Credit Memo
A workflow that drafts a credit memo while distinguishing provided financial data, estimates and missing information.
My role and the decisions
I built a multi-agent pipeline, a researcher, a financial analyst, a risk assessor, and a memo writer, with a human review gate between the analysis and the final memo. Each agent has one job and hands its output to the next.
The main decision was to make the basis of the financial analysis visible. The snapshot labels figures as sourced, estimated or not available. In this fictional sample, sourced means supplied in the test input. Estimates remain explicit and do not fill in undisclosed company results.
The second decision was letting the pipeline say no. I did not want a tool that always finds a way to recommend the deal. On its test company, the pipeline concluded Pass, because the loan book performance and real cash position were undisclosed and no covenant could fix that. A memo generator that is allowed to say no is a memo generator I'd actually trust.
What shipped
A pipeline that researches a company and drafts a concise credit memo with an executive summary, business overview, financial snapshot, ranked risks and a recommendation. A person reviews the analysis after risk assessment and before the final memo is written.
The sample below shows the output from a test run.
Published artifact
Harbor Lending - Credit Memo
One-page memo. Synthesized from research.md, financials.md, risks.md. Harbor Lending is a fictional/synthetic company used for pipeline validation; assessed here as a real credit decision on the facts provided.Executive Summary
Harbor Lending is a seed-stage fintech offering embedded working-capital loans to SMB e-commerce sellers. It has $3.2M ARR, $6M in equity raised ~8 months ago, and no existing debt. The ask is a credit facility to this lender. Recommendation: Pass — Harbor is a credit business whose own credit quality and cash position are both undisclosed, so there is no basis on which to underwrite it today.
Business Overview
Harbor originates embedded working-capital loans to SMB e-commerce sellers, earning take-rate / interest income on loan volume, likely via platform/marketplace distribution partners (the embedded-finance pattern). Category analogues include Shopify Capital, Parafin, and Clearco, though no direct competitors and no differentiation were provided. Funding: a single $6M seed round ~8 months ago; investors, valuation, and lead are [not available]. Management team — founders, backgrounds, underwriting track record — is [not available].
Financial Snapshot
| Metric | Value | Tag |
|---|---|---|
| ARR | $3.2M | [sourced] |
| YoY growth rate | not disclosed | [not available] |
| Take rate on loan volume | 2–8% | [estimated] |
| Net margin post credit losses | ~1.5–3% | [estimated] |
| Annualized default rate | ~3–12%, segment-dependent | [estimated] |
| Cost of loan funding | ~6–12% | [estimated] |
| Monthly burn | ~$100K–$150K/mo | [estimated] |
| Implied runway | ~40 months (rough sanity-check only; likely overstated) | [estimated] |
| Existing debt | None | [sourced] |
| Equity raised | $6M seed, 8 months ago | [sourced] |
Only ARR, cap structure, and raise timing are company-level facts. All margin, loss, burn, and runway figures are industry benchmarks, not Harbor data. The ~40-month runway is internally inconsistent (2x the 12–24 month seed norm), signaling the benchmark understates true spend — treat as unreliable.
Key Risks (ranked)
- Undisclosed loan-book performance + unknown real burn/cash — disqualifying. Zero data on the only thing that matters for a lender: no default rate, delinquency, charge-offs, loan-book size, or funding cost. Real burn/cash is also unknown; the runway figure can't be trusted. We would be underwriting a credit business blind to both its credit quality and its ability to absorb stress.
- Management/team risk — total blind spot. No names, backgrounds, or underwriting track record. For a seed-stage credit underwriter, the team's credentials are the asset, and there is nothing to assess.
- Cyclicality/market risk — structurally high and pro-cyclical. SMB e-commerce sellers default exactly when Harbor's funding cost rises and equity thins. The top of the 3–12% default range likely erases the 1.5–3% net margin entirely, with no disclosed reserves or hedging.
- Concentration risk — unassessed. Not disclosed; embedded lenders typically lean on one or two platform partners, so plausibly high but unverifiable.
- Covenant risk — moot until #1 is resolved. The needed covenants (liquidity floor, max default rate, min post-loss NIM, tangible-net-worth floor) cannot be set or monitored with no baseline. No existing debt is the one genuine positive.
Recommendation
Pass. Harbor is a lender whose own loss rates, loan-book, and cash/burn are all undisclosed — you cannot underwrite a credit business blind to its defaults and its runway, and no covenant can be structured to protect against it. This can be reopened only if Harbor delivers actual loan-book performance (defaults, delinquency, charge-offs, book size, funding cost) and an audited cash/burn statement; absent those two disclosures, there is no basis to lend.
Outcome
In this test run, the pipeline recommended passing because the information needed to underwrite the lender was missing. The financial tags make the gaps visible to someone who did not take part in the research.
What I'd do differently
I'd add a test company where lending could be supported with specific conditions, then check whether the pipeline distinguishes that case from a deal with too little information to underwrite.