AI and Lenders: Who’s Liable if LLMs Err?
Private credit firms can still whiff if algorithms do the work, but the onus is on them.
Lenders are leaning on artificial intelligence to score borrowers, monitor portfolios, and automate workflows that once took analysts weeks. Momentum is only building: More than half of private credit portfolio managers—54%—plan to deploy AI in underwriting, according to a March PwC survey of 120 global firms.
But as AI takes on more of that analytical heavy lifting, firms face a tough question: When an algorithm makes a mistake, who bears the blame?
For credit risk expert Naeem Siddiqi, author of Intelligent Credit Scoring and senior risk advisor at SAS, the answer is clear: Don’t fault AI; it’s just a tool.
If the large language model, or LLM, miscalculates a number or uses a prohibited category like race or religion, “then the lender is liable,” he said in an email. The courts already tested that principle — that a company can’t hide behind its own algorithm. Guess what? The company lost.
‘An Emerging Discipline’
Take Moffatt v. Air Canada for example. One of the airline’s customers used its chatbot in 2022 to ask about bereavement fares following a death in his family. The chatbot told him he could book a full-fare ticket and apply for a refund within 90 days, advice that contradicted Air Canada’s actual policy requiring passengers to submit such requests before travel.
When the customer tried to collect, Montreal-based Air Canada argued it shouldn’t be held liable, effectively treating the chatbot as a separate entity responsible for its own statements.
The British Columbia Civil Resolution Tribunal rejected that defense, found Air Canada liable for the error and ordered the airline to pay $812.02 Canadian dollars, including CA$650.88 in damages plus interest and fees.
Siddiqi said the ruling set a precedent: “Companies can’t argue that the AI is a separate independent entity that frees the firm from liability.”
He pointed to a broader wave of AI-related litigation in the U.S. where legal exposure extends far beyond chatbots and the airline industry. Currently, there are copyright suits against LLM developers, including Anthropic. However, in other scenarios, the company wielding the tech bore the brunt of scrutiny.
Last year, facial-recognition company Clearview AI faced privacy litigation while software firm Intuit and HR tech firm HireVue received a discrimination complaint alleging their AI hiring tools disadvantaged a deaf, Indigenous job applicant.
“This is an emerging discipline,” Siddiqi said, “but it’s safe to assume the lender is liable for discriminatory decisions made on its behalf, whether by a human or an AI.”
Risk Sits With Lender
For private credit firms racing to deploy AI across underwriting and portfolio monitoring, the early case law sends a signal: The technology can do the work, but it doesn’t absorb the risk. That still sits with the lender.
“Legally and regulatory-wise, the buck stops entirely with the lender,” said Omar Abassi, founder of Newport Beach-based lending tech startup LoanFlo AI.
So far, regulators such as the Consumer Financial Protection Bureau, the Office of the Comptroller of the Currency and the U.S. Department of Housing and Urban Development have made it clear: You can’t delegate your compliance obligations to a software vendor, Abassi said.
If an AI algorithm introduces algorithmic bias, violates the Equal Credit Opportunity Act, or fails to provide legally compliant adverse action notices, regulators sue or fine the lender—not the AI company.
Because of this legal exposure, some lenders require vendor platforms to provide audit trails showing exactly what the AI read, and regular back-testing to prove the AI model does not inadvertently produce discriminatory outcomes.
Treating AI Like an Employee
That gap between high market interest and actual operational risk is top-of-mind for technology leaders building loan administration tools.
“There’s a general enthusiasm in the market around AI … and firms are very excited about diverse capabilities,” said David Yahalomi, chief operating officer and co-founder of Tel Aviv-based loan-management platform Hypercore. “But this technology is a statistical-based technology … it can make mistakes, and we’ve all seen that.”
Rather than viewing AI as a replacement for decision-makers, Yahalomi suggested lenders treat AI like a new hire who requires guidance and thorough review.
“We should treat it like it’s an employee,” Yahalomi said. “Even if you feel like you’ve trained your best agent … think about it like you gave a deal to your best person five minutes ago. Would it give you the correct answers, or does it need proper time to actually go and research?”
Ultimately, Yahalomi cautioned against granting agents final authority over deals: “We should not treat it as a person that makes calls … you shouldn’t treat it as an executive.”
What’s Next
The balance between strict regulatory oversight and day-to-day workflow is where human teams feel the pressure most. As LLMs become more ubiquitous, too few humans are taking on too much work and leaning heavily on AI-driven underwriting.
“Underwriters are definitely taking on too much work in traditional setups and being overworked in many cases, which leaves more room for human error,” Abassi said. But don’t expect AI to replace credit risk assessment; instead, it’s closing the gap so that fewer underwriters can underwrite many more loans and be less stressed as a result.
“Eventually, the AI will be so good that human underwriters won’t be able to keep up,” he added. “AI agents will be the ones reviewing the other AI’s work. We aren’t there yet, but ultimately it’s on its way.”
Anthony Noto covers corporate finance and private credit. Contact him at anoto@gfmag.com.
