Human oversight in Agentic AI: Building robust and auditable enterprise workflows

Introduction

The integration of GenAI solutions into enterprise environments is progressing rapidly, with AI adoption widely recognized as essential for maintaining competitiveness. GenAI Agents signals the next step in AI adoption, but as we’ve detailed in previous reports[1], [2], [3], there are several necessary components to make Agents meet enterprise expectations. At Moody’s, we believe that efficient enterprise AI relies on three foundational pillars: quality data, strict but justified guardrails, and continuous human intervention (“human in the loop”).

In this article, we’ll examine the rationale for maintaining humans in the loop within Agentic workflows, underscore the limitations of unchecked automation, and advocate for the adoption of streamlined tools that facilitate rapid and auditable checks. Drawing on the experience of Moody’s and the distinctive capabilities of Research Assistant, we’ll explore how cross-check mechanisms serve as essential safeguards, helping enable both accuracy and transparency of AI-driven results.

The need for human oversight

Agentic solutions within enterprise environments offer substantial advantages in processing large datasets, identifying and aggregating relevant information, and generating outputs that would otherwise require significant manual effort. However, these systems frequently exhibit significant variability in their output, whereby users or internal applications submit inputs and receive results, yet have limited visibility or control over the processes occurring between these stages and datasets used to produce them. In high-stakes environments such as underwriting, the consequences of erroneous or biased outputs are simply not acceptable, impacting not only financial outcomes but also regulatory compliance and institutional reputation.

Human judgment serves as a critical counterbalance to these risks. Experienced professionals possess domain proficiency, advanced reasoning skills, and contextual awareness that AI agents cannot fully replicate. For example, Agents are prone to treating information from different sources with similar importance which may lead to ambiguous or even misleading results if the data in these sources is contradictory. In contrast, human analysts can challenge assumptions, and flag ambiguities, adding clarity to the Agentic output. Maintaining a human presence in the loop is therefore not merely a precaution, but an operational necessity.

Risks of unchecked automation

Due to the probabilistic nature of Agents, fully agentic workflows can pose substantial risks, and inadvertently introduce vulnerabilities into enterprise processes. We know that relaxed data requirements, where information is unstructured, contradictory, or limited, will likely result in outputs that fall short of industry standards. But even well-defined rules cannot anticipate the diverse range of situations a real business might face; it is almost inevitable that there will be moments when guardrails fail to hold a model within its intended context. When this happens, an Agent’s output can collapse into generic text with little factual substance. In these situations, human intervention becomes the critical safeguard, able to catch and correct the errors that an automated system will miss.

Moody's Research Assistant: enhancing oversight and efficiency

Keeping a human in the loop is an integral part of AI adoption; therefore, the next step is to enhance the tools that support a human analyst checks and corrections of an Agent’s output. Two situations are most likely:

• A human analyst is to check the Agent’s result which looks correct but either the workflow is so critical that all the data must be double checked, or the result seems correct, but some data points look strange or contradicting.
• A human analyst sees that the Agent’s result is too generic and lacks the required information either in a particular section of the report, or across the whole report.

Moody’s Research Assistant exemplifies a best-in-class approach to integrating AI within Agentic workflows. Designed for use by professionals in credit risk management, lending, and asset management, it provides vast capabilities for validating Agentic AI outputs. Research Assistant responses are based on Moody’s highly reliable and relevant data estate, helping practitioners check and compare the results delivered by Agents against a verified source. Moody’s GenAI solutions always provide citations, which may add an additional layer of confidence to the result. If the Agentic result fails to meet requirements in one or multiple sections of the report, Research Assistant can help fill the gap and deliver missing information. Importantly, Research Assistant can be used in tandem with any Agentic solution, whether internally built or vendor-provided. In this framework, Moody’s GenAI tool serves as a secondary line of defense, empowering users to corroborate the results of their Agentic solutions and utilize multi-source verification, drawing insights from a comprehensive array of Moody’s proprietary datasets. This process not only enhances confidence in decision-making but also produces detailed audit trails that can be reviewed by both internal and external stakeholders.

Augmented intelligence: a real life example

One of the most compelling use cases in the financial sector is credit risk assessment. In this industry, reports must always be aligned with enterprise requirements while covering multiple dimensions of counterparty credit risk. Credit assessment must usually include sector research, financial analysis, and peer review, as these are the key factors which define overall credit risk.

Sector review and peer analysis are particularly challenging for Agents. A target company may operate in several sectors, it may be a holding company where sector attribution may be ambiguous, or a sector may be so narrow that the available information is very scarce and not enough to make any meaningful judgement. The situation with peers might be even more complicated: an Agent needs to not only correctly choose several competitors from the same sector but also have enough qualitative and financial data for such competitors to make peer review valuable.

In both these cases human intervention is highly impactful. A human analyst will use his own judgement for sector mapping. With the help of a GenAI solution like Research Assistant, he can then guide an Agent to the right direction or completely rewrite the sector analysis part. The same applies to peer review: human assessment in picking correct peers may dramatically improve the quality of the result, and using Moody’s Research Assistant might be very handy to narrow down the list of peers, check data availability, and add necessary details to the peer review process.

This synergistic approach ensures that AI remains an assistant to, rather than a replacement for, a human analyst. It fosters transparent, adaptive workflows that align with organizational values and support long-term success.

The three pillars of enterprise AI excellence

Efficient and robust AI adoption in enterprise settings is built on three key pillars:

• Quality Data: Robust, relevant, and accurate data underpins all effective AI-driven processes. High-quality datasets minimize the risk of misinterpretation and bias, providing a trustworthy basis for decision-making. This may include context engineering, or embedding domain-specific knowledge, regulatory requirements, and business priorities into AI workflows.
• AI guardrails: a set of rules that keep Agents within pre-defined data boundaries, ensure auditability, trace sources, and highlight where human intervention is necessary.
• Human-in-the-loop: having a proficient analyst equipped with the skills and tools needed to understand, check, and effectively utilize Agentic output is essential.

Moody’s approach addresses all three pillars, offering curated datasets, context-aware engineering, and user-centric assessment through its Research Assistant platform. In doing so, it delivers an additional layer of protection and efficiency for professionals adopting Agentic workflows.

Conclusion

As organizations increasingly integrate Agentic solutions into their operations, the imperative to keep humans in the loop becomes ever more pronounced. In critical domains such as underwriting and risk assessment, human oversight is essential to ensuring the accuracy, fairness, and accountability of AI-generated outputs. The adoption of streamlined, auditable tools – such as Moody’s Research Assistant – provides the transparency, control, and cross-verification needed to navigate complex workflows with confidence.

The future of enterprise AI is rooted not in the pursuit of complete automation, but in cultivating intelligent partnerships between human professionals and advanced technologies. By maintaining robust oversight and equipping professionals with advanced yet accessible tools, organizations can unlock the full value of AI while safeguarding against risk, ensuring compliance, and supporting ongoing innovation. Moody’s Research Assistant stands as a testament to the power of this approach, offering a practical, effective solution for enterprises embracing the promise and responsibility of Agentic AI.

About the author:

Dmitriy Turmyshev serves as Associate Director – Solutions Specialist at Moody’s, where he supports initiatives focused on Moody’s GenAI product suite. In this capacity, he partners with financial institutions, asset managers, and large corporates to optimize the integration of Moody’s data, content, and technology into credit-related workflows. Dmitriy’s role centers on enabling clients to incorporate advanced GenAI solutions into their risk assessment processes, driving efficiency and accuracy.

Prior to joining Moody’s, Dmitriy accumulated over 13 years of experience in credit research across a broad spectrum of fixed income asset classes, including bonds, syndicated loans, and private credit transactions. This extensive background has equipped him with deep analytical expertise and a comprehensive understanding of credit markets and risk evaluation.