Argus is a review agent. It takes your function to 100% coverage by completing the standard reviews itself and flagging the complex deals for your reviewers — with every threshold set by your own credit policy, and every conclusion signed by a person.
Argus was formerly named CRAA. Same product, wider mandate.
AI can produce the work. Only a human can own the conclusion.
No jargon. Six steps, in the order they happen — inside your bank’s own environment.
The data warehouse, the document repository, the loan system. Read-only. Nothing new for your lenders to fill in; nothing for borrowers to see.
Financial statements, loan agreements, covenant certificates, appraisals, memos — the same file your reviewers would pull, read completely, every time.
Your credit policy, your rating definitions, your risk appetite — the rules your board and committee already approved. It brings no opinions of its own.
The spreads, the ratios, the narrative, the rating rationale — the production work that consumes most of a review team’s year.
Anything near a line — a rating boundary, a covenant, a policy limit — goes to your reviewers, with the workup attached. Your thresholds decide what “near” means.
Nothing becomes a rating until a qualified person makes it one. The agent produces work; your reviewers own conclusions.
Annual review of a slice of the book was never the goal — it was the most a fixed team could reach. The rest of the portfolio goes unexamined between cycles, and everyone knows it.
Seasoned reviewers are retiring faster than they are being replaced. The judgment that made sampling defensible is exactly the resource in shortest supply.
A credit graded in March under one reviewer and re-graded in November under another can move for reasons that have nothing to do with the borrower. Drift is invisible until an examiner finds it.
Whether or not you use AI, supervisory teams increasingly expect a documented position on it. “We don’t use it” is a policy statement that still has to be defended.
Continuous Credit Review is full-portfolio coverage with risk-allocated human depth. The question changes from “was this file in the sample?” to “did anything about this file warrant a person’s attention — and if so, did it get it?”
| Sampling regime | Continuous review | |
|---|---|---|
| Coverage | A risk-weighted slice, annually | The whole book, every cycle |
| Human depth | Spread evenly across the sample | Concentrated on triggered files |
| Between reviews | Deterioration waits for the next cycle | Event triggers surface it same-day |
| To the examiner | “Our methodology was sound” | “Here is every file, and who looked at it” |
The full argument is set out in the Creditboard Standard.
Argus began as an assistant: you opened it, handed it a credit, and asked for a review. It is now an agent — and the difference is not a feature, it is the direction of the workflow.
An assistant reviews what you bring it, when you bring it. Coverage is still rationed by human initiation — the tool made each review faster, but the book was no more covered than before.
Argus works the portfolio on its own schedule: it takes in financials and documents as they arrive, recomputes the ratios, evaluates every trigger, drafts the reviews, and queues what needs a person. Your reviewers open a morning queue of files that genuinely need judgment — with the workup already attached.
Agentic in one precise sense — it initiates and completes production work without being asked — and in no other. It cannot widen its own scope, change a threshold, or own a conclusion. The rules that govern it are your policy, and the tiers below are how its work reaches people.
Escalation is not a model feeling uncertain. It is a published set of triggers — your thresholds, applied uniformly — that decide which files a person must examine. Argus never concludes; it drafts, flags, and routes.
No triggers fire. Argus drafts the full review — spreads, ratios, narrative, rating rationale — and a reviewer signs off in batch. A named human still owns every file.
Soft triggers fire. Argus drafts the review; a reviewer examines and signs the file individually before it enters the workpaper.
Any hard trigger fires. The file is routed to a human reviewer for full review, with the agent’s workup — computations, extractions, the triggers that fired — attached as raw material, not as a recommendation.
A deterioration or event trigger fires — covenant breach, servicing deterioration, adverse development. The file alerts the review team the same day, ahead of any cycle.
5–10% of Tier 1 files are randomly routed to full human review, every cycle, forever. Random audit is how the tier boundaries stay honest — and the standing answer to “how do you know it isn’t missing things?”
This is the training’s RACI exhibit applied at portfolio scale: the machine is Responsible for drafts; a human is Accountable on every row.
Seven trigger categories cover rating proximity, leveraged lending, cash flow and liquidity, CRE, documentation, events, and model integrity. Three examples below; the full descriptive framework is in the Standard. Every threshold in every category is configurable to your institution’s policy — the categories are ours, the numbers are yours.
The rating is computed twice — with and without high-risk EBITDA addbacks. If the answer changes, the file stops being straightforward and routes to a human reviewer with both computations attached.
A near-term maturity where the refinance case fails a market-terms takeout test — current rates, current vacancy, actual NOI — is escalated. An unsupported proforma is treated as absence of support, not as evidence.
A covenant that cannot be tested from the file — missing certificate, stale financials, undefined terms — is treated as failed, not as passed. Silence in the file is a trigger, never a comfort.
Every bank has its own credit policy, rating definitions, and risk appetite. Argus does not arrive with a rating philosophy; it enforces yours, consistently, across the entire book.
Your best reviewer’s judgment, applied to every file. Your policy, without drift.
Argus is not one black box. It is three layers, each doing what it is genuinely best at — which is what makes the output defensible to a model-risk team and an examiner.
Ratios — DSCR, leverage, coverage, LTV — are computed from the raw financials, not trusted from free text. Escalation triggers and hard policy gates are enforced in code, not by persuasion. Same inputs, same routing, every time.
Within those guardrails, the model does the production work that is un-staffable at full-book scale: spreading, extraction, red-flag synthesis, and a draft narrative in examiner style — each assertion tied back to a specific data point in the file.
The reviewer accepts, or overrides with a documented rationale. Nothing is a rating until a qualified person makes it one. Every decision — draft, triggers fired, routing, human sign-off, override reason — is stamped and written to an immutable audit trail.
Because a human owns every conclusion, Argus is a coverage and consistency system — not an autonomous decision model. Where scrutiny applies, we bring evidence, and client data is never used to train a model.
Against a library of expert-adjudicated borderline files, the share the agent correctly escalates is tracked with a target of at least 98%. This number is reported, not asserted.
5–10% of files the agent judged straightforward are routed to full human review anyway, every cycle. Agreement between the human and the agent — target at least 95% — is the ongoing answer to “how do you know it isn’t missing things?”
When the random audit or an examiner finds something the agent should have escalated, the case is adjudicated and a new trigger is added. The framework only tightens.
Model purpose, limitations, validation evidence, and change history are documented in the structure your model risk team already uses.
Every review stamps its inputs, the triggers evaluated, the routing decision, the draft, and the human sign-off. Nothing is overwritten; everything is reconstructable.
Batch sign-off on the clear tier, individual sign-off on the watch tier, full human review above that. There is no tier in which a conclusion belongs to the software.
Loan-review and consulting firms aren’t just competition — they’re the sharpest users. Argus turns a fixed-capacity practice into a scalable one.
The agent does the first-pass grind, so reviewers spend billable hours on judgment — expanding margin without expanding headcount.
Offer always-on coverage between annual engagements — something no manual practice can deliver today.
Strict multi-tenant isolation per client institution, under your own identity and workflow.
A forward-deployed engineer builds Argus on your own cloud, as an integration with the systems you already run — no license, no data leaving your boundary. The how, the unification layer, and the security answers live on their own page.
Deployment →Coverage pressure, lean teams, and examiners asking about AI governance — written directly to the person who owns the review function, with what you can take to your governance committee today.
For review leaders →Senior credit review practitioners assess your current coverage, encode your policy into Argus’s operating rules, and stay on retainer as your risk appetite evolves.
Advisory →Bring a file you already know the answer to. In one working session with a practitioner you’ll see the draft, the triggers, the routing — and exactly where your judgment still decides.
Advisory only. Argus’s output is for informational purposes and requires reviewer judgment. The tool does not constitute a formal credit determination and is not a substitute for the independent judgment of a qualified credit professional. Every conclusion is owned by a person, not the software.