Most banks would accept black-box AI if it performs better, survey finds

Adoption is near universal, governance is not

The report, titled Why AI at Scale Needs Governance at Speed, is based on a global survey of 300 banking leaders across 31 countries. It was conducted by FT Longitude, a thought leadership agency owned by the Financial Times, with data and AI company SAS and its strategic partner Microsoft. Fieldwork took place in June and July 2026. Twenty percent of respondents were C-suite executives, and 66% represented banks with more than $50 billion in assets.

According to the findings, 299 of the 300 banks use AI in credit and risk decision making. Some 82% say they already use generative AI for decision making, and 41% report using agentic AI. Yet 82% also say they lack fully operational and embedded governance processes for AI models, and 26% have no dedicated AI model risk policy.

Other control gaps are listed in the report. Only 19% have real-time AI model drift monitoring, and 36% continuously retrain or update their models. Some 42% have no defined "kill switches" to deactivate AI models that cross preset boundaries.

Black boxes and ungoverned models

The survey points to a tension between performance and transparency. The 63% of leaders willing to use a black-box model, one with little or no decision-making transparency, would do so if it offered better performance, returns or liquidity than an explainable option.

Many banks appear to be acting on this already. A full 92% of respondents reported that at least one large language model in their institution was developed or operates outside formal governance frameworks.

Tyler Pichach, Global Head of Payments, AI Strategy and Go-to-Market at Microsoft, warned that the short-term boost from a black-box model comes with a cost. In his words, its "lack of visibility, repeatability and explainability spells trouble from a regulatory and risk perspective."

Autonomous decisions are coming

Few banks let AI decide on its own today. Only 10% of respondents say AI makes autonomous credit or risk decisions at scale. Expectations are far higher: 76% expect that within five years, and 25% expect it within two.

AI is already embedded in many processes. According to the survey, 59% of banks have partially or widely embedded AI in decision-making, and 53% say AI influences departmental decisions at least monthly. AI is partially or widely embedded in 63% of credit risk processes and 55% of asset and liability management and risk steering processes.

Canadian digital bank Tangerine, which began using AI in lending decisions in 2023, illustrates the balance. Nishant Taneja, its Head of Credit Risk, said that in retail credit portfolios "more than 90% are AI-based decisions with very little human-in-the-loop methods," while for high-risk portfolios the bank has "intentionally kept the human in the loop, until we have more confidence in the technology."

Regulators urge caution

Regulators are also signalling restraint. Marilin Pikaro, Director of Innovation, Conduct and Consumers at the European Banking Authority, told FT Longitude: "Ambition should match its capacity." She warned that even compelling AI use cases should not advance faster than an institution's ability to manage the risks.

Aurelio Maccario, Head of Group Risk Management at UniCredit, said in the report that governance means setting boundaries and rules, and that within them a bank is "ready to run."

For banks, the findings suggest a widening gap between ambition and controls as autonomous AI approaches. SAS argues that institutions that advance technology and governance together will be able to decide faster without sacrificing transparency, oversight or regulatory confidence. SAS will present related work at RiskMinds International 2026 in London on November 16-19.