What a global survey of 300 banking leaders reveals about AI, performance, and the accountability gap.
Sixty-three percent.
That is the share of banking leaders — 300 of them, across 31 countries, in new research by FT Longitude, SAS and Microsoft — who said they would use a black-box AI model, one with little or no transparency into how it decides, if it delivered better performance than an explainable alternative.
Read that again. Not “would tolerate.” Would choose. Knowingly.
Now place that number next to the others in the same report. Only 18% of these institutions have fully operational, embedded AI governance today. Forty-two percent have no defined kill switch to deactivate a model that crosses its boundaries. Nineteen percent monitor model drift in real time. And 92% acknowledge that at least one large language model is already operating inside their institution outside formal governance frameworks.
The decisions are being delegated faster than the accountability for them.
This is not a technology problem
A decision that cannot be explained cannot be owned.
When a client asks, “Why was my credit refused?” someone has to answer. “The model decided” is not an answer — it is an abdication dressed up as innovation. Explainability is not a feature for data scientists. It is the mechanism that connects a decision to a person who can be questioned, corrected, and held responsible. Remove it, and responsibility does not transfer to the machine. It evaporates.
I have watched this pattern before, in a slower form. For two decades, the industry optimized what could be counted — targets, volumes, quarterly numbers — and discovered, repeatedly and expensively, what happens to trust when the human being in front of the client can no longer explain or defend what the system demands. The KPI era taught front-line employees to serve the metric. The AI era risks teaching entire institutions the same lesson, at machine speed.
Performance is measurable this quarter. Accountability only becomes visible on the worst day — the day an unexplainable model makes a consequential mistake, and no one in the building can say why.
The line is not “never” — it is “proportional”
Opacity is not always wrong. Fraud detection does not owe anyone an essay. Pattern recognition on low-stakes, reversible decisions can be delegated with a clear conscience.
The ethical failure begins where the stakes become personal and the decision becomes contestable: a refused loan, a closed account, a flagged customer. Those decisions deserve a human being who can answer for them.
One Canadian institution in the report describes the balance plainly. At Tangerine, more than 90% of retail credit decisions are now AI-based, with very little human involvement — but for high-risk portfolios, the bank has deliberately kept the human in the loop “until we have more confidence in the technology.” Autonomy proportional to the stakes. Explainability proportional to the consequences. That is the line.
What “answering for it” actually requires
Governance is not a committee and a PDF. It is a short list of unglamorous disciplines:
- An inventory. You cannot govern models you have not listed. When 92% of institutions find models running outside their frameworks, the first problem is not ethics — it is eyesight.
- A named owner. Every consequential model should have one human being whose job is to answer for it.
- A kill switch, and the nerve to use it. Forty-two percent of institutions do not have one. That is not a technology gap; it is a decision, made in advance, to let the machine run.
- Monitoring that matches the autonomy. The more independent the model, the closer the watch — yet only 19% monitor drift in real time.
- Governance at the speed of adoption. The report’s own title says it: Why AI at Scale Needs Governance at Speed.
None of this slows a serious institution down. As one European risk chief puts it in the report, governance means you have done your homework, created the boundaries, set the rules — and then you are ready to run.
The bill always arrives
Three in four banking leaders expect AI to be making autonomous credit and risk decisions at scale within five years. That future is coming regardless of what any of us write about it.
The only open question is whether, when the first unexplainable decision causes real harm — and it will — the institution can still do the one thing banking has always run on: look a client, a regulator, or a court in the eye and explain itself.
Performance is reported quarterly. Trust compounds for decades, and can be spent in a single decision nobody can explain.
A decision nobody can explain is a decision nobody owns. And in banking, unowned decisions are how good institutions end up in bad headlines.
Jean-Philippe St-Jacques is a financial services executive and Partner at 5 Mindset Consulting. He writes on leadership, ethics and AI governance.
Source: FT Longitude, SAS and Microsoft, “Why AI at Scale Needs Governance at Speed” — global survey of 300 banking leaders across 31 countries (reported October 27, 2026).

