AI and advanced analytics are reshaping how federal financial agencies approach regulation, compliance, and risk management. As agencies move from reactive processes to more predictive approaches, they have new opportunities to identify emerging issues earlier, strengthen decision-making, and improve service delivery.
MeriTalk recently sat down with Le Tran, managing director at Maximus Federal, to discuss how agencies can adopt AI responsibly while maintaining trust, transparency, and accountability. Tran discussed the role of shared services, explainable models, workforce readiness, and human oversight in helping federal financial agencies use AI responsibly to support mission outcomes.
MeriTalk: How can shared or common service models help agencies and financial institutions accelerate AI adoption while still meeting rigorous testing, validation, and governance requirements? What makes a centralized AI capability effective in practice?
Tran: Shared service platforms can help agencies consolidate validation and testing, reduce redundant work, and centralize the tools and infrastructure they need. That is especially important in financial regulatory environments, where agencies are responding to hundreds of regulatory changes each year while also managing increased workloads and smaller federal IT teams.
There is a real opportunity to use centralized AI-enabled solutions to improve efficiency, but agencies also have to address common blockers. A lot of times, they want to adopt AI but run into long procurement timelines, the authority to operate (ATO) process, and security concerns. FedRAMP AI solutions and ATO-approved tools can help speed adoption because they give agencies a more trusted path to begin using AI.
I am seeing agencies make progress. One agency we support has been effective in using ATO-approved tools, centralizing its data, and building a repository of AI use cases. These steps have made it possible for staff to begin using approved tools in their day-to-day work, and the agency can measure proof-of-concept outcomes before scaling the use cases that improve productivity or service delivery the most.
Centralized implementations can support machine learning for fraud detection and risk management, natural language processing to pull insights from regulatory documents, and predictive analytics once the right data foundation is in place. The NIST AI Risk Management Framework is also a helpful reference as agencies build safe, scalable infrastructure.
MeriTalk: In financial regulation and compliance, how should agencies balance model performance with explainability? And what practical steps can they take to identify and reduce bias when fairness itself can be defined in different ways?
Tran: Clear AI explanations are essential because agencies need to show how a model arrives at a decision. In high-stakes environments, performance matters, but regulators, customers, and staff also need to understand what data informed the output and why the model reached that result.
Agencies also need visibility into how models are behaving. That means actively monitoring AI systems, looking for patterns in the training data, and making adjustments when something is not working as intended.
AI cannot operate by itself. Human-in-the-loop review at key points helps validate data, reduce bias, and improve models over time. That review should include people from different backgrounds so the model is tested from multiple perspectives.
Accountability frameworks also help agencies address fairness. Fairness can be defined in different ways depending on the use case, so those definitions need to be tied to legal expectations and mission requirements. For example, if the use case involves financial compliance or fraud detection, the questions subject matter experts ask during review need to be specific to that environment. The oversight has to be tailored to the work being performed.
MeriTalk: As AI takes on more decision-support activities and potentially autonomous functions, how should regulators and institutions determine which use cases are truly high risk? And what does meaningful human oversight look like at scale?
Tran: Agencies should start by defining measurable risk criteria. That includes questions such as: What is the potential financial impact on the customer? What is the transaction or data volume? How much autonomy does the AI system have? Those criteria need to be defined at the beginning of the project, along with how the risks will be measured.
From there, agencies need a classification framework to determine which applications cross the high-risk threshold, how they should be assessed, and what level of human oversight is appropriate. That work requires cross-functional collaboration across technical, regulatory, business, and mission teams.
The goal is to protect customers while making AI use, scoring, and oversight transparent.
MeriTalk: Looking ahead five years, which compliance functions are most likely to be automated, and where will human judgment remain essential? As more institutions deploy AI tools, how should leaders think about systemic risk and stress testing?
Tran: AI can really help with large volumes of data and repetitive compliance functions. That includes auditing, processing reviews of different types of data, and real-time monitoring. Whether the data is on paper, in a database, or in another format, high-volume reviews are where AI can reduce the manual burden.
As agencies get better at that, predictive models will start to shift the work from detection to prevention. Instead of only finding problems after they occur, agencies can begin identifying patterns earlier and using that information to prevent issues.
But human judgment will always be essential. Humans need to define the problems that AI can solve. AI is not going to do that on its own. Right now, many agencies are still focused on the basics, such as getting their data organized so models can be trained. As people provide more judgment, review, and oversight, that data will be used more effectively.
Stress testing will also be important as AI use increases. Agencies need to make sure their systems can accommodate the volume and complexity of AI-driven requirements, and leaders need to understand how tools behave across institutions.
That is also why workforce programs matter. Once people understand the tools, they can better determine which tasks can be done by AI, which require humans and AI together, and which remain uniquely human.
MeriTalk: What organizational capabilities are most important for successful AI adoption in federal financial agencies, from data infrastructure and talent to governance of third-party models and tools?
Tran: Data infrastructure quality is the foundation. Federal financial agencies manage large volumes of highly sensitive data, so the infrastructure has to be sound, secure, and able to protect privacy.
Once that foundation is in place, transparency becomes very important. Agencies need to build trust across regulatory staff and the public. If people understand how AI is being used, and agencies can make that usage publicly available when possible, there is less fear or concern about what AI is doing.
The teams also need the right mix of expertise. Technical staff are important, and regulatory and business experts need to be involved as well. They provide checks and balances and often serve as part of the human-in-the-loop process.
Workforce readiness is another critical capability. Staff need training, and leaders need enough AI literacy to explain what they want AI to do so that mission goals are reflected in IT roadmaps and operational plans. Secure infrastructure, trained people, transparency, and strong oversight all help build trust as AI becomes more widely used across federal agencies.