AI hit 55% median accuracy on real QuickBooks month-end checks — guardrails, not prompts, fixed it
Samtyang · reddit · 2026-11-13
A developer building accounting automation for trade contractors tested frontier and open-weight models unassisted on real anonymized QuickBooks month-end close checks: median raw accuracy was 55%, with subtle handoff traps (stale invoice totals, unlabeled supplier bills auto-coded to overhead, deferred-vs-recognized revenue confusion) rather than wild hallucinations. Prompt tweaks barely helped; what worked was moving math and permissions out of the prompt — code verifying every dollar figure against underlying workpapers, read-only analysis with staged proposed entries showing before/after balances, and human approval plus live read-back checks before anything touches the ledger. The author solicits other production-grade hard stops from ERP/field-service agent builders.
More from coding & agent
- Open-source flight search MCP server refuses to answer what it can't confirm, curbing model hallucination — Fast_Entrance7839 · 2026-11-13
- Amplitude launches Wave: AI that decides what to build next from product data — testingcatalog · 2026-11-13
- jxnlco: when AI makes you faster, go deep on new things and grow your taste — jxnlco · 2026-11-13
- Vibe-coded space-defense game ORBITAL SIEGE shows Three.js gameplay from one prompt — kimmonismus · 2026-11-13
- AI-generated WebGL2 shader bends light around a black hole on the first try — kimmonismus · 2026-11-13
- Sententos puts a personal agent in your Mac's notch that can take over your screen — ycombinator · 2026-11-13