Insider Brief
- Financial institutions can use AI to improve efficiency and fraud detection, but the Financial Stability Board recommends measuring results, assigning accountability and matching controls to potential risks.
- Firms should evaluate AI’s full costs and performance, including human review, error correction and ongoing monitoring, before expanding deployment.
- AI agents require limits on system access and authority, while reliance on outside providers calls for clear contracts, assessment of shared dependencies and tested fallback plans.
Artificial intelligence can speed financial work and strengthen fraud detection, but its benefits depend on firms’ ability to measure performance, limit its authority and intervene when it fails.
For banks, insurers and investment firms, that makes AI adoption a management decision extending well beyond the choice of a model. A system that produces a credit assessment in minutes may still require substantial review. An agent that completes a task successfully may have taken unauthorized steps along the way. Buying software from several vendors may leave a firm dependent on the same underlying technology provider.
Those are among the practical implications of the Financial Stability Board’s (FSB) recent report, “Sound Practices for Responsible Adoption of Artificial Intelligence.” The international body, which coordinates the work of financial authorities and standard-setting organizations, sets out 12 practices covering how institutions select, use and oversee AI.
The guidance is nonbinding and does not establish an international standard, but has value for business leaders lies in the questions it provides for evaluating AI investments, assigning responsibility and deciding when automation has earned a larger role.
The report also cautions against treating inaction as risk-free, as institutions that fail to adopt AI could become less competitive or less capable of detecting fraud and cyberthreats.
The challenge is to identify where AI delivers a measurable improvement and establish controls proportionate to the consequences of failure. A tool that sorts internal documents and a system that influences loan decisions need different levels of scrutiny.
Measure the Finished Work
The FSB’s case studies show why speed is an incomplete measure of AI’s value.
One medium-sized regional commercial bank used generative AI to extract financial data, prepare credit narratives and support risk detection in small-business lending. Generative AI produces text and other content in response to instructions.
The bank reduced the time needed to generate a preliminary credit proposal from more than a day to 15 minutes, according to the report. Yet financial data extraction accuracy was reported as more than 80%, while credit approvers considered more than 70% of the generated content valuable reference material.
Those figures describe different aspects of performance. Faster preparation does not establish that every extracted figure is correct or that the proposal is ready for approval.
For an institution considering a similar application, this means it’s important is to measure the full process. That includes the time employees spend checking source documents, correcting errors and handling unusual cases, as well as the time saved generating a draft. A useful assessment would track the cost and quality of the completed work.
Another case illustrates how responsibilities can be divided. A global asset manager developed an AI application that reviews investment documents and prepares an initial risk assessment linked to relevant rules and source material. It does not approve or reject investments. Accountability remains with the risk manager.
That arrangement gives reviewers evidence they can check while directing more of their time toward judgment-intensive work.
The FSB cautions that its case studies illustrate reported experiences. They are not expected outcomes or benchmarks for other institutions.
It also recommends considering the total cost of AI, including computing, energy, staffing and contingency arrangements. Usage costs can rise rapidly as deployment expands.
A pilot’s business case therefore needs to account for what happens when hundreds or thousands of employees begin using it. Review capacity, support costs and the expense of maintaining a backup process belong in that calculation.
Complexity should face the same test. The report describes a bank that postponed deploying a machine-learning model in its trading business after testing found it did not consistently outperform the existing non-AI model. Its training data also failed to represent the data it would encounter in operation.
The useful decision in that case was recognizing that the proposed replacement had not demonstrated sufficient value.
Assess What AI Can Touch
An AI system’s assigned task provides only part of the picture of its risk. Its access to information and ability to act also matter.
The FSB distinguishes between an application’s importance to an institution’s operations and the harm it could cause. An agent handling a minor administrative task could still create substantial damage if it gains access to sensitive records or acts outside its authorized environment.
Agentic AI can plan and execute sequences of tasks with varying levels of independence. That makes permissions a central business issue as a system that drafts a recommendation has a different risk profile from one that can change a record, contact a customer or initiate a transaction.
The report warns that risks associated with access to tools and systems can compound as permissions expand. New access needs to be assessed alongside the authority an agent already holds.
Based on this, managers should conduct specific, rigorous reviews before expanding any deployment. What additional information can the agent retrieve? What can it change? Which actions require approval? What happens if an instruction is misunderstood?
The FSB recommends combining human oversight with technical restrictions that define what an agent can and cannot do. These may include limiting access to necessary systems, requiring approval before consequential actions and preventing a workflow from advancing when a required condition has not been met.
One life insurer’s experience shows why these controls need testing. During development, its agents could be coerced into attempting to access unauthorized databases and systems. In one case, an agent tried to delete data files, according to the report.
The insurer added controls, deployed agents with constrained autonomy and limited their access to the minimum permissions needed. Staff also tested whether subordinate agents stopped when the parent agent stopped.
The lesson extends beyond insurance. Before granting an agent greater authority, an institution needs evidence that its boundaries hold under pressure and that stopping the system halts all relevant activity.
An inventory can make that possible with institutions recording AI applications, owners, dependencies, approvals and limitations, according to FSB. Some banks with more limited adoption can use spreadsheets.
The immediate priority is reliable visibility into where AI operates, who is accountable and what each system is allowed to do. The record should also capture AI features embedded in existing software, which may arrive through vendor updates.
Make Sure Oversight Can Lead to Intervention
A human reviewer provides protection only if that person can recognize a problem and has the authority, information and time to respond.
The FSB explicitly warns against assuming that human involvement guarantees effective oversight. Employees may place excessive trust in automated outputs. Sustained reliance on AI may also weaken the expertise needed to challenge it or operate without it.
That creates a practical staffing question for organizations counting on AI to reduce manual work. Which skills must remain available to verify decisions, investigate failures and maintain service during an outage?
The report recommends preserving opportunities for employees to exercise judgment and retain relevant knowledge. It also identifies several forms of oversight, ranging from approval of individual decisions to periodic intervention and high-level control over an autonomous system’s authority.
The appropriate arrangement depends on the application. Requiring approval for every routine action may undermine the value of automation. Allowing a consequential action to proceed before anyone can review it may leave little opportunity to prevent harm.
An internationally active bank’s fraud application offers one example of this division. Its agent identifies emerging threats and proposes detection rules, but the fraud analytics team reviews and approves every new rule before implementation. The bank reported that the agent contributed to developing or updating three-quarters of its card fraud rules and helped reduce fraud losses by more than 20% in the first half of its 2026 financial year compared with the same period in 2025.
Those reported results do not establish what another bank would achieve. They illustrate a workflow in which AI accelerates analysis while people retain control over changes affecting live operations.
Oversight also requires standards established before deployment. The FSB recommends setting performance thresholds in advance where possible and monitoring trends over time.
For a document tool, that could mean evaluating whether important facts are omitted as well as whether included facts are correct. The report warns that generative AI can produce fluent, apparently complete responses while overlooking relevant information in lengthy inputs.
For an agent, evaluation must cover intermediate actions as well as the final result. Completing an assignment does not demonstrate that every step was appropriate.
The report also distinguishes recordkeeping from explanation. Logs may show what information a system accessed and what actions it took without fully explaining how it reached a decision.
For customer-facing applications, institutions need to consider whether affected people receive appropriate information and can seek an explanation, human assistance or review. The level of explanation warranted for a fraud alert may differ from that needed for a consequential lending decision.
Examine the Supplier and the Fallback
Outsourcing AI changes how an institution manages risk, but the FSB says institutions remain accountable for AI outcomes.
Vendor review therefore needs to extend beyond product features and headline performance. The report recommends contractual arrangements addressing data use, sharing and retention; material changes in AI behavior; service availability; underlying technology dependencies; and permitted or prohibited uses.
A practical procurement review would ask whether customer information can be used for training, how model changes are communicated and what support is available when performance deteriorates. It would also establish how responsibilities are divided if an autonomous agent takes an unauthorized action.
Certifications and external audits can provide supporting evidence, the FSB says, but may be insufficient on their own to satisfy an institution’s due diligence and monitoring responsibilities.
Supplier diversity also requires examination below the surface. Different applications may depend on the same cloud platform, foundation model or other infrastructure. Several contracts can therefore conceal a common point of failure.
The FSB recommends assessing these dependencies and incorporating them into continuity planning and exit strategies. Relevant scenarios include cloud disruptions, declining model performance, vendor discontinuation and infrastructure failures.
The test is whether a critical service can continue at an acceptable level when AI becomes unavailable or unreliable. That may require another model, a manual process or temporarily reduced functionality.
Smaller institutions need not reproduce the structures of a global bank. The report emphasizes proportionality and allows firms to build on existing risk, cybersecurity and supplier-management processes.
A workable starting point is to identify consequential AI uses, assign accountable owners, define acceptable performance and test the response to failure. Expansion can then follow evidence that the application delivers value and remains within its boundaries.
That approach treats reliable operation, customer outcomes and recoverability as part of the AI investment itself. Those costs help determine whether an apparently impressive demonstration can become a useful financial service.
Read the complete report here.
