RevParPro · AP Invoice Pipeline

What a hotel AP module looks like in 2026, and whether Innrly's design made sense

Deep research brief for the in-house AP build decision · run 2026-07-20 · prepared by Woz

105 research agents 23 sources fetched 25 claims adversarially verified 20 confirmed · 5 refuted

Verdict: Innrly's structure was right. Its extraction engine was obsolete before they built it.

Innrly's marketed AP workflow (submit, review, approve, push to QuickBooks, vendor profiles, duplicate and fraud identification) matches the canonical 2026 shape used by every leading platform, verified against Innrly's own live pages. Their public materials never mention OCR, extraction method, or confidence scoring anywhere. The per-vendor regex parsers that produced four months of wrong totals and missing line items were a hidden implementation choice, not part of the design.

Ace's instinct is confirmed: failing at OCR in 2026 was a choice. Per-vendor template OCR is considered obsolete across the industry, and in the one head-to-head benchmark found (AIMultiple, updated 2026-01, six tools), a general-purpose vision model beat every dedicated OCR product on invoice extraction accuracy. The in-house plan (vision extraction plus confidence scoring plus arithmetic checks plus human approval) is the current state of the art, not a bet on something exotic.

One more thing the research settled: Innrly's "Pay" stage was marketing. Their real workflow ends at approval plus accounting sync, exactly where our V1 ends. We are not shipping less than what THM actually used.

01The canonical 2026 AP pipeline

Capture
email / scan / upload→ Learned extraction
~95% confidence gate→ GL coding memory
line level, learns from corrections→ Duplicate check→ Approval routing
amount / property / dept→ Payment
cut from V1→ Post to books
near real time, with class/location

Verified consistent across hotel-vertical platforms (Ottimate, Inn-Flow, M3, Nimble) and horizontal leaders (Stampli, Bill.com, Yooz, Ramp). Cutting payment matches Innrly's actual scope; cutting PO matching is defensible because THM runs no procurement system (M3 delivers matching via its Reeco partnership, which presupposes one).

02Scorecard: standard vs Innrly vs our spec

Capability (2026 table stakes)	Innrly (as shipped)	Our spec, pre-research	Our spec, now
Email-in invoice capture	Yes	Yes	Yes
Layout-proof extraction	No, per-vendor regex	Yes, vision model	Yes, vision model
Confidence gate + human review	Promised late, never shipped	Yes	Yes, 95% initial bar
Learned GL coding, line level	Partial, GL-in-description complaints	Yes, vendor_gl_map	Yes
Duplicate detection	Marketed, never answered Junaid	Yes, sha256 + fuzzy key	Yes, + QBO DocNumber layer
Approval routing beyond flat queue	Broken roles, self-approval	Single approver only	Threshold + escalation + delegate
Mobile-usable approval	Not evidenced	Unspecified	Responsive web required
Per-action audit trail	Not evidenced	Yes, 276/277 pattern	Yes
QBO class/location tagging	Descriptions did not even sync	Unspecified	Line-level class, dimension mirror
Source PDF attached to the Bill	Not evidenced	Stored, not linked into QBO	Attached via QBO Attachable API
Post on approval, not batched	Up to a day, sync button broke	Yes	Yes
Payment execution	Marketing only	Out of V1, deliberate	Out of V1, deliberate

Innrly column sourced from the THM email record (thread "AP Module - Innrly", 42 messages, 2026-03-12 to 06-11) and Innrly's live marketing pages, both read 2026-07-20. Spec columns from the AP Invoice to QBO Pipeline Spec in the vault.

03What changed in the spec today

AddedQBO's own duplicate guard as a second layer. The Bill API rejects duplicate DocNumbers (21-character cap, error 6140), but only if the company-file preference is on. We confirm the preference per property, catch the error path, and normalize long invoice numbers. Our own vendor + invoice number + amount check stays the primary defense.
AddedLight approval routing. Still one named approver in V1, but with a dollar threshold rule, an escalation ping to Discord when an approval stalls, and a delegate field. Asad being on leave July 20 to 31 the very week we scoped this is the live proof a delegate field is not hypothetical. The approval screen must work on a phone.
AddedClass tagging on posted Bills. The books use QBO classes today, so extracted lines carry them, and the vendor sync mirrors each property's class and location lists alongside vendors and the chart of accounts.
AddedOriginal invoice PDF attached to the QBO Bill via the Attachable API, so the source document travels with the transaction. The upload mechanics were already proven in-house in June (98 franchise invoice PDFs harvested through the same API).
Already hadDuplicate detection, audit trail, post-on-approval, confidence-gated review. The research confirms these as table stakes; the spec had them before today.
Kept outPayment execution, PO matching, PMS coupling. All three confirmed as defensible cuts for THM's shape.

Cost of the additions: roughly one extra build day across Phases 3 and 4 (Woz estimate). The 8 to 12 day envelope from the spec holds.

04Claims that died in verification

Refuted claim	Vote	Why it matters
Stampli explicitly supports multi-company QBO out of the box	0-3	Buying a horizontal tool would not solve our 3-QBO-company plus 12-tracked-entity world cheaply. Strengthens build over buy.
OCR-plus-templates is still the accepted incumbent in the hotel vertical	0-3	The vertical has already moved to learned extraction. Vision extraction is the mainstream, not a gamble.
M3's six-stage procure-to-pay flow is the canonical model	0-3	It is delivered through a partner bundle and assumes a PO system. Not a template for us.
Nimble ships AP automation as a paid bolt-on	0-3	Pricing-structure claim did not hold. Removed from the analysis.
Innrly's duplicate detection was a working built-in capability	1-2	Marketed on the site, but Junaid asked for it on 2026-03-17 and never got an answer. Treated as unproven.

05Open questions the research could not close

Real-world touchless rate on THM's actual invoice mix. The 95% figure is vendor-stated. The HOURP pilot month is the calibration instrument: run it, measure, then set the review threshold.
How to structure the tracked-only lane for the 12 non-QBO entities. No vendor pattern covers AP without an accounting system behind it; the THM Invoice Tracking sheet is the requirements doc, and we design that lane ourselves.
What hotel operators complain about most in AP tools. Review mining did not survive verification. Anecdotal signal: complaints cluster on payment modules and vendor lock-in, which supports keeping payment out of V1.
Whether Innrly's failure generalizes to other hotel-vertical vendors using similar extraction stacks. Matters only if we ever consider buying again.

06Key sources

Intuit QBO Bill API reference (primary; DocNumber, DepartmentRef, minimum Bill fields)
Innrly features page (primary; marketed AP structure, verified verbatim 2026-07-20)
Ottimate AP automation guide (canonical stage model, 95% threshold, routing dimensions)
Inn-Flow hotel AP explainer (hotel-specific five-stage model, escalation, audit trail)
Stampli QBO integration page (dimension sync, 2h import / 5min export cadence)
AIMultiple invoice extraction benchmark (vision LLM vs dedicated OCR, updated 2026-01-22; small sample, directional)
HotelTechReport hotel accounting category (duplicate detection across top-ranked platforms, updated 2026-05-22)

Method caveats: most structural findings rest on vendor marketing and education pages; cross-vendor consistency substitutes for a neutral survey. The 95% threshold and Stampli cadence are single-vendor statements, corroborated but not independently benchmarked. The AIMultiple benchmark covered 20 invoices and used an older vision model, so treat it as directional. All pages fetched 2026-07-20.