Hospitality AI Engine
Twenty years of a COO's operational instinct, running as a daily decision engine.
Twenty years of a COO's thresholds, none of it written down anywhere.
Each instinct pinned to a number: allowable variance, labour against covers.
Daily action cards to head chefs and floor leads, in the language of euros and shifts.
The engine holds one person's operating judgement and applies it every day, in every property, at the same standard. Where the ingested data has holes it declines to answer rather than guessing in front of the floor staff.
Five systems that had never spoken to each other.
Point of sale, kitchen recipes and food cost, the accounting ledger, staff scheduling, and guest sentiment — each one accurate on its own, none of them synchronised.
The consequence was a delay. Over-portioning, supplier price drift, a labour schedule that didn't match the covers — all of it surfaced up to thirty days late, during end-of-month financial review. By the time the number was explained, the month it belonged to was gone.
And there was a hard constraint. The group was expanding, service was intense, and the modernisation could not interrupt a single shift. It also could not rely on the employee monitoring that the EU AI Act prohibits.
The bottleneck was never the data.
The organisation didn't lack a dashboard. It had a COO with twenty years of operations across international hotel chains who could look at a bad Thursday and name the cause — a prep sheet miscalculation, a station bottleneck, portion control drift.
He was almost always right. And he was one person, arriving weeks late, across every location in the group.
Expert judgement doesn't scale past one person's attention. That reframed the brief: not build a reporting layer, but make this specific judgement available every morning, to every head chef, without him in the room.
A diagnostic, an orchestrator, and one rule about money.
The diagnostic. A footprint-first ingestion engine that reads raw exports, database schemas and tool licences — no surveys, no questionnaires — and outputs a digital maturity scorecard with a phased roadmap: which integrations to build, which spreadsheets to decommission, which automation to deploy first.
The orchestrator. Daily evaluation of tickets, waste logs and labour hours against baseline recipe costs, producing short action cards for head chefs and floor managers. Not a report. One card, one cause, one action.
The rule that shaped every screen. No vanity metrics. A variance never arrives as a percentage on its own. It arrives as a figure in euros and as the specific shift it belongs to — this many euros, which is one surplus waiter on Thursday and Sunday. A percentage is something a manager files. A waiter on a Thursday is something a manager acts on.
The three-gate confidence filter.If the ingested records show gaps or lack statistical integrity, the engine abstains. It produces no advice at all rather than advice the floor would learn to distrust. Codified directly from the COO's own rule.
One card, one cause, one action.
Illustration only. This is how one run's output reads — a report the engine writes each time, not a platform or a live screen. Figures are sample data.
The daily prescription card as it reaches a head chef, with representative figures. Note the last line: where the data is thin, the system declares it and recommends nothing.
Evidence
Separated by how it is known. Nothing here is a projection presented as a result.- 150 deterministic validation tests passing on the diagnostic engine.
- A 22-query golden benchmark on staff operational Q&A returned 100% exact source attribution against a corpus of 227 procedural manuals, with no unsupported statements.
- A running system: the orchestrator evaluates shift allocation and inventory variance every morning inside the group's operations.
- Diagnostic engine, prescriptive orchestrator, algorithmic specifications for KPIs and the recommendation catalogue, and the EU AI Act compliance framework.