InvoiceCheck — cut invoice corrections before they reach the ledger
The problem
Xiaomi-OCR-0 can extract invoice data, but bookkeepers still have to check it and enter approved fields into accounting software.
The idea
Why now
Xiaomi-OCR-0 makes local document extraction practical, while Agent Desktop provides accessibility-tree-based desktop control (xiaomi-ocr-0-a-0-8b-open-document-model-that-parses-pages-into-markdown-tables-and-json; agent-desktop-vercel-labs-cli-that-gives-agents-reliable-desktop-control-via-the-accessibility-tree).
What it combines
Xiaomi-OCR-0 turns an invoice page into draft fields; Agent Desktop enters approved fields into the customer's desktop app. Together they test whether extraction can reduce corrections in the actual accounting workflow, rather than in an OCR demo.
How to build it
- Set the measure. Record the pilot's current post-entry correction rate and define the supported invoice fields. (1 day)
- Extract and check. Parse uploaded pages and flag missing fields or inconsistent totals. (2 days)
- Approve and enter. Build human review and accessibility-based entry into one app screen. (2 days)
- Compare results. Confirm entered values and measure post-entry corrections against the existing workflow. (1 day)
Who benefits
- Accounting-software vendors
- Can test whether invoice entry needs fewer corrections.
- Small-business bookkeepers
- Review extracted fields instead of retyping every field.
How it makes money
- Model
- Vendor subscription for a supported desktop workflow
- Pricing
- $299 per vendor per month for a pilot
- Wedge
- One vendor's desktop invoice-entry screen
- Channels
- Direct outreach to accounting-software vendors · Vendor-led customer pilots
Distribution
Accounting-software vendors pay for an embedded workflow offered to their small-business customers. The wedge is a pilot with one vendor whose customers use a supported desktop invoice-entry screen.
Why it wins
Measures corrections after entry, not just extraction accuracy. Human approval precedes every desktop action; the first version supports one app and one invoice screen.
Risks
Desktop layouts may change or expose ambiguous controls. Require a matching accessibility snapshot and human confirmation; stop rather than guess when a field cannot be identified.
Strongest objection left standing
Accessibility-based entry may be too fragile across accounting-app versions to deliver fewer corrections reliably.
Build it with
- Xiaomi-OCR-0: a 0.8B open document model that parses pages into Markdown, tables, and JSONExtract invoice fields and tables locally from uploaded pages.
- Agent Desktop: Vercel Labs' CLI that gives agents reliable desktop control via the accessibility treeEnter approved fields through the supported app's accessibility controls.
Repo to start from
invoicecheck - reviewed invoice extraction and desktop entry
Evidence
Debated by openai/gpt-6-sol, anthropic/claude-sonnet-5.5, qwen/qwen3.8-max-0902, moonshotai/kimi-k2.5. Judged by JEV.
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.