When AI Starts Acting, Financial Control Has to Become More Visible
For the past few years, most conversations about artificial intelligence in finance have focused on what AI can understand.
Can it detect fraud? Can it assess credit risk? Can it summarize financial information, identify unusual activity, or help a company make better decisions?
This week, the conversation felt different.
The more important question is becoming: what happens when AI does not simply analyze financial activity, but begins to participate in it?
Financial institutions are already using AI to support fraud detection, compliance, customer service, credit assessment, cybersecurity, and risk management. At the same time, agentic systems are becoming more capable of completing multi-step tasks, accessing business systems, and acting with less direct human involvement.
That shift creates real opportunity. It can reduce repetitive work, shorten response times, and help smaller teams manage information that would otherwise require much larger operations.
But it also changes the meaning of financial control.
When a person prepares a payment, approves an expense, or changes a financial record, a company can usually identify who made the decision. When an AI system participates in the same workflow, responsibility may become less visible. The action might involve a model, an application, an external data provider, a cloud platform, and a human reviewer.
If something goes wrong, knowing that “the AI did it” is not an adequate explanation.
Recent discussions from the Bank for International Settlements emphasize that financial supervision can no longer focus only on whether banks are using AI responsibly. Supervisors also need to consider whether institutions remain operationally and strategically resilient in an economy increasingly shaped by AI.
This distinction matters.
Model governance asks whether an AI system is accurate, explainable, and appropriately monitored. Operational resilience asks what happens when that system fails, becomes unavailable, receives incorrect information, or produces actions faster than people can review them.
For small businesses, the same issue appears at a different scale.
A small company may use connected tools for bookkeeping, payments, customer communication, expense management, research, and reporting. Adding AI can make those tools more useful, but it can also make the relationships between them harder to see.
A financial assistant might identify an invoice, recommend payment, categorize the expense, update a forecast, and notify the business owner. Each step may appear reasonable on its own. But the company still needs to know where recommendation ends and authorization begins.
This is why AI readiness is not only a technology question. It is an operating-design question.
Before allowing an AI system to take financial action, a business should be able to answer several basic questions:
What information can the system access?
Which actions can it recommend?
Which actions can it complete?
What requires human approval?
How are unusual actions identified?
Can the company reconstruct what happened afterward?
What happens if the system or its provider becomes unavailable?
These questions may sound procedural, but they are becoming part of the infrastructure of trust.
FINOS highlighted similar priorities in its recent work on AI and open source in financial services. As agentic AI moves toward production, financial organizations need common approaches to identity, authorization, provenance, orchestration, and secure access to enterprise systems.
In practical terms, an AI agent needs its own clearly defined role.
It should not receive broad access simply because broad access is convenient. Its permissions should reflect the specific job it is expected to perform. Its actions should be recorded. Spending or transaction limits should be visible. Higher-risk decisions should move through an approval process that people can understand.
This is particularly important because AI is also changing the speed of cyber risk. Frontier models can help defenders discover vulnerabilities and respond to incidents, but they can also reduce the time between finding a weakness and exploiting it.
When response windows become shorter, companies cannot depend entirely on improvised decisions. They need clear escalation paths, recovery procedures, access controls, and reliable records before an incident occurs.
At ToNoisy LLC, I see a practical role for business analytics in this transition.
A useful dashboard should not only report revenue, expenses, or cash flow. It should also make automated activity visible. A company may need to track which actions were suggested by AI, which were approved by a person, which were completed automatically, and which were stopped for review.
For an early-stage business, this does not require an enormous governance program. It can begin with a simple approval table, clear access levels, transaction thresholds, exception alerts, and an activity log.
The goal is not to place a person inside every automated step. That would remove much of the value of automation. The goal is to ensure that human judgment is positioned where consequences become meaningful.
AI in finance will probably become more autonomous. The question is whether business controls will become clearer at the same time.
The companies that benefit most may not be the ones that automate the greatest number of tasks. They may be the ones that can explain what their systems are allowed to do, recognize when something unusual happens, and intervene before a small mistake becomes a financial problem.
As AI begins to act, trust will depend on more than intelligence.
It will depend on visible boundaries, traceable decisions, and clearly assigned responsibility.
Main Resources
Bank for International Settlements, “Supervising banks in an AI-shaped economy,” delivered September 18, 2026.
https://www.bis.org/speeches/20260918-supervising-banks-ai-shaped-economyFINOS, “AI and Open Source in Financial Services,” covering machine-scale cyber risk, agentic workflows, and common industry standards.
https://www.finos.org/2026-ai-open-source-in-finance-reportBank for International Settlements, “When machines attack: frontier AI cyber threats and policy responses in the financial sector,” published September 9, 2026.
https://www.bis.org/publications/fsi-paper-28-when-machines-attack-frontier-ai-cyber-threats-and-policy-responses-financial-sectorEuropean Central Bank, “A new age of capital: growth, sovereignty and AI,” delivered September 14, 2026.
https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260914_2~a3f0efbee4.en.html