Claims & Close: one page

The idea

Expense claims are the flow every employee of every customer touches. Let a model read each item, let code apply the policy on what it read, and measure every candidate model (a small open model Dvar can host, a low-cost frontier model, a mid-tier one) on the same deterministic scorer for accuracy, critical failures, time and cost, before anything ships.

What the demo shows

• Claims: an employee scans, the claim fills itself, a policy note appears before submitting; the approver sees one line with the flags.
• Ask your spend: plain-English questions answered with a line and a chart; the model writes only a query, so it cannot invent a figure.
• Month-end close: a whole synthetic company-month through each model: share closed with no review, critical failures, latency, cost per company-month, the exceptions queue.
• Honest labels: results come from run run-20261022172600-test, an internal benchmark on synthetic data, and show where each model wins or loses. No model has a critical failure on this run, so none is blocked. Fastest: Mid-tier frontier; lowest cost per item: Low-cost frontier.

A four-week evaluation

1. Week 1Pick the steps together (claims reading, GL coding, receipt matching are our suggestion). Agree the policy checks and the critical failures with finance ops. We set up Dvar Evals in your environment and connect it to traces of the current feature.
2. Week 2Build the gold set: a few hundred redacted or synthetic receipts and transactions from your markets, labelled by your finance ops team, with the hard cases you care about. Freeze the instruction on a dev split.
3. Week 3Run your current model, a Dvar-hosted small model and a low-cost frontier model through the same scorer. Tune the small model on its misses (GL codes, local formats) and run it again.
4. Week 4Read the results together: closed with no review, critical failures, latency, cost per company-month. Decide per step: keep, switch, or route the hard cases. The scorer stays as the regression test.

What we need from Aspire

• An owner in finance ops or product to agree the policy checks and the critical failures.
• A labelled sample (redacted or synthetic) from the markets that matter, and the current feature's outputs on it if you can share them.
• Whether receipt OCR happens before the model or inside it (a vision model), so the evaluation measures the step where the cost is.
• A place to run Dvar Evals: your cloud, or ours with synthetic data.