Solutions

AI-Powered Invoice Verification
for Lark Expenses

Kopi uses Claude AI vision to read every invoice — PDF or image — and cross-validate amounts, dates, and vendor details against the expense form automatically. Errors are caught before they reach the approver.

🔍

AI field extraction

Claude AI reads invoices and extracts seller name, invoice number, date, line items, subtotals, tax, and total — even from scanned images and handwritten receipts.

⚖️

Amount cross-validation

The invoice total is compared against the expense form amount. Any mismatch — including currency conversion errors — is flagged red for human review.

📅

Date validation

Invoice date is checked against submission date. Future-dated receipts, date gaps exceeding your policy, and suspicious timing are automatically flagged.

🔄

Duplicate detection

Kopi checks if the same invoice has been submitted before. Duplicate claims are flagged before approval — eliminating a major source of expense fraud.

🏢

Vendor matching

Vendor name on the invoice is compared against the expense form. Mismatches — like submitting an Uber receipt for a hotel claim — are caught automatically.

📊

Confidence scoring

Each extracted field gets a confidence score. Low-confidence extractions (e.g., unclear scans) are escalated for human review rather than auto-approved.

Free for Singapore SMBs in private beta

No credit card. No time limit. Set up in 10 minutes.

Built on Mirror

One agent reads every policy doc — across every domain.

Invoice verification is just one rule in Mirror's vocabulary — alongside amount caps, attachment requirements, vendor allowlists, and routing escalations. Every rule traces back to a verbatim line in your source doc.

01 · Onboarding

Docs → spec in 3 minutes

Drop a PDF, form screenshot, or FAQ doc. Five Claude stages extract every field, rule, and routing path with a verbatim source quote. No hallucinated rules.

02 · Runtime

Decisions in 4ms

runTemplate(spec, request) → Decision is pure code, no LLM at decision time. The spec is data, the runtime is deterministic. Three-color verdict with full routing trace.

03 · Channels

Web · Email · Slack · Lark

A submission can come from anywhere. Reviewers see the same triage card across all surfaces. Web + email is the default; Slack and Lark are accelerators for already-connected tenants.

Citation contract — non-negotiable

Every field, rule, and routing approver Mirror emits carries a verbatim quote from your source document. If she can’t cite it, she drops the item. No fabricated policies — your AI is anchored to your own words.

Read the docs →Architecture, full pipeline detail, page map.