An in-house AP module for RevParPro replacing Innrly's failed AP product: invoices come in by GM forward or upload, a vision model reads and codes them, the GM approves on screen, and bills post to QuickBooks (or a tracked lane for the 12 active entities that run on the Excel sheet today). Built role-based because it ships as a product for management companies; THM is tenant #1. Payment execution stays out of V1.
Basis for the build call: ~144 QBO bills a month across HOURP/HOUZN/HOUUS (12 months of QBO data) plus ~75 a month on the manual tracking sheet, which showed $611,165.81 past due when parsed on 7/20. The research verdict on Innrly stands: sound structure, obsolete extraction engine (their own emails admit per-vendor regex parsing), which is the layer we invert with vision extraction.
The backtest is Ace's revision: hundreds of historical invoice PDFs (QBO attachments plus the tracker's Drive folder) already have known-correct answers in the books, so extraction accuracy gets measured in days, not discovered during a pilot month.
| Decision | Locked call (Ace, 2026-07-20) | |
|---|---|---|
| 1 | Approval chain | GM approves own property's bills; org (Ace/Alyssa) co-signs above threshold. Accounting facilitates, never authorizes spend. |
| 2 | Threshold | $2,500 default. Calibrated same day: 3.6 org co-signs a week at full rollout, 2.1 during the pilot (12-month bill distribution). |
| 3 | Accounting's touch | Exceptions only: low confidence, new vendor, math mismatch, suspected duplicate. Clean bills skip accounting. |
| 4 | Stalled approvals | Remind at 48h; escalate at day 5 or 3 days before due date, whichever first. |
| 5 | Notifications | Daily digest; escalations break through individually. |
| 6 | Pilot shape | HOURP August, GM Michael approving from day one, manual entry parallel. |
| 7 | Unreadable invoices | Accounting keys them in-app beside the image; nothing bounces to vendors. |
| 8 | Auto-post | Schema supports trusted-vendor auto-post; OFF in V1 until pilot data proves accuracy. |
| 9 | Intake | Per-property internal addresses + smart generic. GMs forward or upload; vendors keep emailing GMs as today. |
| 10 | Tracked lane | The 12 active sheet entities onboard right after the pilot month; the Excel tracker retires. |
| 11 | Near-term extras | Org-wide AP aging dashboard + missing-bill alerts. Later: statement reconciliation, exports. |
| 12 | Roles, not people | Submitter, AP Reviewer, AP Approver, Org Admin, Vendor Steward, GL Map Manager. Every tenant assigns its own humans. |
Baked-in invariants no tenant can disable: nobody approves a bill they submitted or coded; editing amount or vendor resets approval; when the same person captures and approves, accounting adds one verification touch; delegation is time-boxed and audited; one-person shops get an honest control-exception stamp, never fake segregation.
| Reviewer | Verdict | Top catch, and what changed |
|---|---|---|
| Warren (CFO) | Build with changes | GM both forwards and approves: fixed by defining forwarding as capture, not submission, plus a second touch when capturer = approver. Also added: vendor-splitting velocity rule, remit-to fraud diffing, coding-risk gate. |
| Walton (COO) | Build with changes | Nothing fired on silence: intake heartbeat and zero-intake alerts now land in Phase 1. GL map pre-seeded from 12 months of history so August is not an exception avalanche on Asad's first week back. |
| Andrej (GPT) | Build with changes | The QBO function we planned to clone has live bugs (undefined variable, "bill" excluded from its logging contract): posting service gets built fresh. Approval enforcement moved into the database, keyed to bill versions. |
| ChaseAI (Gemini) | Build with changes | A model grading its own confidence hides its blind spots: answered with the historical backtest plus a month of field-by-field comparison against Asad's manual entries. |
| Wynn (CRO) | Build with changes | Three cheap columns (expense category, service period) so AP data can feed cost-per-occupied-room benchmarking and franchise-fee revenue checks later. |
Zero STOP verdicts. Ace's two post-review revisions: (1) prove extraction on historical examples up front instead of relying on the shadow month; (2) shadow window shortens to gates-met, realistically 1 to 2 weeks, because QBO's own duplicate check is exact-match and preference-dependent, so it stays the second net, not the first.