Where AI fits in the workflow
A practical framework for separating tasks AI can accelerate from tasks that require a human accountability moment.
- Classify credit review tasks into high fit, conditional fit, and off-limits
- Explain why covenant extraction is a better AI candidate than final covenant compliance approval
- Identify human-only accountability moments in the credit lifecycle
Not every part of a credit review is a good candidate for AI assistance. Some tasks are perfect for the machine’s ability to organize text; others require human judgment, institutional authority, or accountability that a machine cannot hold.
This module provides a simple framework for deciding where AI fits in your workflow.
The Fit Framework
We classify tasks into three categories:
- High fit (AI can assist): The work is production-oriented and directly checkable against source.
- Conditional fit (AI can assist under named controls): The output contributes to judgment, requires sequencing or enhanced verification, or could anchor the reviewer.
- Off-limits (Human-only accountability moment): The task’s value lies in the authorized professional or committee owning the conclusion.
High Fit: Production and Organization
High-fit tasks are those where the machine does the heavy lifting of reading, extracting, or organizing, but the result is easy for a human to verify against the original document.
Examples:
- Extracting maturity dates, interest rates, or reporting requirements from a loan agreement
- Preparing an initial spread from a clean financial statement
- Summarizing the stated purpose of a loan from a credit memo
Why it fits: The machine is excellent at finding and formatting information. If it makes a mistake, you can spot it immediately by checking the source document. The verification is straightforward.
Conditional Fit: Contributing to Judgment
Conditional-fit tasks involve some level of interpretation or synthesis. The machine can help, but you must apply specific controls to ensure you don't accidentally delegate your judgment.
Examples:
- Flagging potential add-backs in a financial statement
- Drafting a preliminary issue statement based on a policy violation
- Identifying factors that might influence a risk rating change
Why it's conditional: The machine can surface patterns or draft language, but it doesn't understand your institution's risk appetite or policy nuances. If you accept its draft without independent thought, you've anchored yourself to its (potentially flawed) conclusion.
The control: Always form your own preliminary view before reading the machine's draft, and independently verify any logic or calculation it presents.
Off-Limits: Human Accountability Moments
These are the tasks where the institution requires a named, authorized human to take responsibility for the outcome. AI cannot hold this responsibility.
Examples:
- Assigning the final risk rating
- Approving the severity level of an issue
- Signing off on the final credit review report
- Concluding whether a borrower is in compliance with a covenant
Why it's off-limits: The value of these tasks isn't in the drafting; it's in the ownership. A regulator or internal auditor needs to know which human stands behind the decision. A machine cannot be held accountable, so it cannot perform these tasks.
The Simple Rule
AI may be responsible for producing a draft artifact. AI may never be accountable for a credit conclusion.
Without scrolling back to the map: for each step, is the machine a high fit, a conditional fit, or off-limits?
| Lifecycle step | Fit | Why the line is here |
|---|---|---|
| ScopingScoping and portfolio risk assessment | Conditional fit | The machine assembles migration trends, concentrations, and prior findings fast. Where the risk actually is — and what the review must be able to say at the end — is a judgment about your institution that leadership owns. |
| SamplingSampling methodology design | Conditional fit | It can propose stratifications and compute coverage. The methodology is what makes your coverage claim defensible, so the design rationale must be yours, written down, in your words. |
| SamplingSample selection against approved criteria | High fit | Mechanical execution of an approved methodology. Verify the pull matches the criteria before locking — a selection error contaminates everything downstream. |
| IntakeDocument intake, OCR, and data extraction | High fit | The highest-volume, highest-fit task — and the biggest failure surface. OCR transposes digits, merges columns, and drops negative signs, so extracted figures carry a verification tier, not trust. |
| AnalysisFinancial statement spreading | High fit | Structured, checkable, source-anchored work. Spreads feed ratings, so every machine-produced figure gets independently recomputed — recomputed, not sampled. |
| AnalysisEBITDA bridge and quality-of-earnings flags | Conditional fit | It assembles the bridge and surfaces add-back candidates at volume. Whether an add-back is credible is a judgment; the machine will carry an aggressive one forward without blinking. |
| AnalysisCovenant compliance testing | High fit | Computation is high fit; the trap is definitions. Amendments, cure periods, and defined terms bite — verify the definition and the period, then the arithmetic. |
| AnalysisDSCR / LTV / borrowing base recalculation | High fit | Deterministic arithmetic over extracted inputs. Fit is high precisely because verification is cheap: recompute from source and compare. |
| AnalysisCollateral and valuation review | Conditional fit | Summarizing appraisals is fine. Adequacy — staleness, method, market — is judgment, and a fluent summary of a three-year-old appraisal reads deceptively current. |
| ConclusionRisk rating recommendation | Conditional fit | The machine may present rating-relevant factors — after you have formed your own view. Sequence matters: a suggestion seen first anchors the human who sees it. |
| ConclusionAssigning the final risk rating | Off-limits | The rating is the review’s conclusion. A named human assigns it and answers for it. A number nobody owns is worthless in an examination. |
| ConclusionAccrual status / classification decisions | Off-limits | Direct accounting and regulatory reporting consequences. Made from source, by accountable humans, every time. |
| ConclusionConcluding on ACL / allowance adequacy | Off-limits | An institutional opinion with board-level consequences. Humans form it from verified inputs, full stop. |
| DocumentationWorkpaper narrative drafting | High fit | Its strongest documentation use — and the fluency trap. Edit until the narrative says what you concluded; then it is your narrative. Never sign prose you have not read end to end. |
| DocumentationEvidence citation to source documents | Conditional fit | It can propose citations; language models fabricate them under exactly the conditions that make them useful. Every citation gets opened — assertion, document, page. |
| DocumentationIssue / finding drafting | Conditional fit | Draft language from the workpaper, yes. Severity, root cause, and what management must actually do are the reviewer’s words and the manager’s call. |
| DocumentationApproving the issue and its severity | Off-limits | Approval sets what the institution is formally told to fix. That is an accountability moment, and accountability is human. |
| QAQA of AI-generated output | Off-limits | AI output is sampled as its own population, independently, from source. The grader and the graded cannot be the same system. |
| ReportingReview report drafting and committee exhibits | Conditional fit | Aggregation of already-verified workpaper content. Every figure in the report traces to a workpaper figure that already carried its own verification. |
| ReportingSign-off and issuing the report | Off-limits | The signature is the product. It certifies that a named human stands behind every conclusion. There is nothing here for a machine to hold. |
| GovernanceRegulatory examination support | Conditional fit | It retrieves its own logs — prompts, versions, verification records — quickly. The explanation of why your use of it is sound comes from the function’s accountable head, in the room. |