How Banks Can Trust AI: What Tech Leaders Are Saying and What It Means for Credit Teams

Banks can trust AI the way they trust any powerful tool or new employee: by limiting what it can do, checking its work against the bank's own rules before it takes effect, keeping decisions with named people, recording every step and keeping the ability to stop it at once. Trust comes from control, not from the model.

That view is not only a banker's caution. Many of the people who build these systems are saying much the same thing in public. The sections below summarize what they and bank regulators have said, then translate it into specific controls.

Anthropic CEO Dario Amodei, in his April 2025 essay "The Urgency of Interpretability", wrote that people outside the field are often surprised to learn that "we do not understand how our own AI creations work." He noted that this opacity keeps AI out of many high-stakes financial settings because its makers cannot fully set limits on its behavior, and he pointed to mortgage assessments, where decisions are legally required to be explainable, as an example of a legal blocker. He set a goal for Anthropic that interpretability can reliably detect most model problems by 2027.

Microsoft CEO Satya Nadella, in an essay titled "Models as Insider Risks in the Super Intelligence Era" published in October 2026, argued, as reported by OfficeChai and The Deep View, that companies should treat AI models like insiders with powerful access, keep the controls outside the model, separate the supply of intelligence from authority over it, and keep the ability to contain a model's actions. He also argued that a model provider's assurances do not relieve a company of responsibility for what AI does on its behalf.

Google CEO Sundar Pichai wrote in February 2026 that AI should be pursued boldly and approached responsibly, and that "trust is the bedrock of adoption." Google DeepMind CEO Demis Hassabis, speaking at the India AI Impact Summit the same month, described current systems as having jagged intelligence, very good at some things and poor at others, and named consistency as one of the gaps that remain. OpenAI CEO Sam Altman, on the OpenAI Podcast in June 2025, was widely reported as saying people place a high degree of trust in ChatGPT even though AI hallucinates, and that it should be a technology people do not trust that much.

Federal Reserve Governor Michael Barr, in an April 2025 speech on AI, fintechs and banks, said generative AI systems may still hallucinate and that answers can differ for the same query, which is tough to square with banking, "where decisions must be well-controlled, numerically and legally precise, explainable, and replicable." He said successful use of generative AI in banking requires "getting the guardrails right."

JPMorganChase CEO Jamie Dimon, in his letter to shareholders published in April 2026, wrote that "AI will affect virtually every function, application and process in the company." He also wrote that AI will introduce serious new risks, and that "These risks are real, but they are manageable if companies, regulators and governments prepare."

The banking agencies' revised model risk guidance of April 17, 2026 (SR 26-2, OCC Bulletin 2026-13) states that generative and agentic AI models are novel and rapidly evolving and are not within its scope. It adds that a bank's own risk management and governance practices should determine the controls for tools it does not cover, and the agencies said they plan a request for information on banks' use of AI. In other words, for now the controls are the bank's to design and to show.

Read together, the statements point to the same few ideas. None of them says AI should not be used. All of them say that trust has to be built around the model, by the organization using it. The table maps each idea to what it means for a credit team.

Whether the AI comes from a core provider, a fintech or an in-house team, a credit team can ask for the same set of controls. Each one should be something you can see working on your own files, not a line in a contract.

A control that lives inside a model, such as an instruction in its prompt to follow credit policy, depends on the model following it. That is the very thing the bank cannot be sure of. A control outside the model is a check the model cannot skip: the step is proposed, the bank's rules are applied, and only then does anything happen.

Keeping controls outside the model has a practical benefit too. Models are replaced often. If the bank's rules, signers and records sit in their own layer, the bank can change models without rewriting its controls, and it can compare models on the same rules and the same records.

Trust should be earned on the bank's own files, in steps, with evidence at each one. A simple sequence works for most credit workflows.

Some versions of trust are really just hope. A credit team should be wary of each of these.

The statements summarized in this guide come from the sources below. Paraphrases are ours. Direct quotes were checked against the original page where it was available. For Satya Nadella's essay, Demis Hassabis's summit remarks and Sam Altman's podcast remarks, we relied on the press reports listed.

Aarvion Risk OS is built around controls that sit outside the model. Every AI step in credit work is checked against the bank's own credit rules and delegated authority before it takes effect, and is allowed, held for the right signer, blocked or stopped. Each step is recorded with who proposed it, what evidence it used, which rule applied, who approved it and when.

Figures trace to source pages, decisions stay with the bank's named people, and the bank can stop all AI activity at once. Banks, credit unions and lenders typically start with a 90-day pilot on one workflow.