The Challenge
DERAG AG, headquartered in Munich, operates Living Hotels — apartment and business hotels in ten cities across Europe. A large share of their bookings arrive through online travel agencies like Booking.com and Expedia, and every OTA booking carries a commission claim that has to be reconciled against the hotel’s own invoicing.
In theory, that’s simple arithmetic: invoice times commission rate. In practice, it’s a reconciliation puzzle. One reservation can produce several invoices. Invoices get amended, so multiple revisions of the same document exist. Some invoices cover only extras — parking, minibar, breakfast — that aren’t commissionable at all. OTA records sometimes lack the booking reference, and claimed amounts rarely align perfectly with what the hotel actually billed. Until recently, working through this stack of edge cases was manual finance work: find the matching records, compare the amounts, decide whether the claim is correct, escalate the odd ones.
The obvious automation — point an LLM at every invoice and let it match — raises an uncomfortable question. An LLM will always produce an answer, plausible or not. But a finance team can’t file “the model said so” under an audit. What was missing was a principled way to know, case by case, when the machine can decide alone, when human judgment is required, and when the two should work together.
Why NOMYO AI
NOMYO built DERAG a commission calculation pipeline that runs as a scheduled batch job: it pulls invoices from the property management system, groups them with their OTA commission records, matches and calculates — and, crucially, decides for every single case who should decide. The routing brain of the system is hybrid-groundworks, a framework NOMYO implemented from the mathematical paper Uncertainty and Relation — Groundworks to a Theory of Hybrid Systems.
The framework models every case as a point in a 2D epistemic space, defined by two independent signals:
- α (structural predictability) — how well-structured is the case, judging from the data alone? Is there an external booking ID? A single invoice or five? Do the amounts align?
- H (semantic entropy) — how uncertain was the LLM while it reasoned, measured from token-level logprobs and its reasoning trace?
Two curves — the machine curve and the human curve — divide this space into zones, and they intersect at φ = (1+√5)/2 ≈ 1.618, the golden ratio. That’s not a tunable threshold someone picked in a meeting; it falls out of the mathematics as the point where machine inference and human judgment cost exactly the same. Each zone maps to a routing decision: run the algorithm, accept the LLM’s answer, flag it, or hand it to a human.
Phase 1: Measuring the Case Before the AI Runs
Before any LLM call, the pipeline computes α from the invoice’s structural features: does it have an external ID, is it a single invoice, how far do invoice and OTA amounts deviate, is a service line-item breakdown available for verification? A textbook case — reference ID present, one invoice, zero deviation — scores α ≈ 2.9, deep in the BLUE zone, and never touches the LLM at all. It’s resolved with a deterministic calculation. At the other extreme, a missing booking reference pushes the case into the RED zone, where no amount of model confidence substitutes for a human who picks up the phone.
Phase 2: Measuring the AI While It Thinks
For everything in between, the LLM runs — on NOMYO’s encrypted inference endpoint, which returns per-token logprobs and the model’s reasoning trace. From these, the pipeline measures H: token entropy with spike detection on genuine decision points, combined with structural failure signals. The two signals then fold back into the geometry via α_eff = 1 + 1/H, and the case lands in a zone with a routing label — accept, light_review, flag, or human_handover. The model isn’t asked how confident it is; the system measures how uncertain it actually was while working.
Ground Truth as the Final Arbiter
Routing doesn’t end the story. After the LLM responds, a fully deterministic check verifies the result against the hotel’s own service records: does the summed room revenue match the amount the OTA paid commission on? When a case the model was visibly uncertain about passes this arithmetic check, it’s marked auto_resolved — accepted, but with a preserved audit trail showing the LLM wavered and the ground-truth data settled it. And the LLM itself works with a calculator tool for exact arithmetic, so corrected commission factors are computed, never estimated.
The Results
Running the pipeline over a full batch — 317 invoices across 178 reservations, processed against NOMYO’s inference endpoint — produced a precise division of labor:
- 15% of invoices never needed an LLM — structurally clear cases resolved algorithmically, saving roughly 27 minutes of model processing per batch
- 42% were accepted outright — LLM-processed with measured low uncertainty
- 9% were flagged for human verification before filing
- 32% were escalated to humans — mostly genuinely incomplete inputs, like real room revenue with no OTA commission record to match against
- 1% landed on the golden handover point φ ≈ 1.618, where human and machine contribute equally
The structural signal proved predictive in practice: successfully matched invoices average α = 2.46, while unmatched ones average α = 1.64. The finance team now works a triaged review queue instead of raw data — every escalated case arrives with its measured uncertainty and the concrete reason it was escalated. And every routing decision is recorded, making the AI’s behavior auditable line by line.
Knowing When Not to Trust the Machine
Most AI automation stories are about making a model smarter. This one is about making the system honest: it knows the difference between a case that is mathematically simple, a case the model handled with confidence, and a case where the model was quietly lost — even when its formatted answer looked fine.
That is the core thesis of the hybrid-groundworks paper: hybrid human-machine systems don’t need bigger models, they need a principled handover. NOMYO AI turned that theory into production infrastructure — from the epistemic routing framework to the encrypted inference endpoint that serves logprobs and reasoning traces — and it now runs DERAG’s commission reconciliation as a daily, measurable operation.
DERAG AG is headquartered in Munich, Germany. Learn more at deraghotels.de.