Commercial credit memos and approval decisions for lending agents

This page is about lending memos, not accounts-receivable "credit memos" (billing credit notes), which belong with procure-to-pay and order-to-cash records. It sits in the agent, workflow and domain-reasoning data hub and focuses on commercial and business lending.

What a commercial credit memo contains that agents learn from

A commercial credit memo is a structured argument: facts about the borrower, numbers from the spread, a judgment about risk, and a proposed structure with conditions. Moody's describes the memo as the point where origination data, financial analysis, market context and credit judgment converge, and notes that at many banks it is still assembled by hand from siloed systems [1]. That hand assembly is exactly why memos are valuable training data: they record how an analyst selected, reconciled and weighed evidence.

Most memos share these sections, whatever template the lender uses:

• Request and purpose: facility type (revolver, term loan, CRE mortgage, equipment, SBA 7(a)), amount, tenor, pricing, use of proceeds.
• Borrower and guarantor overview: ownership, management, industry, NAICS code, relationship history, existing exposure.
• Financial analysis: historical spreads, DSCR, fixed charge coverage, funded debt to EBITDA, global cash flow including guarantors, working capital and liquidity, projections versus actuals.
• Collateral: appraisal values, loan-to-value, borrowing-base advance rates, lien position.
• Risk rating: the internal grade (often a 1 to 9 or 1 to 10 scale mapped to pass, special mention, substandard, doubtful) with the rationale.
• Structure and conditions: financial covenants, reporting requirements, conditions precedent, guarantees.
• Policy exceptions: each exception, its mitigant and who approved it.
• Recommendation and decision: analyst recommendation, approver or committee outcome, modifications made at committee.

Microsoft's scenario for credit memo generation shows the target workflow an agent automates: pull customer documents, calculate ratios, compare results to internal credit policy and draft a note with decision status [2]. Training data that only contains final memo text teaches drafting; data that links memo text to the spread and the policy teaches analysis.

Why the approval decision matters more than the memo text

The decision record, not the memo prose, is the label that makes this data useful for evaluation. A memo recommends; the approval authority or credit committee decides, and committees routinely change structure, add conditions, lower amounts or decline against the recommendation. If you license memos without the final decision and the committee's modifications, you can train a writer but you cannot measure whether an agent reaches the same credit conclusion.

Ask for the decision as structured fields, not just as a signature page. Useful fields include decision outcome, approving authority level, date, final approved amount, conditions added at approval, exceptions approved, and dissent or minutes excerpts where the lender keeps them. For the general pattern of capturing outcome plus reasoning, see decision records with rationale and approval and rejection records.

Risk ratings deserve care as labels. The federal banking agencies' uniform definitions treat special mention as potential weaknesses that deserve management's close attention but are not an adverse classification, while substandard means well-defined weaknesses that jeopardize liquidation of the debt. Each lender maps its internal grades to these categories differently, so ask for the mapping and the rationale text, because that weighing of evidence is what an evaluation set should test.

Declines and adverse action reasons in business lending

Declined applications are the most informative and most often missing part of a credit memo dataset. Survivorship is the usual failure mode: a lender's loan system holds booked loans, while declined or withdrawn deals live in a separate pipeline tool, email or nowhere. An agent trained only on approvals never learns where the policy line sits.

Regulation B (12 CFR 1002.9) has separate adverse action notification rules for business credit, with different handling for smaller and larger businesses; confirm current text with counsel. Practically, smaller-business declines often carry a stated reason or a right-to-reasons disclosure, while larger-business declines may have only internal memo language. Ask suppliers which declines have recorded reasons, in what form, and whether those reasons match the memo's analysis. Consumer-side reason codes are covered separately in credit decision records with adverse action reasons.

Fair lending review is part of buying this data. If the agent will influence decisions, check whether any memo fields (owner demographics, location proxies, free-text commentary) could leak protected characteristics into training, and decide before delivery which fields are dropped rather than masked.

Linking memos to loan performance

Memo data becomes evaluation-grade when each approved deal links to what happened next. Risk ratings are a forecast; performance is the outcome. The most useful post-decision fields are rating migrations with dates, covenant compliance certificates and breaches, waivers and amendments, past-due status, nonaccrual, troubled-loan modifications, charge-off and recovery.

Two cautions apply. First, performance windows are long: a three-year term loan approved in 2024 has not fully seasoned by October 2026, so mark censored outcomes explicitly. Second, outcome fields in operational systems are not automatically reliable labels; downgrades can lag reality and charge-off timing reflects accounting policy. See verifying outcome labels in operational records before treating any of these as ground truth.

Annual reviews are an underused source. They repeat the memo's analysis a year later with actual results, so they let you compare projected versus realized DSCR and check whether the original analyst's stated risks materialized.

Formats, systems and the spreading problem

Credit memo data rarely arrives as clean text; expect Word or PDF memos, Excel spreads and exports from loan origination systems. Lenders commonly spread financials in dedicated tools (Moody's CreditLens and Abrigo's lending products are examples) or in bank-built Excel templates, then paste tables into the memo. The memo and the spread can disagree after late edits, so ask for both and a flag for which is authoritative.

Plan for these formats and joins:

• Memo documents (DOCX or PDF) with section structure preserved, not flattened OCR.
• Spreads as XLSX or CSV with line-item mapping to a chart of accounts, period end dates, and whether statements were audited, reviewed, compiled or tax-return based.
• Source documents the analyst used: financial statements, tax returns, rent rolls, accounts receivable agings, appraisals. These are the inputs for document-to-system entry pairs.
• Workflow events from the origination system (stage changes, reviewer comments, returned-for-revision loops), which give trajectory data similar to spreadsheet task trajectories.
• A stable deal key linking memo, spread, decision, booking and performance, without exposing the lender's real loan numbers.

Revisions are valuable: a credit officer's edits to an analyst's draft are correction data, close to human override and correction logs. Ask whether version history exists in the document management system.

Confidentiality, privacy and permissions

Commercial memos are confidential business records about third parties, so the supplying lender's right to release them is the first diligence question. Borrower financials are typically covered by loan agreement confidentiality clauses and the lender's own confidentiality obligations, and memos often include guarantor personal financial statements, personal tax returns and credit bureau excerpts. Where consumer nonpublic personal information is involved, Regulation P restricts how information received from a financial institution may be reused and redisclosed [3].

De-identification for this data goes beyond names. Business names, addresses, EINs, NAICS code plus city plus revenue band, unusual collateral descriptions, and narrative commentary can identify a borrower in a small market. NIST notes that traditional de-identification has inherent limitations compared with formal privacy methods [5], so pair removal with generalization (revenue bands, regional rather than city geography) and a re-identification review of free text.

If the agent will be used in decisions about individuals, note that, as of October 2026, Colorado's SB26-189, signed May 14, 2026, replaced SB 24-205 and, from January 1, 2027, requires developers of automated decision-making technology used in consequential decisions to give deployers technical documentation [4]. Keep provenance documentation for training data in a form you can hand over, such as a Data Card covering sources, collection, annotation and intended use [6].

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Buyer request template for credit memo data

A precise request names the loan types, the decision fields and the joins you need, not just "credit memos." Dataset license terms are often missing or wrong even in public collections; one audit found license omission above 70% and error rates above 50% on popular hosting sites [7], so write the permitted uses into the request from the start.

Illustrative example: invented to show structure; it does not describe an available dataset.

Specify the evaluation split before delivery. A held-out set by lender, vintage and loan type is more honest than a random split, because memo templates and credit culture vary by institution and leak style cues across a random split.

Related owner pages cover adjacent data: underwriting files for complete loan files, financial transaction data for account activity, and finance buyers for the wider industry view. Mortgage-specific condition clearing is covered in mortgage underwriting conditions data.

How SourceX handles credit memo requests

SourceX sources operational datasets from US companies on request, including document-heavy finance and legal workflows; it holds no stock, and a request does not guarantee a match. Buyers describe the data they need, SourceX looks for US businesses that hold it, and every release is approved by the supplying company. Each dataset is rights-reviewed for ownership and consents, personal details such as names, emails, phones and account numbers are removed or replaced before delivery with the method recorded and a sample checked, and no method is perfect. You can describe your credit memo requirements to SourceX using the template above.

Source commercial credit memo data for your agent

SourceX runs a Find, Assess, Agree, Transact and Manage process, and nothing is contracted until a supplier agrees. Datasets are delivered under a license defining records, uses, term and delivery, through private, access-controlled workflows. To start, describe the credit memo and decision data you need.