For businesses, the limits of agentic finance depend on how reliable the results are, how much it costs to verify them, and who is accountable for the consequences. Some limitations are diminishing as large language models (LLMs) improve. Others require changes to processes, contracts, and governance structures. However, the vast majority still require human involvement.
For example, in August 2026, DBS announced the deployment of an agentic system for preparing credit assessments, used by around 1,500 employees. Specialized AI agents perform more than 70 tasks, but their output is only a draft document. Employees then review it, applying their expertise and understanding of the specific client and the circumstances of the transaction.
In short, the theoretical potential for a high degree of automation doesn’t necessarily mean it’s practical to implement. Assessing the capabilities of agentic systems requires an understanding of their key limitations. This makes it particularly important to examine what agentic finance can’t do yet.
Guarantee Reliability in Nonstandard Scenarios
Successfully completing a task doesn’t necessarily mean an AI agent will be equally reliable across repeated runs or when the initial conditions change. In its guide to evaluating AI agents, Anthropic distinguishes between a system’s ability to produce the correct result at least once and its ability to perform a task consistently across repeated attempts. This distinction is critical for financial operations because rerunning a task doesn’t undo the consequences of an earlier error.
For example, when matching a supplier invoice against a contract, an agent may handle a standard transaction without difficulty but struggle with a partial delivery, an invoice adjustment, or a previously paid advance. Assessing its reliability should therefore account for:
Consistency of results. How reliably the agent performs the same task across repeated runs.
Exception handling. Whether it can account for changes in transaction terms instead of simply extracting data from documents.
Handling uncertainty. Whether the system can recognize missing or conflicting information and escalate the decision to a specialist when it can’t safely proceed with the transaction.
Combine Complex Analysis With Tight Deadlines
Response quality, processing time, and cost need to be evaluated together. This is particularly important for tasks that require an agent to retrieve information sequentially from multiple systems, as speeding up individual computations doesn’t guarantee that the overall process will meet time expectations.
In July 2026, Microsoft researchers published the results of a benchmark study involving 251 test queries covering financial liabilities, company performance metrics, and the preparation of analytical reports. The comparison of general-purpose agentic systems found that they performed significantly worse under a 60-second time limit than when given 300 seconds and additional resources to process each query. This was because agents often needed to make multiple sequential requests to external systems to obtain the necessary information and couldn’t always complete them within 1 minute.
For businesses, this means evaluating AI agents against actual requirements for speed and quality. In some cases, it may therefore be more efficient to assign an agent to verify specific metrics in a company’s financial statements than to delegate the entire assessment of its financial condition, which entails the additional costs of lengthy analysis and exception handling.
Verify Transaction Completion Through Payment Confirmation
Even a properly authorized and executed payment doesn’t guarantee that a purchase achieved its intended purpose. For example, an AI agent may purchase a dataset from an approved supplier, stay within the allocated budget, and receive payment confirmation, only to discover later that the information provided doesn’t meet the order requirements.
This distinction is reflected in the study Authority–Inference Separation in Agentic Finance, published in August 2026. Using public records from the Base blockchain, the authors demonstrate that transaction data can confirm that a payment was settled but doesn’t, on its own, prove that the supplier fulfilled its obligations to the buyer.
Fully automating a transaction therefore requires a separate verification that the delivered product or service meets the agreed terms. An AI agent can participate in this process, identify discrepancies, and initiate a claim. However, successful payment processing alone doesn’t guarantee that the counterparty fulfilled its obligations.
Guarantee ROI Through Automation Alone
The economic impact of deploying agentic systems can’t be measured solely by the number of labor hours saved, as reviewing and approving AI agents’ decisions also takes employees’ time.
When evaluating an agentic AI project, businesses should consider:
Cost of running models and accessing external data
Time spent reviewing results and handling exceptions
Costs of integration, maintenance, and repeated testing
Costs of correcting errors and restoring operations
A useful metric in this context is the total resources consumed per successfully completed task, including all verification steps. This should be compared with the performance of existing business processes, including automation methods already in use.
For routine financial calculations performed according to predefined rules, deploying agentic systems may be unnecessary. AI agents offer greater economic potential in processes that require analyzing diverse types of information and independently determining the sequence of actions.
The difference between operational efficiency and financial returns is illustrated by Deloitte’s State of AI in the Enterprise 2026 report, based on a survey of 3,235 executives across 24 countries conducted between August and September 2025. Among respondents, 66% reported improvements in productivity and efficiency, 40% reported cost reductions, and 20% reported revenue growth. It’s worth noting that the survey covered enterprise use of AI solutions in general.
Maintain Predictability After Updates
Successfully testing an AI agent doesn’t guarantee that it will maintain the same level of reliability after a model update, the integration of new tools, or changes to data sources. Financial institutions therefore need to monitor how these systems perform in practice and document changes that could affect their behavior.
In its 2026 Annual Regulatory Oversight Report, the Financial Industry Regulatory Authority (FINRA) highlights the challenges of auditing multi-step agentic decisions and recommends retaining records of model versions, prompts, agent actions, and their outcomes. This information makes it possible to reconstruct the sequence of operations and determine the conditions under which the system made its decisions.
A more systematic approach is proposed in the study Agent Reliability Profiles in Financial Services, published in October 2026. The authors propose assessing each agent’s reliability based on its specific operating conditions, level of autonomy, permitted actions, and established restrictions. The approach remains a research concept, but it supports the conclusion that test results establish a system’s reliability only under specific conditions, not across every possible use case.
For businesses, this means reassessing an AI agent’s reliability after significant changes to its configuration. Applying previous test results to an updated system or other financial processes could lead to underestimated risks and flawed decisions about the appropriate level of autonomy.
Relieve Financial Institutions of Responsibility
Using AI agents to execute financial transactions doesn’t relieve businesses of their responsibility to comply with applicable regulatory requirements. FINRA’s 2026 report explicitly states that the use of generative AI doesn’t eliminate firms’ obligations regarding supervision, recordkeeping, communications, and fair business practices.
FINRA also clarifies that outsourcing a regulated function to a third-party provider doesn’t relieve a firm of its responsibility to comply with applicable requirements. Consequently, commercial use of an agentic system doesn’t, in itself, transfer regulatory obligations to its developer. The allocation of contractual liability for specific damages remains a separate issue.
Effective oversight of AI agents requires clearly defined employee authority and responsibilities beyond simply approving transactions. Employees must have access to the necessary information, the ability to suspend operations, and the means to address the consequences of errors. Otherwise, human oversight may become a mere formality.
The future development of agentic finance will depend on businesses’ ability to ensure the reliability of autonomous operations and manage the associated risks. The level of autonomy granted to AI agents should therefore be determined by demonstrated results in specific financial processes, not by the technology’s potential capabilities.
