AI is already making financial-security decisions. Who governs the next action?

AI is already operating close to the action boundary

Artificial intelligence is no longer merely being tested around the edges of financial services. It is already assessing transactions in milliseconds, examining hundreds of millions of payments and autonomously containing cyber threats.

This changes the governance question. The issue is no longer simply whether AI can detect financial or cyber risk. It is increasingly what the system should be authorised to do after it detects that risk.

That distinction matters because detection, recommendation and action carry very different consequences. Three public deployments illustrate how far the technology has progressed.

Mastercard: scoring transactions in less than 50 milliseconds

Mastercard says its Decision Intelligence service helps banks assess 143 billion transactions each year. Its enhanced Decision Intelligence Pro technology examines relationships between accounts, merchants, purchases and devices, with the resulting risk assessment produced in less than 50 milliseconds.

Initial Mastercard modelling reported an average 20% improvement in fraud detection, reaching as high as 300% in some circumstances. The company also reported that its analysis showed false positives falling by more than 85%. These are Mastercard's reported and modelled results, not independent PF Systems measurements.

The important operational detail is what happens next. The AI improves the risk score supplied to the bank. That score can inform a real-time decision to approve or decline the transaction.

The model provides intelligence at machine speed. The bank still needs rules determining how that intelligence affects the customer's payment. A high-risk score might justify declining the transaction, requesting additional authentication, temporarily holding it, allowing it while generating an alert or referring it for investigation.

Those are not merely modelling choices. They are decisions about authority, customer impact and financial consequence.

HSBC: checking approximately 980 million transactions every month

HSBC has publicly described using its Dynamic Risk Assessment system, co-developed with Google, to detect financial crime. The bank says it checks approximately 980 million transactions each month.

According to HSBC, the system identifies two to four times more financial crime than its previous approach, produces 60% fewer false-positive cases and has reduced analysis that previously took several weeks to a few days. These are HSBC's reported operational results.

This is a substantial deployment rather than a laboratory demonstration. It also shows why detection and action should not be treated as the same function.

AI can identify unusual activity, connect patterns and prioritise cases. But a resulting investigation may affect a customer, restrict an account or lead to information being supplied to law enforcement.

The stronger the detection system becomes, the more important it is to define which findings create an alert, which can affect a live account, which require an investigator, what evidence the investigator must see, who can authorise a restriction and how the institution records the basis for the resulting action.

Better detection does not remove the need for governed authority. It makes the handover from intelligence to action more consequential.

NKGSB Bank: autonomous containment within seconds

A more direct example comes from NKGSB Co-operative Bank in India. According to a published Darktrace customer case study, the bank deployed self-learning AI across its digital environment alongside an autonomous-response capability.

Darktrace reports that the system can move beyond producing an alert and autonomously contain malicious activity within seconds. The bank's security personnel are then notified through its threat visualisation and mobile systems.

This is much closer to the AI action boundary. Containment might be entirely appropriate when the system restricts one suspicious connection or compromised device. The consequences change if a proposed response would isolate a critical banking service, disable legitimate users or interrupt a dependency shared with another organisation.

The operational question is therefore not whether autonomous response should exist. It is which responses may proceed automatically, within what limits and under whose authority.

The NKGSB account is a vendor-published customer case study rather than an independent evaluation, but it provides a concrete example of autonomous cyber response being used within a banking environment.

These deployments operate at different distances from action

Mastercard produces a real-time transaction risk assessment, while the bank determines how that assessment becomes approval, refusal or additional verification. HSBC detects and prioritises possible financial crime, while authority remains over whether a case progresses to investigation, restriction, reporting or another intervention. NKGSB's deployment can autonomously contain activity, making the permitted scope of containment and the conditions for human intervention central operational questions.

The examples are different, but they reveal the same progression: AI observes, assesses, recommends or initiates a response, and then a consequential change becomes effective.

Governance becomes most important between the third and fourth stages. That is where a persuasive model output can become a refused payment, restricted account, isolated system or interrupted service.

The FSB is warning that the consequences may propagate

On 31 August 2026, the Financial Stability Board warned G20 finance ministers and central-bank governors that frontier AI could change the speed, scale and economics of cyber risk.

The FSB highlighted the interconnected nature of finance, including reliance on shared infrastructure, cross-border services and concentrated technology providers. A disruption beginning within one organisation could therefore affect others.

Its warning rightly emphasises vulnerability management, operational resilience, recovery and the ability to restore critical systems following a serious incident. Recovery, however, addresses what happens after disruption. Authority addresses which automated response is permitted to become effective before or during it.

As AI-enabled security systems act faster, an organisation may have less time to discover that a response was too broad, relied on incomplete information or affected the wrong dependency.

What a governed deployment could look like

Consider a financial institution that already uses an AI security product to propose containment actions. PF Systems would not need to replace the threat-detection platform, the bank's identity systems or its incident-management tooling.

A controlled integration could begin with three actions: isolate an individual endpoint, disable a service credential and restrict traffic to a payment service.

During a shadow deployment, the existing security product would continue operating through the institution's established process. PF Systems would independently evaluate copies of its proposed actions without executing them.

The governed result might allow isolation of one endpoint, narrow a network-wide block to an affected connection, require authorised review before disabling a privileged credential, refuse an action outside the approved incident plan or halt the process when the affected environment cannot be identified reliably.

The evidence would show the original proposal, the authority applied, any required approval, the governed result and the action that ultimately became effective.

This is the PF Systems proposition in operational terms: not another threat detector, fraud model or recovery platform, but a modular authority-and-evidence layer positioned between an AI-generated proposal and consequential execution.

• Is the affected asset identified reliably?
• Is the proposed response within delegated authority?
• Is its scope proportionate to the detected threat?
• Would it affect a critical or shared service?
• Is an incident commander required?
• Is essential information missing or contradictory?

A useful benchmark is disagreement, not just speed

The first shadow pilot should not attempt to prove that PF Systems makes better cyber decisions than the institution or its existing security provider. It should measure where the systems disagree.

Those measurements would give the institution evidence about whether its stated authority actually reaches automated operations. They would also expose rules that appear clear in policy documents but become ambiguous when applied to real actions.

• How often would the existing system act automatically?
• How often would organisational policy narrow that action?
• How many proposals lack sufficient context?
• Which actions consistently require additional authority?
• How often does the approved action differ from the original proposal?
• Can reviewers reconstruct why each result occurred?
• What latency does governance add to the existing workflow?

Financial resilience now includes control over machine-speed response

Mastercard demonstrates AI contributing to transaction decisions in milliseconds. HSBC demonstrates AI operating across hundreds of millions of monthly transaction checks. NKGSB Bank provides a published example of AI autonomously containing cyber activity within seconds.

The financial sector has therefore moved beyond asking whether AI can detect risk at scale. The next challenge is ensuring that speed does not silently become unlimited authority.

PF Systems does not claim to prevent fraud, stop cyberattacks, guarantee financial resilience or establish regulatory compliance. Production deployment would require appropriate technical, security, operational, legal and independent-assurance work.

Its proposition is narrower and increasingly practical: when AI proposes or initiates a consequential response, the institution should retain control over what may become effective—and preserve evidence of why.