For commercial banking analysts, the job description has long masked a fundamental tension.
According to nCino, the title suggests analysis, but a significant portion of the working day is consumed by something else entirely: pulling data from disconnected systems, reconciling inconsistencies, and reformatting information into something workable before any real interpretation can begin.
nCino recently discussed how agentic AI takes the manual data assembly, returning analysts to credit judgment.
According to McKinsey, banking operations staff spend roughly 80% of their time on coordination and rule-based tasks such as collating information and drafting credit memos. Agentic AI is beginning to shift that equation.
Unlike conventional generative AI, agentic AI can carry out multi-step workflows autonomously, retrieving data from multiple sources, applying logic, generating structured outputs, and flagging exceptions, all without requiring a human prompt at each stage. This represents a meaningful step beyond both legacy chatbots and the rule-based automation that banks have experimented with in recent years.
A different kind of automation
Those who witnessed the last major automation cycle have understandable reason to approach this sceptically. Robotic process automation (RPA) promised to relieve analysts of time-consuming manual processes but consistently fell short. Its brittle, script-based architecture struggled whenever variability entered the picture, and in commercial banking, variability is the norm rather than the exception.
Agentic AI operates differently. Where RPA followed rigid instructions that broke down at the first unexpected input, an agentic system pursues an objective across a dynamic, multi-step workflow, adapting as conditions change. When a borrower’s tax return arrives in an unfamiliar format, RPA halts. Agentic AI adjusts and continues. McKinsey has cited this shift as one that “appears to be for real,” noting that early use cases have already reduced manual workloads by between 30% and 50%.
Three workflows in particular stand to benefit most: credit review preparation, covenant monitoring, and portfolio risk trending. All three involve variable inputs, multiple systems, and exceptions that require judgement. They are also precisely the workflows where RPA most frequently failed.
Credit review prep: from hours to minutes
A standard commercial credit review typically draws on at least four systems: the core banking platform, the loan origination system, a covenant tracker, and the CRM. Before any analysis can begin, an analyst may spend several hours logging into each system, running separate queries, exporting and re-keying data, and reconciling conflicts. Only once that groundwork is complete does the actual analytical work start.
With agentic AI handling the preparation stage, that process compresses dramatically. The system retrieves current data across source systems, normalises formats, surfaces conflicts, and delivers a structured document with exceptions already flagged. The analyst opens the package and moves directly to interpretation.
McKinsey’s QuantumBlack team documented a potential productivity improvement of more than 60% from an AI-assisted credit memo system drawing on at least ten data sources, with projected annual savings exceeding $3m. Those savings derived from eliminating manual data gathering, not from any improvement in analytical quality. The critical caveat, however, is that the efficiency gains depend on how unified the underlying data architecture already is. When data remains scattered across disconnected systems, the agent inherits the same fragmentation the analyst faces, and the assembly work recedes into the background rather than disappearing.
Covenant monitoring: exception-led rather than calendar-led
Covenant testing has historically been a calendar-driven process for practical reasons. Constructing the data infrastructure to test a portfolio of covenants is a multi-hour exercise, making quarterly or semi-annual cycles the norm. Continuous monitoring is preferable in principle, but the manual setup required has made it unworkable at scale.
Agentic AI makes continuous monitoring the default. The same testing logic that would previously have been applied once a quarter runs continuously in the background. When a threshold is breached, the analyst receives a prioritised alert with the relevant data already assembled: borrower trends, recent financial filings, comparable history, and the relationship context that frames the breach.
The analyst’s role shifts from running the test to interpreting its output. A breach can indicate a data quality issue, a one-time accounting adjustment, or genuine credit deterioration, and distinguishing between them requires borrower knowledge and credit experience that no system can replicate.
Portfolio risk trending: real-time rather than retrospective
Quarter-end portfolio reviews carry an inherent lag. By the time the review is complete, analysts are reading the recent past rather than monitoring what is developing in real time. Stage duration anomalies, concentration shifts, early delinquency signals, and unusual draw patterns may have been building for weeks before they surface in a periodic review cycle.
Agentic AI replaces that rhythm with continuous pattern detection, surfacing signals as they develop with supporting context attached. Peer benchmarking adds a further dimension, enabling analysts to assess whether a shift reflects a broader market trend or something specific to their institution, context that can sharpen the interpretation of any individual signal.
The limits of the technology
The efficiency gains outlined above scale directly with how unified the underlying commercial data already is. When data is consolidated within a single platform, agentic AI retrieves once, normalises once, and delivers a complete picture. The more fragmented the architecture, the more the assembly work is redistributed rather than eliminated, and the smaller the practical time saving.
This makes data consolidation the foundational question, both for institutions evaluating these technologies and for analysts assessing vendor claims. The relevant question in any procurement conversation is how much integration work occurs before the AI starts, and how much the AI itself is expected to absorb. The answer reveals whether a platform genuinely compresses hours or merely moves them elsewhere.
What remains with the analyst
Agentic AI handles the assembly and surfaces the signal. The work that remains belongs to the analyst, and it is not incidental.
Exception interpretation is the most immediate. A flagged covenant breach or an outlier risk score requires a judgement call that draws on relationship context and credit experience. Examiner accountability is the most consequential. When a regulator asks why a risk rating moved, the analyst who reviewed the output and made the determination is on the record. That responsibility does not transfer to the system.
The third capability is harder to name but easier to recognise: the experience-derived instinct to identify when an AI-generated output is simply wrong. An analyst who has built hundreds of credit memos understands how a well-constructed one reads and where a borrower’s story tends to break down. When AI handles assembly, that expertise becomes more valuable, not less.
Across credit review prep, covenant monitoring, and portfolio risk trending, agentic AI changes how the working day is structured. Monday morning still begins with a credit review. The difference is that the package arrives already assembled, exceptions flagged, and the hours previously spent building it are now available for interpreting it.
Read the full nCino post here.
Copyright © 2026 FinTech Global
