Banks want autonomous credit decisions by 2031. Only 18% have the governance.
The SAS/FT Longitude survey finds 76% of banks expecting autonomous risk decisions by 2031, 92% running LLMs outside formal governance, and 63% willing to trade transparency for performance. A new preprint, Quant-Arena, offers a sealed benchmark — and no agent beats the market.
- The SAS/FT Longitude survey found 76% of banks expect autonomous risk and credit decisions at scale by 2031 — a quarter expect the shift within two years — while only 10% make autonomous credit or risk decisions at scale today. Self-reported m:…1919
- Only 18% of banks have fully operational, embedded AI governance processes; 92% operate at least one LLM outside formal governance frameworks; 42% lack defined kill switches; 19% have real-time model drift monitoring. Self-reported m:…1919
- 63% of banking leaders would use a black-box model if it offered better performance than an explainable alternative. Self-reported m:…1919
- Jasper says financial services firms that scale AI successfully put brand governance in place before scaling content production, because AI adoption moves the compliance bottleneck downstream rather than eliminating it. Self-reported m:…3031
- In a preprint, the Quant-Arena authors benchmark LLM agents on the full quantitative research loop — hypothesis proposal, signal implementation, backtesting. GPT-6 Astra scored highest among nine frontier models, but none beat an equal-weight market strategy in the 2024 out-of-sample test. Self-reported m:…2751
The 2031 bet
Banks are placing a dated bet on autonomy. According to the SAS/FT Longitude survey, 76% of banks expect autonomous risk and credit decisions at scale by 2031, and 25% expect to make the shift within two years. Only 10% say AI makes autonomous credit or risk decisions at scale today — so this is a five-year sprint from a standing start. Already, 82% of banking leaders say their organizations use generative AI for decision making, 41% report using agentic AI, and 59% have AI partially or widely embedded in decision-making processes. m:…1919
The concrete version of the bet already exists: Tangerine uses AI for more than 90% of decisions in its retail credit portfolios. Marilin Pikaro's warning on the findings: its ambition should match its capacity. m:…1919
The gap
The capacity is the problem. Only 18% of banks have fully operational, embedded AI governance processes today. 92% of surveyed banks operate at least one large language model outside formal governance frameworks. Only 19% have real-time AI model drift monitoring, just 36% continuously retrain or update their models, 26% lack a dedicated AI model risk policy, and 42% lack defined kill switches to deactivate models that cross preset boundaries. That is the governance posture of an industry planning to hand credit decisions to machines within five years. m:…1919
And the industry knows what it is choosing. 63% of banking leaders would be willing to use a black-box model if it offered better performance than an explainable alternative — as the survey's summary puts it, 63% would trade transparency for better performance. The implication is uncomfortable: the governance gap is not just lag, it is partly preference. m:…1919
Governance as the product
The firms closing the gap are buying governance as infrastructure, not bolting it on. Jasper's case for financial services marketing teams applies to the broader pattern: firms that succeed with AI scale by putting brand governance in place before scaling content production — approved claims, disclosures, language to avoid, brand voice, audience context — because AI adoption moves the compliance bottleneck downstream rather than eliminating it. "The constraint didn't go away. It moved downstream and got bigger." m:…3031
The alternative to governance-first is governance-never, and the survey suggests that is where most banks are drifting: LLMs operating outside formal frameworks, no kill switches, no drift monitoring. For bankers, the question is not whether autonomy arrives — 76% say it does — but whether the governance gets embedded before the first incident writes the rules. m:…1919
A sealed exam for the quant desk
Meanwhile the agents themselves are getting a sealed exam. In a preprint posted this week, the authors of Quant-Arena introduce a benchmark that evaluates LLM agents on the full quantitative research loop — hypothesis proposal, signal implementation, and backtesting. Among nine frontier models tested, GPT-6 Astra scored highest. But none of them — not the best one — outperformed an equal-weight market strategy in the 2024 out-of-sample test period. m:…2751
Put the two together and the picture sharpens: an industry willing to trade transparency for performance, planning autonomous credit decisions by 2031, whose agents cannot yet beat buy-and-hold on a sealed test. "Ambition should match its capacity" is not just a quote — it is the entire risk register in one line. m:…1919 m:…2751
What is still unresolved
- Specific regulatory enforcement actions against banks operating LLMs without governance
- The long-term impact of black-box AI models on financial stability and market liquidity
- Whether Quant-Arena's 2024 out-of-sample results generalize to other market regimes
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.