The deals too small to underwrite

A middle-market credit analyst costs somewhere between $120,000 and $200,000 fully loaded, and is hired for judgment. Ask one where the day actually goes and the answer is spreading: transcribing financial statements and tax returns into the lender's template. Practitioners commonly put it at a third to half of the job; we have not found an authoritative published figure, which is itself telling about how much attention the task gets.

Whatever the exact share, the work creates no differentiated value. It is extraction from PDFs into a fixed schema.

There's a quieter cost too. When spreading is manual, add-back treatment drifts. Two analysts working the same tax return produce different adjusted EBITDA, because owner compensation, non-recurring items, and related-party transactions each involve a small judgment that nobody wrote down. The committee sees a number and cannot see which convention produced it.

We scoped the first phase narrowly and deliberately: intake and spreading only. No credit judgment.

• Intake and completeness classifies every document in a submission and identifies exactly what's missing. This alone removes a large share of the back-and-forth that stalls files, because the follow-up request is specific on day one instead of on day six.
• Spreading extracts and normalizes financials into the lender's own template, with every figure traced to its source document and page. Click any cell and you get the tax return line it came from.
• Adjustment log itemizes every add-back separately, with its justification, each one individually reversible, so the committee can toggle a convention off and watch the ratios move rather than arguing about a number.
• Covenant monitoring came in a later phase, tracking covenants against incoming borrower reporting and detecting breaches when they occur rather than at quarter-end, which is the difference between having remedies and having a conversation.

Ratios, coverage tests, and reconciliations are computed in tested code, invoked as tools. The language model extracts inputs and explains outputs; it never calculates a debt service coverage ratio.

This is not a stylistic preference. A confidently wrong fixed charge coverage ratio in a credit memo is a decision-grade error that looks exactly like a correct one, and it will pass committee. Every computation in the system is verified arithmetic with the formula shown, and every extracted figure carries a confidence score so a low-confidence input can't silently produce a high-confidence ratio.

The interesting build is credit analysis: risk rating, quality of earnings, structure recommendations. We deliberately didn't start there, and the reason is worth stating plainly.

Anything that contributes to a declination pulls in adverse action obligations: the lender must supply specific principal reasons, which means every negative factor has to be traceable to a stated, permissible, non-proxy variable and reconstructable after the fact. That is real architecture, not a disclosure. Spreading and completeness checking carry almost none of that surface. They're mechanical, so they ship in weeks rather than stalling in review for months, and they happen to be where most of the hours actually are.

The boring agent is usually the one worth building first. It carries the most hours and the least risk.

The credit committee decides. Every credit decision, every declination, every adverse action notice remains entirely with the lender's people. The system produces decision support with its work shown, which is both what the lender wanted and what the regulatory posture requires.

Spreading time per file dropped from 2.6 hours to about 25 minutes of review, and time from complete file to committee from 6 days to 2. The same eight analysts now underwrite 2,250 deals a year against 1,450 before, and the lender now quotes deals down to $75,000, a segment it previously declined on economics alone.

The throughput number is the easy one to report. The segment expansion is the one that changed the business.