then your whole trust and reputation can be gone in a few minutes.
Banks run on trust, but AI agents are not there yet. Oliver Bussmann is the former group CIO of UBS and advises banks across Europe.
Roughly 50% of financial institutions are already using uh agents in there in an environment. I think most of them are proof of concept. Um, um, I would say, uh, most of the banks are exploring, uh, agent in the software development. And, uh, so, I think we are seeing a big change from, I would say, copilot work with, uh, uh, using, um, uh, agentic, um, uh, AI, agentic AI to, uh, autopilot solution with, uh, using agent in in different function from from back office to IT, to marketing, et cetera.
So, it's a it's a big change, as especially, I would say, the last the last couple of months.
So, would you say that banks are actually embracing agents? Or is it still in practice, or is it still kind of theoretical? You know, it seems like a good idea.
It is taking time because, uh, you know, I'm also working in the non-regulated business with software companies as a as board chairs, et cetera, and the adaption of agentic AI is much faster because the financial service industry is a regulated industry. And then every time if you implement, uh, a new technology, like, you know, a couple of years, uh, software as a service, cloud or blockchain, uh, you have to, uh, adjust your, um, uh, risk method, your control environment, your processes.
That means the regulator is expecting, um, you that you, um, make sure that these new technologies are, uh, used in, uh, you know, following the laws, regulation and policies, and so, that so, that means we will see a lot of testing, proof of concept, uh, but, a huge production cases. Always takes time to make sure everything is bulletproof.
Are the obstacles technology or compliance and regulatory issues or talent? Uh-huh.
The technology is there. I can I can, you know, working with software, uh, companies that, yeah, yeah, are going through the transformation, the change to last, um, I would say five months after, you know, the stock market, um, uh, uh, you know, lowering down the the valuation of software businesses. There is a really huge change on these businesses to embrace, um, uh, agentic AI, meaning vibe coding or the next generation of software development, and we see a huge change in in that non-regulated business from copilot in, uh, doing software development, uh, with twenty forty percent productivity gain to factor to four x more output.
So, I I would say, uh, regulated institutions see this, you know, huge advantages, you know, there there seeing this from a from a software development perspective, but that requires that, you know, procedures, controls, risk frameworks, policy has to be, uh, adjusted. And, um, especially if you work out of Europe, European regulation and laws are much more restricted than the the US one, so that puts even more effort, um, uh, for the financial institutions. So, I would say everybody is working on that.
Um, and for me, the best example is that, uh, using agent is is also from a cyber cybersecurity perspective. Now, with my thoughts, um, uh, being available, you know, you use, uh, those kind of agent to scan your entire software libraries and and landscapes in in, uh, a couple of days and and try to understand any potential issues. So, so, I would say very selective, um, I think in certain non-customer-facing areas even more, uh, but the adoption has to follow first that these kind of control and risk frameworks are in place to make sure it is safe and in compliance with regulation.
In other words, the technology works, the technology is in place, people have it available, but there needs to be an alignment with the existing compliance and regulatory environment, and people are trying to figure out how to make the technology work within that environment.
absolutely, absolutely, that that more. I think you have to distinguish two things. One is making sure that, you know, in Europe, there's the European AI that you, uh, uh, have an inventory of all your AI cases, including agent cases, that depending on the risk classification of the use cases, you have to put more human controls and testing in place, uh, and so that has to be bulletproof. The second topic is also, but I see from a non-regulated, uh, software environment, you have to still put some guardrails, uh, making sure that these agents are acting in a, you know, uh, in the same consistent, traceable, auditable way, you know, that, um, if you work with the agent that the the output
is is in the same, uh, design principles, um, um, coding principles, et cetera. So, there's a lot of effort also making sure that, you know, um, these kind of agent produce, uh, the expected outcome and, um, so that requires also a lot of governance controls oversight.
Folks, you can ask your questions. If you are watching on LinkedIn, just pop your questions into the chat. If you're watching on Twitter X, use the hashtag CXOTalk. Take advantage of this opportunity to ask Oliver Bussmann pretty much whatever you want. We have a a simple and interesting question on LinkedIn from, uh, Vibhav Mandra Wadkar, who says, How are the guardrails being configured?
If I look from a typical software company perspective, I think the the engineering team and head of engineering is is, you know, the job description is changing from, you know, actually doing the software coding, et cetera, into defining how these agent are being defined and set up and, um, also that is sufficient testing and oversight in place. So, the the role of the engineering team is changing rapidly, and their major role is making sure there's high quality, uh, according to these expectations from a compliance security point of view, and, um, making sure that this almost like the production factory is producing, um, good quality outcome as expected.
He follows up with a question, another really great one, he says. What are the cores core KPIs that have helped? He asks about UBS, but just banks in general. What are the core KPIs that help banks after implementing agentic AI?
You know, if I look deeper into, um, uh, the software development, because that's something that I would say from an agentic, uh, AI perspective is the most advanced one is is is, and that was a topic also during my time at UBS, um, how to measure, um, developer, uh, productivity, uh, efficiency, code quality. Um, I have to say there are now tools over the last ten, thirteen years in place that measure this kind of, um, productivity, quality and security of the school, and that's now been adjusted also for the agentic AI environment.
So, whatever you do, um, and it's pretty standard already in financial service, baselining the productivity is is absolute necessary. Um, and then you, you, then you can bulletproof is the way of automation that you have in place. Um, is that is a really factor two or three or four that everybody is is talking about, uh, is a real, because that's a proof point that I would say is whatever you can do measure in that tools now in place, do it, and it gives you much more confidence that this this is really a game changer.
So, have your baseline set up in advance so that when you implement agentic AI, you can actually have a a point of comparison.
yeah, absolutely, you need that.
Give us some examples of some of these baseline measurements.
The baseline is how many, um, um, um, lines of code you you develop, um, their quality KPIs, um, how the the code is being structured, how is it being, um, in line with the, uh, coding standards, et cetera. So, so these kind of KPIs, um, they're very sophisticated, and, uh, these these tools that are m measuring that, they're they're they're, they're working on that kind of KPIs over the last ten years. So, it's, I would say, it's pretty pretty standard already, and, uh, uh, that's the same now.
You have to apply this not only for a human, uh, development, you have to measure also how an agent, a a machine, is now producing this kind of, um, uh, quality of code.
And this is from Arcelon Khan, who's a long time listener, and, uh, Oliver, sounds like you remember him from your last appearances on CXOTalk.
Yeah, it's ten years already, you know, it's, uh, it's a fantastic journey of, I would say, you know, thank you so much for having me, by the way.
Uh, well, thank you for thank you for coming back. And Arcelon Khan says, should we deregulate banking for the AI area, a AI era? So, maybe you can talk about this issue of regulation and the restrictions and what can be relaxed or not relaxed.
If you're global and like UBS, uh, you know, it's a global, um, operating wall that means you try to to utilize as much as possible the same application, uh, the same infrastructure, et cetera. Um, regulation drives fragmentation. Uh, so you you have to demonstrate that in each jurisdiction that you follow the policies, regulation laws, um, best practices, et cetera, that that makes, um, production and work, uh, complex. And, uh, what we see right now, if I look from an AI perspective, the European regulation is very restrictive, um, and very risk-based classification, um, uh, controls.
By the way, it's not only for, uh, financial institution, also normal corporations to follow, uh, significant fines, and then we see the opposite trend in the US, for example, that, uh, uh, assume that, you know, uh, companies will take care about the safety audit, et cetera, and, um, so that's less restricted, and, uh, the administrative overhead to do that is significant lower. So, that's something if you're running a global institution, um, that that is is is, uh, is something that you have to, uh, take into account.
And this is on LinkedIn from Rafael Simonowitz, and he says, Oliver, when giving AI agents write access to core ledgers like SAPs for HANA, how should boards combat automation bias to keep humans truly in command?
If you write into ledgers, I think that that has to be, um, uh, tested, has to be traceability, audit logs, et cetera, to make sure if there's if there's an incident, um, that, uh, that you can roll back and figure out what happened. Uh, and I think there will be also, um, audit agents, so the the audit function going forward is not only that, you know, you bring in topic by topic and audit team and go deeper and see if there's everything, uh, computed in the right way.
I think the future is also that your oversight in, you know, in banking, called second, uh, um, line of defense or third line of defense, you know, will be part of the production, uh, so that's, that's it. And in audit firms like, uh, PWC already, and and Deloitte, they're bringing their own agent now into auditing the end reporting also, so that you see there's a change, not only that, you know, you bring, uh, an agent into produce, also, then the controls will be also more automated, and so that you can act immediately if there's something, uh, is if there's a major deviation point of view.
But to what extent are banks able to be hands off on their agents and trust that agentic audit report as opposed to require a manual or or human review? Because we all know the the the problem with generative AI and agents is they can make stuff up.
That's the point that I mentioned before, is the engineering team has to put guardrails into these agents, uh, making sure that the design standards are being followed, uh, certain coding standards, uh, certain testing, uh, each time if you develop a new function that certain test cases are built in going through. So I would say the engineering function is changing to really got wheels, quality assurance, et cetera. Then you have again also the other function, risk and audit, constantly monitoring the environment, you know, and and and there will be always a supervisor, human supervisor in the loop.
So the discussion going forward will be how much you you know how many agent you you you manage as a department head in the future, um, in in respect to how many people you're managing, so, so, the the the discussion is already changing, uh, from from that perspective, that still, as a manager, I need supervision, not only about my own people, also for the agent in my working environment.
I know Rivian uses, uh, AI agents as part of their financial processes. They have very specific points of review, so the human in the loop is very explicitly and very carefully built into the process.
you have to define exception or, uh, topics, um, uh, workflows that, you know, run into a problem and cannot be fixed through AI. That that usually triggers, you know, from a a cons payable perspective, um, you can, you know, uh, uh, scan invoices and and map them, uh, you know, that's usually ninety ninety-five percent, and then AI try to bring it to 100%, but there will still a percentage or two out there that you need human oversight and control perspective, and I think everybody is now trying to calibrate that.
Um, and, um, and so that's something that I would say in a non-regulated industry is easier in in a regulated industry that the regulator usually says, you know, there's a black and white, no, so you there are rules. And and with the agentic AI, you know, there will be a situation that certain rules, um, because the environment has changed, will apply differently, and that's something the the at least the European uh, uh, regulators are a little bit afraid of that, is to move away from a very rules-based workflows to, um, bring other parameters into the workflow that drives a a different outcome.
Well, the idea of handing essentially regulation over to the firms themselves, uh, is potentially problematic for obvious reasons. We have, uh, another interesting question, and I encourage you guys ask your questions. If you're watching on LinkedIn, pop your questions into the chat. If you're watching on Twitter or X, use the hashtag CXOTalk and and at M. Krigsman, me directly as well. And this is from Chris Peterson, and he says, in the fin in the financial realm, is agentic AI built on LLMs and GPTs, or embracing other forms of AI to get the transparency and explainability that auditors demand?
I would say it's not only limited to to, um, large language models, agentic, um, um, um, or or to. I would say machine learning is still a big portion of a lot of, um, functionality in, uh, in AML, anti-money laundering or some analytic capabilities. I would say I would not exclude that, so, um, even I would say the combination with blockchain, everybody knows that I was pretty, I'm still pretty active in the blockchain, uh, DLT community, is a pretty good way because then, um, uh, you have a bulletproof lock of your decisions, not that you can't manipulate, and so I would not exclude any other technology, uh, being part of a much higher level of automation and, uh, driving better
outcomes. So definitely, I would not exclude that.
We have another interesting and short question, uh, great question from Vibob Bob Vibov, uh, Mondre Wadker, and he says, Is there any specific scenario or use case where agentic AI is not suited to a regulated industry?
I would say the regulator, if you ask me, you know, in front, sitting in front of the regulator, um, everything that could be biased into a credit decision or, uh, security oversight, that means, uh, monitoring, uh, people on the street, that's something the regulators, let's say, from a European perspective, I said, classify this as a high-risk topic, um, is something that that, um, I would be careful, because then you you have to put a lot of effort to make sure that there is not misuse of this kind of information or decisions, et cetera.
So, so, I would say, I I like the classification, the way how y- European governments are looking at these cases, and if they're, um, high-risk cases, and these are the the most important one, credit decision on on consumer, et cetera, that, you know, goes beyond, uh, uh, information that is usually available and drives decision, that's something I would be careful, yeah.
It sounds from what you're saying that the regulators are very concerned about the judgment decisions made by a AI as opposed to the AI making mechanical mistakes.
yeah, yeah, absolutely, because at the end they're also very keen, you know, from a European perspective, kind of consumer protection, client protection, et cetera, making sure that these kind of technologies are not, uh, driving a disadvantage for, um, consumer from that perspective. So, that's something I would also that angle is maybe not so important in other jurisdiction.
I'm halting on this particular point because most technical technology discussions around agentic AI focus on the mechanical aspects, hallucinations, and getting things making things up. Whereas in this instance, you're looking at the higher level set of judgment issues, yeah, yeah, it's the real intelligence.
suppose, I think I think I think the the I think the i independent, how risky, um, um, uh, how risky, um, um, AI use case is, um, the biggest biggest value that, you know, a bank has to take care is trust. So today, if you act as a financial institution, whatever you provide produce has to be, you know, according to the highest level of trust. No, because I, as a bank customer, have to rely on on the bank that they're processing all my re my business in the right way, and it's it's, it's accurate, it's it's traceable, it's um, audible, et cetera, and and if, you know, I'm waiting for an incident in the industry that that hallucination is or
some other issues that that drives to, you know, an incident, um, then then your whole trust and reputation can be gone in a few minutes. No, you saw you saw that KPMG a couple of weeks ago was on the pressure, they put a research report out there and about usage of AI in financial services, and there there was a lot of hallucination in there, uh, uh, wrong data, and, you know, financial institution had to push back on that.
So that has an impact on trust on, uh, can I can I confidence, can I use this kind of services now? And so that's something I would say, you know, if you look at big crisis in the banking industry, bank run is based on people. Customers are losing trust, and so, so, that's the worst case scenario that you, as a as a bank, has to be super careful that whatever you do embracing emerging techniques like AI, making sure there's still that the trusted relationship between you and your customers are still intact and will be not not jeopardize that just to maximize certain certain, uh, uh, profits, for example.
We have another question now from Arcelon Khan, who's on Twitter, who says, Are banks creating their own AI infrastructure, buying chips and so forth, or are they just using Google and others for AI services?
It's only one bank, JP Morgan announced that they are also building their own infrastructure with, uh, chip provider, then not the typical Nvidia, et cetera. Um, but that's pretty, I would say it's more an ex, uh, ex, uh, exemption, I think, because you need significant investments, um, people, uh, going into this hardware topic, uh, and, uh, only large bank with, you know, JP Morgan, US banks are usually, after, you know, among above 10 billion, uh, dollars of budget, IT budget per year, so there there is scale, et cetera, but that's not the normal spend level for, um, uh, uh, uh, normal, uh, national banks, uh, so, so I would say that's more an exception, and, uh, um, more tech players,
uh, with the volume and scale can can do that.
This is from Alexander Yin on LinkedIn, who says, What is your confidence level that agents in a in the same LLM, say, Claude, for example, or ChatGPT, would be independent enough to quality control their own work by another agent in the same model, or do you need to use agents from different companies, or does it make no difference?
If I apply, you know, my experiences, uh, uh, in my day-to-day, uh, work life, um, using different large language models, cross-model, ve- uh, validation for super important, uh, task is mandatory because it gets you, it gets you a much higher confidence that whatever you produce is is correct and validated. You know, the for, uh, again, important outputs is always, uh, human checks and reviews is, um, like, like, the important research, uh, report that we talk about, KPMG is mandatory.
In production, yes, ab- absolutely. If you automate more, um, you you use different models or, um, teams that operates, uh, agent in a totally different way, not like, like, the auditors or audit firms are doing.
And this is from Mari John N, who says, if you need to make a prediction, how will entry-level roles change as agentic AI advances? So, the impact of agentic AI on entry-level jobs.
I'm tech advisory board member of a non-profit profiting organization that helping over 100,000 students to, um, to get the right jobs, et cetera, and we see in the last couple of years that demand for junior software engineers, uh, uh, dropping, you know, the job advertisement are down by 40%. What we are seeing still is that tech companies, um, hiring, uh, juniors, uh, graduates in, uh, you know, like, I think Salesforce announced that they are still hiring even more than last year.
The difference, what I see, is they're looking, uh, for, uh, graduates with I-AI background, and they're using this for their new AI project to automate certain business function. I see, uh, in high-tech firms, and also now in banking, this kind of centralized FDE, uh, forward deployed engineer model that goes into business department, understand the the breakdowns, the process improvements, and use these kind of, um, agentic AI models to to automate this as much as possible, and so they need the juniors to, um, to work on that and then build expertise, yeah, within the first three to five years, to become a supervisor of this kind of topic.
So, so you can't stop hiring juniors because at the end, uh, someday we need them as a as a, uh, as a very senior person. On the other side, the the work, project, work of financial institution and high-tech firms in transforming business requires, um, uh, employees with really deep understanding about AI models and the way how AI is working, so that's something I would say is an opportunity. On the other side, I have to say there is fear out there.
You know, if you talk to the younger generation, to my family members, et cetera, they're saying, you know, I'm studying and what will be my futures, uh, uh, joby? My message is don't be afraid, be a first mover, spend time with these models, apply this to the maximum, and new expertise will be in high demand. No, the same also for the, uh, older generation, uh, so, um, people, you know, um, at the end of their career, or, uh, they, they are, they're saying, oh, I don't have to learn that.
It's almost like I'm my, I'm, I'm comparing this, like, ten years ago, saying, I'm not using the iPhone. This is something that maybe is not so so important, you know? So, that's something I would say is learning is a lifelong, um, experience and must do, and, uh, um, so don't be afraid that that gives open up to more opportunity, but definitely job description are changing and changing mean now we have different kinds of jobs. If I look at the software companies that I'm I'm supporting, we have a discussion now, what's the role of the CTO, head of engineering, head of product management and sales, because work is being shifted from the engine room into more the the pre-sales organization.
So, so i it's natural that, uh, through technology, you know, and we saw this, you know, Mike, we in the business for so many years, it's changing, and, you know, and requires an open mindset and the willingness to learn.
If you're a mid-level programmer, software developer inside a large organization, isn't it undeniably true that your job is going away? Because, uh, AI tools can do software programming, so development fabulously well.
First of all, there is a need for really good, uh, engineers to making sure that these agent are being defined in a professional way and maintained and expanded. That that is definitely way. Second topic is is what I see is we are moving into, you know, you know, we're being through this, uh, from a very standardized SaaS model into more, I would say, AI customize a custom SaaS model. So your engineer, uh, capability will be required in more on-site customer projects to understand the problem and define pro prompts and and input for then your agent product factory.
So, the demand for, uh, software engineer then on site within your business environment, uh, will will be higher. So, so, message is, I think, um, if I look at studies saying the, uh, amount of certain work that's very traditional will come down, will be compressed, but there's more demand for digitalization and customization. Now that we couldn't do because engineering was always the bottleneck, the coding bottleneck, that now is going more to customer-facing, understanding the problem and and developing, uh, new functionality.
Instead of months, we're talking about days now. That's the game changer that everybody's now trying to figure out how to do that, and and this kind of change will lead to more business, and engineers will then be more project driven than working on core software functionality.
Well, there's no doubt that the key here is having an understanding of the technology so that you can manage that technology, and, most importantly, having a a deep understanding of the business and the workflows and what your customers actually care about, so that you can create technologies and guide agents to actually solve the most important problems.
at the end, we will go much deeper in the vertical topic, you know, that you need not not only the software, and also domain expertise to to understand problems and fix them, and and again, this will open up more opportunity. So, I'm, I'm personally, I'm not afraid of this kind of change, because it leads to more, um, chances to more businesses, and, um, and, like, you know, who have smooth my cheeks, now the cheeks is moving right now, you have to figure out where is this, you know?
So, that's the that's the best, uh, way to describe that.
I'm not quite as optimistic as you because of everything you're saying, of course, is accurate, but I also understand, as as you do very well, the difficulty that we humans have changing our roles and learning new skills.
Based on my journey, my learning curve over the last thirty thirty, I'm in the business over thirty-five years, to to be not afraid to take something new on. So, my risk profile is maybe different because I lived in different countries, different companies, I've been through a lot of new innovation cycles, and my learning is if you're the front runner, uh, in in testing this and utilize that, it's eye-opening, it's fun, it's motivating, and that opens new doors, you know?
So, if I if I share with you, the the way I I'm also using the technology in my own business, it's eye-opening. So, the productivity gains and work and the quality of work that I'm now, uh, yeah, working with my, uh, companies, with my firms, et cetera, if I compare this with a year ago, has changed so much that, uh, now, if I look at office productivity, we are, we are used to build PowerPoint and reports, et cetera, you know, it's sometimes you question the value of that.
That's now possible in a couple of minutes or hours than days, you know? And if I look at the amount of quality work that I'm now able to produce, so, so, that's the that's an experience that I would say helps me to, um, increase my confidence that if if you do this in the right way, with the right gradation there, in the way how you now use now, what, what, what technology is best for your, your different use cases, it's, it's, it's really, it's, it's fun, you know?
And, uh, so that's something that, yeah, I I wanna encourage everybody to to to, uh, don't miss that opportunity.
So, this is from Mohamed Da-Sougi, and he says, What is the impact of autonomous agents with human in the loop? So, maybe just describe that interaction briefly.
Yeah, the interaction in this autonomous, um, is is almost like you run a business process almost 100%, um, uh, through those agent working, uh, uh, as defined, and, uh, uh, human in the loop is, you know, if there's exceptions, critical cases that need still intervention, and, um, and then also, um, monitoring the overall outcome, is it in line with defined KPIs, et cetera? So, that's, again, the the split is between human and agent, and the numbers of agent per human.
My experience is depending on the criticality of business processes, regulatory requirements, et cetera, but that's something that, you know, will change over day-to-day work.
This is from Yannick Ishimwe, and he says, What is the real threshold that separates an agent that a bank will trust to advise, uh, from one it will trust to execute, and who should own that decision?
It really depends on how these use cases being classified as a risk, risk that, uh, could lead to wrong outcome, impacting customers, impacting businesses, and that drives this classification, right? The level of cross-model validation, human verification, cross-checks, et cetera. So, uh, so that's the that's the logic that I see in my in my in my board discussion is, you know, is the governance in place from a inventory from a risk classification, and is up the appropriate, depending on the risk qualification, controls, cross models, uh, human oversight measurement in place to make sure that there is no no deviation.
This is from Reza Sattari, who says, What would it take for a regulated bank to accept transformation deliverables, target architectures, fit to standard decisions produced by AI agents rather than an army of consultants? Is that a model risk question, an audit trail question, or a culture question? And I'll just add the commentary that Reza has just brought the entire, has just brought the entire consulting industry under a microscope with this.
It's an ongoing discussion, is, you know, as financial institutions are going through their learning curve of how these h agentic AI tools are being used, and then the productivity boost and the quality, um, improvement, um, uh, if you have the right again, if you have the right graduates in way, um, so I think the the the line between external and internal support is shifting. That that's the reason why, you know, software companies and consulting firms like an Accenture, et cetera, they have to change their business model.
So, business model means they have to ramp up also the way how the AI tools being utilized, um, uh, uh, being implemented, also the training of those, uh, uh, consultants, et cetera, because the expectation level now, if you've been through that learning curve on the customer side, the expectation level is much higher, and you do, um, uh, provide services much faster for, uh, reduced cost. You see this already in the legal industry. Legal industry is now using, uh, large language models, totally customized for their jurisdiction.
So, the amount of work to draft new contracts, review them is coming down, and so the customers are not willing to pay anymore $5,000 per per day. So, and that changed the whole pricing, uh, valuation or outcome discussion. So, I would say absolutely right. It really depends on how quickly you can ramp up your expertise and tools, you utilize enough tools inside internally, and then you see if somebody else outside is still up to that expectation level, which is much higher than to uh, currently.
And, of course, the major consulting companies are and have invested billions of dollars in AI training of and tool development because they know their business is shifting rapidly.
You see, consulting firms also, their valuation is down this year, forty and more percent. So, I think there there is a cross industry change going on, and and, uh, and that means there will be winners, there will be losers, and and the the the companies that embracing this, um, have a good chance to survive.
We have another really interesting question from Doctor Alexander Bachmann, who is the Group Chief Technology Officer of Helvetia Balduz Group, and has been a guest on CXOTalk.
Yeah, and we work together, so at Allianz, so yeah, great to hear.
Oh, I didn't realize that. Oh, okay, so oh, so it's a a small world, and Alexander says, Oliver, switching to the customer side of the topic. Do you think, or by when do you think, will customers use their digital twin or agentic AI assistant to execute their banking tasks? And by when will the business model include such things as, uh, machine-to-machine banking?
If I look at the agent to agent, uh, use cases, I think the the, uh, agentic payment topic that Visa and Mastercard working is is the maybe the first one. And here we have exactly the same topic, too, making sure that the agent is really Axal in a Brockleman and not somebody else, and and, um, and making sure that this, uh, transaction that been initiated are legitimate and not been faked. Um, and that's something, uh, everybody is working on that.
Um, do I think we are right now ready for prime time? I'm not sure, because at the end it requires also that there is a digital identity, um, for Alexander, that everybody can verify, and and I know that the European, uh, commission is working on certain governments on on those kind of digital identities, and but that has to be in place that, uh, a digital twin, i is verified, that's you and not somebody else, and, uh, everybody is working on it.
I hope hopefully this will accelerate because, uh, uh, without that verification, it will be hard to to do mass volume of transaction, uh, for agent to agent transaction.
This is from George Mertens, who is a CFO, and he says, Oliver, from your perspective, which AI-driven business models are emerging in financial crime compliance? For example, shared KYC slash AML utilities or compliance as a service. Could these models turn compliance from a cost center into a revenue-generating capability?
The compliance models, utility models, not to do KYC know your clients as a utility, um, it's a it's a known issue in industry for the last ten, fifteen years, and, uh, most of these, um, utility setups failed because there was not enough critical mass, um, sponsoring this kind of development. If we then move on into more decentralized work that, you know, an agent could, you know, reach out to another bank, uh, with the right identity to validate this, maybe that's that's the chance to do that, because, uh, these kind of centralized utility, uh, utili- utility models is hard to establish.
So, so, overall, I would say if you look at the, um, uh, value chain of financial crime, I think agentic AI will help us, uh, because, uh, these financial crime categories are in silos. Now, you have a client onboarding, client refresh, um, transaction monitoring, sanction, et cetera. These are usually specialized application and data silos, and agent, you know, with AI, you can connect them much better, uh, to a three-sixty risk counterpart view. And that's exactly, I think the industry is going is if there's innovation, one trigger that, uh, maybe will th- uh, drive a refresh of a client data point of view.
So, I I would say the technology is really, um, uh, at at game changer for the financial crime, fraud topic, and, uh, we see already companies working on this kind of, uh, embedding this kind of techniques.
And on the subject of technology, Alexander Bachman comes back, and he says, a counterargument is coding is not the bottleneck, it is the organizational change required for agentic process redesign and adoption.
Absolutely, spot on, because, again, um, uh, if you had set up your right agent, coding is not the problem. Is then you move more to the business analysts and business people in the business department that, uh, you know, has to, you know, we have to train and and bring up to speed how to utilize, uh, um, to analyze processes and breakdowns and able to, um, formulate prompts that, you know, leads to software development at the end.
So, I would say, as I said, from a software development perspective, we are shifting the focus into understanding the business, business functionality, business analysts, and then also product manager bringing this kind of, um, uh, uh, requirements into the product. So, that that shift, um, uh, means we will have much more focus on on people, front office, business analysts, product manager dealing with the issue instead of, you know, that the engineering team is becoming the bottleneck, uh, for these kind of changes, and that that again, it's we are coming back to job def- definition, uh, allocation of headcounts to this kind of value chain change.
This is from LogiRoot Runtime AI governance, with verifiable, and I can't read the rest, because verifiable audit evidence. That's his name, okay, and LogiRoot Runtime says, when a regulator or a court asks a bank to prove this agent did exactly what you say it did, and prove nobody touched the record, does today's tooling actually clear that bar, or are we one incident away from learning that having logs and having proof are not the same thing?
you know, it's independent of the technology, you know, certain governments say, you know, whatever technology, using, make sure there's traceability, auditability there, and that you can verify the transaction, and then you can see if there's any misstep, uh, from that perspective. You can go to the next, uh, level of maturity, saying, you know, I'm using blockchain, that nobody can even correct, uh, um, uh, transaction or information past that, uh, production time. So, so, I would say, um, I that that's not a specific agent topic.
I think we always have to prove that, you know, these kind of businesses are falling, uh, according to the standards and and rules of of, uh, the local jurisdiction, and you, you need that evidence in place.
And this is again from George Mertens, and I'm taking this because I like questions from CFOs, and his question is something I was going to be asking you anyways. And and he says, Oliver, from your perspective as a supervisory board member, how has AI changed the way you exercise oversight?
I think the the it has changed. Number one, I think the, um, in if I look at financial financial, um, institution that I'm, uh, in general, I would say the the board agenda focus has shifted the last couple of years into much more operational risk. Operational risk is most of them is is using of technology and operations, um, for the day-to-day banking, and and that means, you know, with the new technology, uh, board members, you know, first of all, have to be up to speed, you know, what does it mean using AI and what are the risks, the limitation, and the way how this being, uh, oversight, et cetera, and their role is making sure, you know, if if if
I talk to the to the ECB as my regulator, uh, they're telling me make sure that, you know, for this kind of topic, the existing regulation, like the European AI, the Dora, um, uh, jurisdiction, how to classify vendors and issues, um, is being followed for this kind of issue. So, you know, as a supervisory board member, what you do is you ask questions, you know, how's the setup? how's the governance? also pull, show me evidence that this is in place, that, um, we have the right, uh, safeguards in place to utilize this kind of new emerging technology.
This one is from Ricardo Warhan, and he says, Oliver, in the past, new technology has enabled the creation of digital native banks like Revolut. Do you expect AI native banks to emerge, and how would they look like?
Digital native banks are also embracing, like Revolut and other banks, this kind of technology from automating processes, also or changing the engagement between customers and and the bank. Um, do I see, um, new software companies calling AI agent software out there and saying we have 200 agent? I see this too. Um, do I think o- over time we will be an agent-based, um, uh, a bank out there? Yes, but what I learned over the last thirty-five years is, um, clients love to talk to a person, to a human person, so I don't think that we will have a fully automated, uh, agent-based bank out there without any human-to-human interface.
And just the trust issue of do are you really going to trust AI agents? I mean, eventually, maybe, but we're not there yet, we're not even we're nowhere close.
Yeah, and, Michael, that's the experience of the last fifteen years. It's a robot advisor. We thought about, you know, everybody is, you know, going online, try to analyze their their financial situation, and you get a re uh, recommendation, and you execute to reshuffle your your hundred thousand hundred fifty thousand dollar, uh, portfolio. That not really. People like to talk to a bank at the end, say, you know, is it a good decision? And then usually you see this kind of execution.
You know, it wasn't too long ago where if you had what appeared to be an error on a bill, for example, from utility company or whatever, people would say, oh, the computer made a mistake. We don't we don't say that too much, uh, now, but the reality is we don't, we don't trust computers. Not not when it comes to issues that really matter, that seem to be going against what we want.
Exactly, and that's something I think we always have to think about, um, the human interface now, and, uh, there will be wealth management, um, private bankers, uh, advisors still in the loop with, uh, with the client, but the way this advice, the financial advisor will be supported, uh, will be much more flexible and customized that, um, um, that today is more standardized and limited.
Important question from Arselon Khan, really fast, please, uh, with so much use of AI, what can consumers do to protect themselves from hallucinations?
From a consumer perspective, my advice is always cross-model, uh, checks. So, whatever I do, and depending on the criticality, I use different models to verify that, and that's, uh, is is is always reducing the risk of, um, any misinformation, et cetera, and then my personal last check, too. So, uh, don't forget to read, uh, important information, outcomes that you wanna share, making sure that everything is correct.
And the last question, a year from now, are banks running agents in production at scale, or are we still having this conversation about all the obstacles and why we can't do it, and why it's a good idea, and we know it's a good idea, but it's we can't really do it.
I'm very optimistic that we'll see, um, a lot of agents across, you know, all function in in production with the bank, because the the upside, the the the productivity gains and the the quality improvements are significant. As I said, the copilot, uh, area is ten ten, twenty thirty percent improvement. Here we're talking about, uh, one, two, three x in certain use cases. So, if you have a front runner of looking at your competition, and if that that institution is achieving that, has a huge impact on performance, cost, uh, uh, cost income ratios, et cetera.
That's something that the peer pressure will will even drive more adoption.
Okay, Oliver Bussmann is the former group CIO of UBS, he was the former CIO of SAP many years ago, which is when I first met Oliver, and he now advises banks across Europe. Oliver, thank you so much for taking your time to be with us today. I'm very grateful to you. Everybody, thank you for watching, especially you folks who ask such great questions. Before you go, number one, connect with both Oliver and me on LinkedIn. We we love that, and number two, subscribe to the CXOTalk newsletter.
We have just I I always say this, and it's really true. We have extraordinary shows with amazing people coming up, and we want you to join us and participate. Thank you so much, everybody, and I hope you have a great day, and we'll see you next time.