Skip to main content
Sevrel
← Blog

Why Verification Matters in CRE Document Analysis

A plausible but wrong lease figure can flow into underwriting, investor materials, and negotiations. AI can accelerate retrieval and comparison, but response speed, retrieval, or a citation does not prove that an answer is correct or complete.

Where Plausible-but-Wrong Numbers Come From

The failure modes worth worrying about are rarely exotic. They are the ordinary complications of commercial documents, and each one produces an answer that looks finished while resting on an incomplete reading:

  • The base rent in the original lease was restated by a later amendment, and the answer quotes the superseded figure with a perfectly valid citation to the original page.
  • An escalation clause is read without the cap or floor defined three sections away, or an expense stop is defined in an exhibit rather than in the article that mentions it.
  • Defined terms diverge from plain English: “Rent,” “Base Rent,” and “Additional Rent” can be three different numbers in the same lease.
  • A scanned page, a handwritten rider, or a poor-quality exhibit was extracted imperfectly, so the passage the model read is not quite the passage counsel signed.

None of these is an argument against using retrieval. Each is an argument for reading the cited passage in context before the number travels anywhere that matters.

What the Architecture Can Do

  • Retrieval: supply passages from documents selected as relevant to the question
  • Source tracking: attach links to retrieved evidence when the answer pipeline produces supported citations
  • Grounding instructions and checks: direct the model toward evidence and flag some unsupported output
  • CRE-aware routing: choose a model and retrieval strategy using configured heuristics

None of these controls guarantees that retrieval found every relevant amendment, that the model interpreted a clause correctly, or that every statement has a valid citation.

The Verification Workflow

  1. Use AI to identify candidate documents, clauses, and comparisons.
  2. Open the linked source and review the relevant language in context.
  3. Check amendments, effective dates, definitions, and calculations.
  4. Have qualified professionals decide how the finding affects the transaction.

The economics of this workflow favor the reviewer. Opening a cited passage and reading the surrounding clause is a small, bounded task; unwinding a wrong figure after it has reached a lender package or a purchase-price adjustment is neither small nor bounded. Verification is cheapest at the moment the answer appears, when the source is one click away and the question is still fresh — which is why the citation belongs next to the answer rather than in an appendix.

It also helps to be deliberate about what gets verified. Not every retrieved passage deserves the same scrutiny: a figure heading into a signed document, a deadline with a notice obligation attached, or a clause that changes who pays for what all sit at the top of the list. A quick orientation question about which documents exist can tolerate a lighter touch. Matching the depth of review to the consequence of being wrong is the judgment the workflow exists to support, not replace.

Material-decision rule

Treat AI output as research assistance, not legal, financial, or investment advice. Verify material figures and clauses against the complete source record.

See Source-Linked Review in Action