Deal Leverage needs an agent that creates a defensible diligence view from a messy data room. This is not a chat-with-PDFs feature. It is a controlled evidence workflow that inventories documents, maps them to a checklist, extracts required facts, scores completeness, and makes uncertainty obvious.
A buyer should be able to open a checklist line—“trailing 12-month operating statement,” “title commitment,” “insurance certificate,” “environmental report,” “tenant estoppel”—and see the exact file, page, finding, date, and reason for its score. The agent cannot say a deal is safe or approve a transaction. It can organize evidence and surface exceptions faster than a human opening every folder.
Use Plan–Execute. The plan is the checklist taxonomy and required evidence, not model-generated improvisation. Workers identify document type, extract relevant facts, compare them to the criteria, and return a score with citations. Missing evidence remains missing; do not reward a confident guess.
The walkthrough is deliberately opinionated: a smaller explicit checklist beats a magical “analyze this data room” prompt. You earn autonomy only after the evidence path is reliable.
Before adding complexity, run this design through a small representative eval and inspect the trace with the operator who will own failures. Make the boundary, escalation, and rollback visible in the product. That discipline will expose more useful work than another round of prompt cleverness.
The input is deal_id plus a data-room manifest. The output is a versioned checklist: each row has criterion, required_evidence, status, score, finding, source_file, pages, quote, confidence, and review_required. Start with 25–40 criteria across financials, property condition, legal/title, leases, insurance, tax, debt, and operations. Make criteria deal-type specific; multifamily and land should not share the same checklist.
Execution plan
First inventory files with deterministic metadata: filename, path, size, date, type, checksum, OCR status. A classifier labels each file type with confidence. For every checklist row, retrieve only likely documents, then call an extractor constrained to that row’s schema. A scorer compares the extracted evidence to explicit rules: complete, partial, missing, conflicting, or not_applicable. For example, an insurance item is complete only if the certificate identifies coverage, limits, expiration, and named insured; an old certificate is not “partial complete.”
Evidence and scoring
Use a 0–100 deal score only as a rollup, never as a substitute for rows. Weight blockers explicitly and cap the total score when a required legal or financial document is missing. Every positive finding needs a quote and page. Every negative finding names what was expected and what was found. Conflicting documents create a review item rather than averaging their claims.
Human review and eval
Require review for all blockers, scores below 70, low-confidence classification, scanned tables, and any conclusion involving title, environmental, legal, or underwriting judgment. Build gold cases from closed deals and known bad rooms. Score document classification, row status accuracy, citation precision, blocker recall, and false-complete rate. The most important metric is not average accuracy; it is how often the agent incorrectly marks a material item complete.
Operational behavior
Persist each step so new documents update only affected rows. Show a trace per row. Do not ask the agent to “re-read everything” after an upload. That is expensive, slow, and impossible to audit.