The Standard
Numbered Sections
1.
Purpose and scope.This Standard defines the conditions under which an AI agent may be used to extend loan review coverage across an entire portfolio while preserving human accountability for every conclusion. It is vendor-neutral: it describes a control architecture, not a product. It applies to second-line credit review and to external loan review practices; it does not address origination, underwriting, or first-line monitoring.
2.
The coverage principle.Sampling is a rationing mechanism, not a methodology preference. A review function samples because human capacity is fixed, and it defends the sample because it cannot defend the coverage. Where an agent can competently draft the production work of review, the binding constraint moves: full-portfolio coverage becomes achievable, and the design question becomes the allocation of human depth. Under this Standard, every exposure in scope is reviewed every cycle; human examination is concentrated on the files that warrant it, selected by deterministic rule rather than by sampling frame.
3.
Definitions.Continuous Credit Review: a review regime in which 100% of in-scope exposures receive a complete draft review each cycle, with escalation to human reviewers governed by pre-published deterministic triggers. Borderline file: a file whose facts sit near a boundary — a rating breakpoint, a policy limit, a covenant threshold — such that reasonable reviewers could differ; in practitioner vernacular, a “cuspy” file. Borderline files are the escalation target. Trigger: a deterministic, institution-configured condition evaluated against file facts, whose firing changes a file’s routing. Tier: the routing disposition assigned to a file by the trigger evaluation.
4.1
Tier 1 — Clear.No triggers fire. The agent completes a full draft review; a qualified human reviews and signs off in batch. Accountability for each file rests with the signing reviewer. Batch sign-off is a workflow efficiency, not a transfer of ownership: a named human is accountable for every file in the batch.
4.2
Tier 2 — Watch.One or more soft triggers fire. The agent completes a draft review; a qualified human examines and signs the file individually before it enters the workpaper.
4.3
Tier 3 — Escalate.Any hard trigger fires. The file is routed to a human reviewer for full review. The agent’s workup — extractions, computations, and the triggers that fired — accompanies the file as raw material. It is not a recommendation, and the reviewer’s conclusion is formed independently of it.
4.4
Tier 4 — Urgent.A deterioration or event trigger fires. The review function is alerted the same day, outside the normal cycle. Urgent routing exists because deterioration does not wait for a review schedule.
5.A
Rating proximity.Metrics near rating breakpoints; ratings adjacent to criticized classifications; rating migration since last review; split or contested ratings. Files near a boundary are where reviewer disagreement concentrates, and are escalated on that basis alone.
5.B
Leveraged lending.Leverage threshold proximity; capacity-to-repay tests; back-end-weighted cash flow profiles; sensitivity of the rating to high-risk EBITDA addbacks, established by dual computation — the rating is checked with and without such addbacks, and divergence escalates the file.
5.C
Cash flow and liquidity.Negative free cash flow after maintenance capital expenditure; liquidity runway measured against burn; fixed-charge coverage computed on a maintenance-CAPEX basis.
5.D
Commercial real estate.Near-term maturity where the refinance case fails a market-terms takeout test; proforma NOI without demonstrated support, which is treated as absence of support; baseline debt-service-coverage ladders.
5.E
Documentation and process.Stale financial statements; open post-closing items; covenants that cannot be tested from the file, which are treated as failed rather than passed; policy exceptions without documented approval.
5.F
Events and environment.Covenant breaches; servicing deterioration; industry watch-list membership; adverse-media flags. An adverse-media flag routes a file for human investigation only — it is never, under this Standard, a conclusion about the borrower.
5.G
Model integrity.Low extraction confidence; inconsistency between documents in the same file; disagreement between the agent’s check and the incumbent rating. The agent’s own uncertainty is a routing input: a file the machine cannot read cleanly is by definition not a clear file.
6.
Institution configurability.The trigger categories above are fixed by this Standard; every threshold within them is set by the institution. The reviewing institution’s credit policy, rating definitions, and risk appetite are the operating rules of the agent, encoded at onboarding and recalibrated as they evolve. An implementation that imposes its own rating philosophy, or that overrides institutional thresholds with vendor defaults, does not conform to this Standard.
7.
The random-audit control.A permanent control: 5–10% of Tier 1 files are randomly routed to full human review each cycle. The control is not provisional and is not retired as confidence grows; it is the ongoing empirical answer to whether the trigger framework is missing what it should catch, and the primary source of adjudicated cases under Section 8.
8.
Governance metrics.A conforming implementation publishes, at minimum: escalation recall — the share of expert-adjudicated borderline files the agent escalates, with a target of at least 98%; and Tier 1 agreement — concordance between the agent’s disposition and the human conclusion on randomly audited clear files, with a target of at least 95%. Every adjudicated miss becomes a new trigger: when human review or examination identifies a file the agent should have escalated, the case is adjudicated and the framework is amended. Metrics are reported to the institution’s governance committee on a defined schedule.
9.
Human accountability.In every tier, accountability for each credit conclusion rests with exactly one named, qualified human. The agent may be responsible for producing draft work product and may be consulted or informed; it is never accountable, and judgment tasks reserve no role for it at all. Deterministic rules route files; humans decide them. The reference allocation of roles across the review lifecycle is maintained as The Credit Board’s RACI exhibit.
10.
Maintenance and revision.This Standard is maintained by The Credit Board and revised through practitioner adjudication: submitted borderline cases, random-audit findings, and examination outcomes. This edition is an exposure draft; comment from practitioners, model risk teams, and supervisory staff is invited before the Standard is finalized.