Find The Nation's Top Attorneys

The challenge

ArcLend underwrote commercial property loans. Each application arrived as two to four hundred pages across leases, valuations, company accounts and title documents, and an analyst spent between three and five hours locating perhaps forty clauses that actually drove the decision.

The obvious framing — have a model read the file and recommend approve or decline — was not viable. This is a regulated decision requiring a documented human rationale, and a recommendation an analyst cannot audit is worse than no recommendation, because it invites deference without understanding.

ArcLend had also already run a failed proof of concept with another vendor. It demonstrated well and was abandoned in pilot: nobody could say whether it was getting better or worse between versions, because there was nothing to measure it against.

What we did

We started by building the scoring set, not the pipeline. Two ArcLend analysts marked up three hundred historical files, tagging where each required clause actually appeared. That took three weeks and became the scoreboard every later change was measured against — the thing the previous attempt never had.

The pipeline retrieves rather than summarises. Documents are chunked with layout awareness so a table is not severed mid-row, embedded, and queried per clause type. The model's job is to locate and quote, not to conclude.

The interface reflects that division. Each required clause appears with the extracted text, a confidence signal, and a link that opens the source PDF at the exact page and highlights the passage. An analyst confirms, corrects or marks it absent, and every correction is written back as new evaluation data.

Nothing is auto-approved. The system produces a structured, cited file that a human signs, which is what makes it defensible to ArcLend's regulator — the audit trail shows a person deciding, with the evidence they saw.

The whole estate is Terraform-defined, with per-request cost tracking and caching around the retrieval layer that cut inference spend by roughly two-thirds once usage patterns were clear.

What changed

Median review time fell from four hours ten minutes to fifty-five minutes, measured across six months of production use. Clause retrieval accuracy against the held-out portion of the scoring set sits at 94%, and the remaining 6% surface as low-confidence rather than as silent misses.

ArcLend put the saving at approximately $1.2M annually in analyst time, and redeployed the capacity into larger deals rather than reducing headcount.

The evaluation harness is the part their team values most. When a new model version ships they can answer whether to adopt it in an afternoon, with a number rather than an opinion.