Credit memos in minutes. Exceptions that cannot slip through.
The analyst's ratio work is computed in code from facts quoted out of the file. Covenant and leverage breaches reach a credit officer with the analysis finished; nothing is approved, priced or waived by software.
Run it
Run a commercial real estate file: the sample deal, a clean control, a covenant breach or an embedded-instruction attack.
Unedited file: the engine runs its committed facts and no model is called. Synthetic data only. Identifiers are redacted before any model sees the text.
How the agents hand off
One file through the fleet: verification, the decision, and a reviewer who samples what was auto-rendered.
Where it fits in the process
The lifecycle of a file, and the first workflow worth automating.
01Origination / Sourcing
Days to several weeks, depending on how fast the borrower sends documentsA banker (the relationship manager) finds, or is approached by, a business that wants to borrow. They talk through what the company needs the money for, how much, and roughly what the deal would look like, then collect the first batch of documents.
Who: Relationship Manager (the salesperson-banker) and the borrower
Automated credit spreading & covenant extraction before committee
Spreading is the slow, error-prone front step whose mistakes poison every downstream ratio.
- Map each statement line to the template
- Compute DSCR, LTV, debt yield, cash flow
- Extract covenants and existing debt
- Analyst reviews low-confidence flagged cells
- Committee keeps the lend-or-decline vote
Who else serves this
8 companies already selling into this workflow, described from their public material, and the gap each leaves for a lender that needs a replayable decision. None is affiliated with Arbiter.
Document-AI platform that extracts and structures data from 100s of document types (bank statements, pay stubs, tax forms) at 99%+ accuracy with human-in-the-loop validation, computes income/cash-flow, and runs fraud detection ("Detect")…
Gap for a lender: Sits below the decision - it verifies and feeds data but does not make the credit decision, so it is a component vulnerable to LOS-native extraction (ICE Analyzers, nCino nIQ) and to lenders building…
Source: ocrolus.comCloud bank operating system (on Salesforce) spanning onboarding, account opening, loan origination, credit analysis, and portfolio management. nIQ Data Recognition auto-extracts data from financial statements/tax docs.
Gap for a lender: AI is largely an assistive copilot/data-recognition layer, not autonomous decisioning - bankers still drive credit decisions; broad-but-shallow across verticals.
Source: ncino.comAI-driven business-lending origination platform automating the SMB/commercial lending workflow from application to decision to closing (incl. SBA automation, product templates), processing $50B+ in loans for 30,000 lenders / 500,000+…
Gap for a lender: Decisioning still largely supports human credit approval rather than fully automating it; SMB/commercial credit is heterogeneous and judgment-heavy, limiting straight-through autonomy.
Source: numerated.comCommercial-lending credit platform for consistent financial spreading, dual risk-rating models (or bank-configured policies), benchmarking, and BPMN/CMMN credit-approval workflow.
Gap for a lender: Primarily a spreading/risk-rating and workflow tool that informs human credit committees rather than auto-deciding; commercial credit resists full automation.
Source: moodysanalytics.comCompliance, credit-risk and lending platform for 2,400+ community banks/CUs: loan origination, Sageworks credit analysis/spreading, CECL/ALM modeling, and financial-crime/BSA-AML.
Gap for a lender: AI is assistive (narratives, extraction, review acceleration) rather than autonomous decisioning; serves smaller institutions with lower volumes; less brand-name credit data than Moody's.
Source: abrigo.comSells "The Agentic Decision Platform": a decision engine, an AI agent manager, a case manager and a context layer for financial services.
Gap for a lender: It is a platform the customer builds decisions on; the credit policy, its validation and its test suite remain the customer's work.
Source: taktile.comSells a "Decision Intelligence Platform" for credit risk, fraud, identity, collections and customer-management decisions.
Gap for a lender: A general decisioning platform: document reading, evidence for each extracted fact and commercial-credit memo work are left to the lender or to partners.
Source: provenir.comFICO's enterprise decisioning platform: decisioning with business rules, mathematical optimization, simulation, real-time analytics and ML, and support for human-in-the-loop AI across the customer lifecycle.
Gap for a lender: Built for enterprises with large decision-science teams; heavy to deploy for a regional bank's commercial credit desk, and extraction from unstructured credit files is not its centre of gravity.
Source: fico.comField notes
Reference material behind the design: who is involved, the documents in the file, the cost structure, the metrics a pilot would move, and the agent specification.
StakeholdersWho touches the file and what each owns
The face of the bank to the business. They bring in the deal, figure out what the client needs, and shepherd it through the bank, but they don't usually do the number-crunching themselves. Think of them as the account owner who sells and negotiates.
The person who does the grunt work: re-types the financial statements into the bank's template (spreading), calculates the ratios, and drafts the credit memo. Most of the manual hours on a deal are theirs.
A senior banker (or a panel of them) who reads the credit memo and votes yes or no. They are the gatekeepers protecting the bank from bad loans; bigger or riskier loans need more senior or more people to sign off.
An outside expert hired to give a professional, unbiased estimate of what a building is worth. The bank needs this to know how big a loan is safe versus the value of the collateral.
Writes the binding legal contract, including the covenants (rules the borrower must follow), makes sure the bank's claim on the collateral is legally airtight, and handles the closing.
The company that wants the money, plus the owners who personally promise to repay if the business can't (the guarantors). Their tax returns and financials are the raw material the whole process runs on.
Documents in the fileWhat each document is and why it matters to the decision
Business tax returns
The tax forms the company files with the IRS, showing its income and expenses. Banks use these because they're hard to fake and give a few years of history.
They are the most trusted source of a company's real earnings and are the starting point for spreading and cash-flow analysis.
A 3-year stack of Form 1120-S returns for 'Maple Street Diner LLC' showing $1.2M revenue and $180K net profit in the most recent year, used to confirm the business earns enough to cover a new $500K equipment loan.
Cost structureWho pays for what, with illustrative figures
More for large or complex properties; a common cause of delay as well as cost.
Checks the property for contamination risk; often required on CRE deals.
The bank's upfront charge for making the loan.
Lawyers drafting and closing the loan; scales with deal complexity.
The labor the AI wedge targets; spreading one entity's return alone is ~30-60 minutes.
Metrics a pilot would moveIndustry reference figures from the sources listed; not results of this demo
How much human time an analyst spends to take one loan from raw documents to a committee-ready memo. Lower is better and means cheaper, faster lending.
How long from receiving the borrower's documents until the finished memo is ready for a decision.
How often a number gets mistyped or mis-mapped when financials are copied into the bank's template. Errors here poison every downstream ratio and decision.
The share of memos that reach the committee complete and correct the first time, without being bounced back for missing data or mistakes.
Cash the business or property generates divided by the loan payments. Above 1.0 means it earns more than enough to pay; banks want a cushion.
Loan amount divided by the appraised value of what backs it. Lower means the bank has more cushion if it has to sell the collateral.
Sources: abrigo.com · propertymetrics.com · occ.gov · crefcoa.com · crefcoa.com · financelobby.com · southstatecorrespondent.com · jpmorgan.com
Agent specificationThe build specification for Arbiter Analyst
GlossaryEvery term used on this page, searchable
Re-typing a borrower's tax returns and financial statements line-by-line into the bank's standardized template so every borrower can be compared the same way.
e.g. An analyst enters each line of 'Maple Street Diner's' 3 years of tax returns into the bank's credit software to build a clean, comparable financial picture.
The structured write-up that summarizes the analysis, ratios, risks, and a recommendation, and is the document the decision-makers actually read.
e.g. A 12-page memo concluding 'recommend approval of a $500K equipment loan to Maple Street Diner with a 1.40x debt-service coverage and a personal guaranty.'
A panel of senior bankers who read the memo and vote to approve or decline larger or riskier loans.
e.g. On Thursday the committee reviews five memos and approves the diner loan with a condition that the owners keep $100K in reserves.
The cash available to pay debt divided by the debt payments owed; it answers 'can they comfortably afford this?'
e.g. A property with $125,000 of net operating income and $100,000 of annual loan payments has a DSCR of 1.25x.
The loan amount divided by the appraised value of the collateral; it answers 'how much skin does the borrower have versus the bank?'
e.g. A $750,000 loan on a building appraised at $1,000,000 is a 75% LTV.
The property's annual net income divided by the loan amount, showing the bank's plain cash return if it had to take the property over, independent of interest rates or appraised value.
e.g. A property earning $100,000 net income with a $1,000,000 loan has a 10% debt yield.
A combined view of all the income and expenses of both the business and its owners (guarantors), to see the total ability to repay.
e.g. A property that barely covers its own loan still qualifies because the owner's separate $200,000 salary is added in to show total repayment capacity.
For a property, the rent collected minus the cost of running the building, before any loan payments.
e.g. An apartment building collecting $360,000 of rent with $135,000 of operating costs has an NOI of $225,000.
A rule written into the loan contract that the borrower must keep following, acting as an early-warning tripwire for the bank.
e.g. A covenant requiring the borrower to keep total debt below 3x annual earnings; exceeding it is a breach the bank can act on.
The dollar limit on a revolving credit line, set by the value of the borrower's invoices and inventory and recalculated regularly.
e.g. With $2M of eligible invoices at an 80% advance rate, the borrowing base allows up to $1.6M to be drawn that month.
A person (usually a business owner) who personally promises to repay the loan if the business cannot.
e.g. The two 50% owners of the diner each sign a personal guaranty, putting their own assets on the line for the $500K loan.
A commercial-and-industrial loan to an operating business for working capital, equipment, or growth, repaid from the company's profits.
e.g. A $250K revolving line of credit a manufacturer draws on to buy raw materials and repays as customers pay their invoices.
A commercial real estate loan to buy, build, or refinance an income-producing property, repaid from the rent the property earns.
e.g. A $3M loan to purchase a strip mall, repaid from the rent its tenants pay.