Eric Bierry said SBS needed six to eight more months to build the technical controls for “agentic” AI. At present, Scottish Building Society, the first client to use SBS’s new AI capability, uses generative AI rather than live agents.

The harder part, Bierry said, was whether lenders would accept agents taking decisions and gradually remove the “human in the loop”. He said governance is one of the hardest issues for institutions because regulators require every decision an agent makes to be explainable and repeatable, with the same event always producing the same outcome.

As a result, SBS is building a “control plane” to set what agents can and cannot do, creating a safe sandbox to test their abilities, alongside an AI gateway from Axway.

Bierry said the work with Scottish Building Society began about two years ago with a request to help the society differentiate itself as a mid-sized lender, at a time when consolidation is expected.

He said Paul Denton, the society’s chief executive, initially wanted to make better use of its data, and SBS added a data platform and a layer making that data readable by AI. He claimed the society’s business teams now write their own prompts to launch campaigns faster and to understand client profitability and churn risk, and that tasks that once took five or six weeks can now be done in seconds.

The software boss said two other building societies were close to signing for the same capability and he hoped to announce one or two names soon. He said up to 30 institutions could eventually use it, as it is an evolution of SBS’s existing core lending platform for mortgages and savings. SBS has 65 UK clients, making the UK its second-largest market.

Speaking to Scottish Financial News, Bierry said: “It’s linked to the ability of the institutions to make it happen and to accept that change. There is a lot of resistance for some of them, and then there is the AI governance inside the institutions, which is making a difference, allowing only a human in the loop, and only for generative AI. But in the future the agent is going to take some decisions, so where is the human in the loop?

“All of this is a major shift, and that shift is taking more time than the speed of the technology. That’s where I think the main challenge for them lies in the future.

“In the past we were adding API orchestration and defining which APIs could be accessed, so it’s really about defining rules. But an API is static, right? An agent is dynamic. So how do we manage all of this compliance and security, avoiding hallucination, or avoiding an agent trying something it’s not allowed to do inside that control plane?

“Anything which is happening with an agent in financial institutions has to be explainable at any point in time. It’s enforced by the regulator, by the way. If you replay exactly the same context, exactly the same event, it has to give, on a deterministic view, exactly the same result or the same decision.”

Bierry admitted: “None of this can guarantee 100% that there will not be a problem. It’s like cybersecurity. But at least you have to demonstrate, first, that you are putting all of this in place, and second, that the audit trail and the guardrails will flag it if something starts to behave differently from what you expect. That’s really where we are working, and I think it’s the most complex thing.

“Having an agent is not so difficult from a technology standpoint. I think guardrails are the most difficult part, and in a regulated industry it’s even more critical.

“Nvidia has announced that they will be releasing, by the way, for free, the first level of control plane, to make sure that everything happening will be at least governed and trustworthy. We think it’s not enough, because Nvidia is not going to specialise in regulated markets and their additional constraints.

“So our responsibility is to build on top of Nvidia’s capability and provide the additional guardrails, linked not to the infrastructure but to business clients.”

SBS is working with Nvidia so that it can test and train its applications on “Nvidia chips”. Bierry said this matters because European and UK clients want to choose between public and private cloud and between AI models, and some will buy their own GPUs. Many clients are happy to use external AI services, but others want to run everything themselves, with full control over the environment. “Whatever the choice made by our clients,” he said, SBS’s applications must behave identically.

Bierry said bank chief executives were risk-averse about legal and data risks and AI errors, and that real-world uses are modest but valuable.

He revealed a French bank is trialling SBS technology can have agents repair about 90% of failed payments that staff previously fixed by hand. Humans are checking the agents at first, but he expects humans eventually to be phased out of this work while keeping an audit trail.

He explained: “The reality in our business is that today it’s doing very basic things. I’m saying basic, but that does not mean it’s not generating a business outcome and benefit. We are moving from tasks taking five weeks to seconds. But on very basic expectations, that’s what’s really relevant.

“This bank is doing a lot of payments, so they have people in the back office repairing payments because sometimes there is a feed missing something and the payment is not executed. It comes into a queue, and people in the back office check the payments, ensure the feed isn’t missing and resubmit them, and the payment is repaired.

“So when 90% of that payment repair is managed by agents, at the beginning you put the human in the loop, and then there are patterns, and maybe in one year there will be no human in the loop to repair them. And you will still have the trust and the audit trail and explainability of what the agent changed to make sure that the payment is processed.

“When you put that infrastructure in place, anyone can bring a new agent capability. So what it means is that many actions you do in your core processing platform, to calculate interest on a mortgage loan, for example, there are some of these capabilities which will be actioned outside the core, so the calls will start to reduce, to become a ledger, or to become a system of record, to secure the deterministic part of any calculation or any result of an action.

“And then have more and more agents accessing this system of record… Today, the big part is the back office, and you have an engaging layer and you have the data. I think that big parts will divide by two or three to be only the ledger. You will be having the control planes, all that capability to go there, and you will be having more and more agents on top. So that shift, I think, will be acting more and more in the next two to five years.

“And you will see new players like us… protecting what I call the fortress, the system of record, which will stay in any case, to keep the banks’ solutions fully deterministic.”