Ron van Kemanade, Group COO, Lloyds, on agents, COBOL, automating fraud detection
"Half a year ago, if I would have gone in and said, 'guys, we have 200 COBOL applications, let's refactor all of them!' they would probably have run for the woods.”
Like many technology leaders (and despite having a Chief Operating Officer title, Ron van Kemenade ultimately owns technology delivery) Lloyds Banking Group’s COO has application modernisation always on some part of his mind.
It’s not the shiniest, sexiest part of the job, but few organisations with a three-century history can go far without transforming entrenched legacy systems.*
The bank has got rid of over 900 since he joined, he says. He plans to prune another 500 away as part of LBG's broad standardisation efforts across the stack.
AI is making it somewhat easier. Critically, it’s almost making it, dare he suggest, a bit more enticing. As he puts it: “[Just] half a year ago, if I would have gone in and said, 'guys, we have 200 COBOL applications, let's refactor all of them!'” his engineers would “probably have run for the woods.”
Now?
“If I do this now, I'm sure they will come up with a plan how to do this [and show] how fast that could be done,” he says, sitting down to catch up with The Stack – some 16 months after we engaged in an expansive discussion about the banking group’s huge digital transformation efforts.
Application modernisation
That confidence is built on recent experience: a little taste of success goes a long way. In the back-end of the mortgages space in particular, the bank had “quite some legacy COBOL code residing on mainframes,” he explains.
Reams of undocumented COBOL with mysterious subroutines are not for the faint of heart, and an “if it ain’t broke don’t fix it” approach often prevails in most quarters, whatever the expense/rigidity of the underlying systems.
But the palpably improved ability of LLMs to make sense of arcane code has engineers and COOs alike intrigued.
As he recalls: “A couple of engineers said, ‘Ron… couldn't we use this as a good reason to start using agentic AI, analysing the code, re-documenting the code; not just refactoring, but actually re-imagining [the application.]
He gave them permission and they made good headway.
"It was a pricing application that has been fully refactored. It is in production. We run it in parallel to the current application. If the… outcomes are [consistently] predictable, the old one will be decommissioned.”
The process took the banking group six months, where in the past that process would have taken formidably longer.
(The mainframe, he hastens to add, intrinsically, “is a very modern, hyper-resilient platform” – it's the legacy applications “that we [need to] chip away at.”)
*Per our 2025 conversation, LBG still runs a suite of AS400 or ‘I-series’ systems, Solaris, HP Unix, HPE NonStop, Unisys OS, IBM Z/OS, alongside modern containerised applications.
Agentic challenges
Pressed on what has been most challenging during his tenure so far, van Kemenade (who joined LBG as Group COO in June 2023) pauses briefly.
Probably the sheer speed of technology change and what that means vis-a-vis the group’s operating model, he says. Plenty of the foundations were already in place, e.g. he had already combined engineering data specialists and product people, started work to standardise platforms, etc.
But the switch from just “LLMs” to thinking about agentic workflows meant a lot of thinking about a lot of still-moving parts. As he puts it to The Stack.
“Your data integrity, your data quality is becoming even more important. [You have to think about] how do you manage and control the interactions between agents.
"Your workforce needs to be reskilled and to a certain extent rebuilt. Your business leadership and senior leadership [need to be] able to reimagine what a great customer journey could look like based on AI.
“It's how you reprioritise your backlog and how you put funding behind it. It's your operating model as an organisation. How do you organise where people might actually not manage just a couple of colleagues, but actually manage colleagues and agents? [These challenges] accumulate, right? They all happen in parallel. It's not like you're going to do this one by one.”
“We needed to truly scramble… to avoid every business case [going] out on its own, [creating] its own little platform, and then we lose scale and control over everything we do. So that was a challenge. Was it unexpected in itself? No, but the pace in which that happened, that was definitely a challenge.”
The bank is building agentic use cases fast, he says, running regular sprints and experimenting in sandboxes. (In 2022, the bank embarked on a £4 billion investment in AI and technology more broadly that its COO oversees.)
Using agents to tackle fraud
Is it easy though?
It is not: “The promise and the practice? There is still a gap!”
“Where an agent starts to interact with other agents, and then you introduce orchestrating agents to manage that interaction, and you introduce compliance and risk agents to monitor the activity itself, that introduces a whole new issue about identity and access management, something that in the current enterprise software world is already a challenge for people…”
Useful use cases are emerging fast, he says.
One is in fraud, where LBG is deploying agents to assess potentially fraudulent transactions and drive faster action across its fraud help desk.
Under the program, LBG says “multiple AI agents will operate simultaneously behind the scenes, carrying out tasks such as identity checks, transaction analysis and scam risk assessment in real time.”
Like the bank’s other AI programmes, this has been built on its multi-agent Envoy platform, which it launched in April.
What is Lloyds' "Envoy" environment?
Not to be confused with the open-source Envoy AI gateway, LBG’s Envoy is a standardised environment to design, build, deploy and operate agents.
Pressed for more detail on the environment after the interview, an LBG spokesperson taps the Envoy team and responds, describing it to The Stack by email as a way to abstract the heavy lifting around agent development, packaging infrastructure, pipelines, security and governance.
“Delivery is governed through a single, standardised CI/CD pipeline powered by Harness… Under the hood, Envoy is built largely on GCP native technology, curated for use within a highly regulated environment, including the Agent Development Kit (ADK) for building agents, Agent Engine for runtime execution, and GCS for audit log storage. Agent and tool marketplaces enable discovery and re-use across use cases and are supported by standard interaction patterns / protocols (A2A and MCP)."
They added: "All model access is decoupled from the agent application and centralised through Cortex, the Group’s LLM gateway, ensuring consistent governance, monitoring and control across every model call.
"Cortex provides API-based access to a portfolio of leading models, including Gemini 2.5, Claude (Opus and Sonnet), Mistral, and an increasing range of smaller, specialised models allowing teams to evolve model choice without redesigning their agents."
As van Kemenade puts it: “Ingesting that transaction history and identifying [issues] with five to six identifiers… All of that can be done in split seconds by an agent against a large language model, and prompting with the right context.
“[That] leaves the agent to deal with the actual assessment:is this now fraudulent or not?"
"This is actually in production, we're learning about the effectiveness of it and to what extent it helps. but you can kind of foresee in the future ‘why did we make the customer call us in the first place? Why didn't we deal with this in the background?’ Going forward… we could deal maybe with 99% [of such incidents] in the background ourselves.”
Another area where the bank is building out capabilities is in digital assets. It’s been working with several other banks, for example, on what they call “Great British Tokenised Deposits (GBTD)” under a UK Finance programme.
Lloyds is also looking at mobilising tokenised real‑world assets to meet margin needs, using tokenised deposits to settle tokenised securities, and connecting on‑chain workflows with traditional accounts and market infrastructure. The bank is also collaborating with LSEG’s Digital Securities Depository (DSD) as the industry moves toward deeper on‑chain settlement capabilities.
Van Kemenade is upbeat about the possibilities: “It feels like we're at that inflection point, where all of a sudden [almost like the] birth of ChatGPT, this is now all happening with stable coins. We see it in the mortgage market, people are looking at use cases; there are national banks working on CBDCs.
“I'm quite proud to say that Lloyds has really taken the lead on this in the UK… then even more exciting is the convergence of AI and digital assets.”
What is the potential at the nexus of those two technologies?
The potential of speed when algorithms can drive more rule-based decisions in markets activities, he suggests, that reduce the need to provide more collateral… “[sometimes you] need to sell a money market fund, then transfer the money (that may take overnight, or even sometimes two days) to be received on the side where the risk, or you have the uncovered risk.”
“If you could imagine that we tokenise that money market fund, and we can literally, like in milliseconds, execute the transfer of part of that money market fund to cover the risk in the contract, you have a richer decision-making environment that could operate in real time; [reduce a lot of friction from] manual activity… where AI makes for richer and faster decisions, and the execution against the uncovered risk uncovered can be done in split seconds”?