Banks and insurers may be softening their decades-long resistance to core IT renewal, as AI raises the stakes and exposes limitations across every legacy decision. That tension, and why that stance needs to change, was at the heart of a dinner discussion that I recently moderated for SoftServe and Google Cloud at London’s House of Lords.

Conversation among the senior technology and operations attendees from across financial services focused first on why attempts at modernisation have stalled. But the wide-ranging debate also explored what AI now demands of older systems, and whether many underlying issues are technical at all.

From that start it was clear the value of the evening would come from the shared expertise around the table, with the first question posed becoming the core of the discussion: “Is something deemed legacy because it is old, or because it is a problem?”

Legacy paradox

Attendees were quick to defend their existing estates. One said legacy systems provide a competitive advantage, refined over decades to solve real business problems that newer software rarely understands. Another pointed out that systems are never really replaced. Layer is added on top of layer, until no one is quite sure what the original platform does. That uncertainty itself becomes a reason not to touch it.

The arguments against modernising piled up quickly. Customers do not care about the back end, said one attendee, and any project that fails the value-to-client test is hard to defend. Cost is another perpetual worry, especially when modernisation can convert Capex into Opex and disturb established financial models.

Decommissioning is also a brutal proposition, one said. It is harder than building, and harder still to get past sign-off committees while everyone is busy running the day job. Several attendees agreed there is a shortage of success stories. This meant that, until peers can point to a clean migration that paid off, the inertia argument keeps winning.

Nevertheless, it was agreed that replacing systems simply because they are old has no merit. But the cost of doing nothing changes once new technologies enter the picture.

AI catalyst

Some attendees acknowledged that AI might now be the external event catalyst to change that calculation. Several said their organisations were under explicit top-down pressure to deploy AI tools. One participant reported that staff had been told their jobs would be safe if they used AI to improve productivity, and at risk if they didn’t use it.

That pressure has not translated into universal enthusiasm. Banking has not fully bought into AI, one attendee said, because the industry doesn’t yet trust it. A counterpoint came quickly when another participant said their bank had several AI projects running and was eager to do more. A third was more sceptical, citing the hype and the pricing that assumed a perfect technology, but one that still made mistakes. “Has anyone really achieved AI success at scale”, the attendee asked, “or are we still just looking at impressive pilots?”

But some cited examples of concrete wins. One attendee described using AI to process documents related to car-finance mis-selling claims, a paper-heavy problem that AI technology handled well. Another said their call centre had moved from reviewing a random sample of calls to having AI flag every conversation that needed a human ear. Instead of reducing headcount, they delivered a better service.

The discussion showed that while, for some, AI was an excuse to modernise, in many cases, it still required a business case. The harder question, some suggested, is often whether the underlying systems can support it.

Questions of trust

Trust underpinned much of the conversation. Trust in new technology, trust between technologists and the business, and the customer trust that financial firms depend on. It is why, my co-host from Google Cloud noted that firms like his have put such a premium on high security standards for data and other assets.

Regulation always shapes the risk appetite in financial services. UK regulators were described as still fearful of firms even trying AI, partly because they want proof that decisions are correct, which a non-deterministic technology could not easily provide. Other attendees were more sympathetic. The regulator was helping firms modernise properly, one said, holding them to standards they might otherwise duck. Internal audit teams, another added, often set a higher bar than the regulator anyway.

But most agreed that the world is moving faster than corporate governance can manage. That makes the choice of risk frameworks as important as the choice of technology. Can the new system be governed? Can the AI be controlled?

One also noted that younger businesses can sometimes carry a ‘technology premium’ in their valuation, as an investor reward for not being weighed down by mainframes.

Older firms cannot replicate that, but they can take a different route. Incremental change is often more sensible than a big-bang replacement, particularly as the people who have maintained mainframes for decades start to retire from the workforce. But, even here, AI itself may help, one suggested, by writing and refactoring older code, including COBOL.

Google supported that conversation by highlighting how adopting modern Agentic Engineering can be utilised in legacy environments by providing support that maintains existing estates but also uses in tandem a modern, more efficient approach.