The Expert View: AI and credit - from pilot projects to production

The experimental phase of artificial intelligence (AI) in credit is over, and the technology must now earn its keep. That much was agreed by senior leaders from major financial services firms at a Business Reporter dinner briefing at the House of Lords, hosted by S&P Global Market Intelligence. But there was little consensus otherwise. Attendees split over whether the spending can be justified, staff will use the tools, and who answers when AI gets it wrong.

S&P Global Market Intelligence first hosted this conversation a year ago, said Clemens Thym, Global Head of Credit and Risk Solutions at S&P Global, but AI had moved so fast since then that it needed to happen again. Clients were excited but also concerned, he said, and every aspect of credit was up for consideration as firms worked out how to make credit decisions future-ready. To support clients, the business has moved quickly to connect its trusted financial intelligence, including credit ratings from S&P Global Ratings, to leading AI platforms as well as its clients’ AI applications and agentic workflows through the S&P Global AI Data Portal.

The bill arrives
There is a growing expectation that AI shows value, one attendee said, and less patience for experimentation. The gains are real: a risk report that once took a week now takes two minutes. AI runs over complete datasets where it once handled samples, which has improved the quality of what one person produces, though it hasn’t reduced the workload.

The harder question is cost. Firms were less rigorous during the experimental phase than with other technology, one attendee admitted, because AI companies made experimentation a loss leader. Almost every banking process could be improved by AI, another said, but is it worth it? Costs are counted in tokens, the units AI providers charge by, and even asking AI to check its answer uses more of them.

The group debated whether banks would move to cheaper models or use tokens more efficiently. One attendee raised a sharper worry: what happens when staff have forgotten how to use Excel and token prices increase further? By then, there is no manual fallback. Against that, one participant argued, the firms that use AI best to drive return on equity will take market share. Private credit and fintech firms are already using AI to identify future risk, another said, though getting that working inside a big firm is harder.

The adoption gap
Adoption is the hardest thing about AI, one attendee said, and experience varies widely between firms. One participant’s previous employer encouraged AI use wherever possible; their new company barely uses it. Another firm has trained staff and given them a mandate to find use cases that cut costs or raise revenue.

But staff may have incentives to hold back. Why adopt AI, one attendee asked, if it will cost them their jobs? Another felt safe because their value lay in talking to clients, which AI won’t do in investment banking, though it may in retail. Plenty of employees see no benefit in their employer’s tools and use personal AI subscriptions instead. This is shadow AI: staff using tools the firm hasn’t approved.

Knowledge is thin at the top too. Many decisions are made by people who are not AI-literate, and with a new model every six to eight weeks, keeping track is hard. Then there is the risk of losing skills. One firm gives graduates a basic chat tool instead of the full Microsoft Copilot, because it wants them to learn. The buck stops with you, as one attendee put it, and if you lose the skills you aren’t AI’s ‘copilot’ anymore, you’re a passenger.

A human must answer for the decision
In wholesale lending, a human must make the decision, one attendee said, because regulation requires an individual to take responsibility. Where AI is given a role, the limits are tight. At one conservative bank, it can approve automatically at a low level but can never decline without a human in the loop. Regulators also expect banks to be able to explain their decisions, and the scrutiny that brings is heavy: at one large bank, 60 groups had to review a single AI implementation, partly to reduce bias.

Accountability divided the room. One view was that you can’t blame a tool: follow it, and the fault is yours. But if faulty information from an AI could lead someone into a bad decision, should they check the whole calculation? And if so, where is the productivity gain? The opposite risk is that people treat AI as gospel, when expertise is sometimes needed to challenge it.

Trust and guardrails
Hallucination remains a concern. However, S&P Global Market Intelligence has developed a tool that gives the same answer to the same question every time and doesn’t stray outside its dataset. That matters because some tasks must be deterministic, producing the same output every time: payroll, for example, must pay everyone the right amount without fail. The same principle applies to financial and credit risk workflows, where users need reliable access to trusted datasets such as Credit Ratings data. Through Deterministic Retrieval and the S&P Global AI Data Portal, users can access credit ratings data and other S&P Global datasets in a controlled manner, helping ensure AI-driven applications are grounded in authoritative information.

Multiple agents can check one another’s work, but what governance sits around them? Firms must set boundaries for any agent they deploy so it doesn’t go rogue, and training in AI governance and risk matters too. Shadow AI makes that harder.

Closing the evening, Jim Elder, Global Head of Market Development and Strategy, Credit & Risk at S&P Global Market Intelligence, said he had enjoyed the passionate conversation. Everyone was further along the journey and starting to implement, he said. Their thinking on cost, he said, was more advanced. Firms were talking about redeploying people instead of laying them off and learning to refine their approaches by breaking tasks into chunks and optimising token use.  One thing is certain, if this group reconvenes one year from now, we can expect even more advanced thinking, adoption and opinions on the topic.