Find your biggest AI opportunities in under 30 minutes.Book a consultation
Back[ Coventa ]Case Study

[ RESTAURANT OPERATING SYSTEM ]LIVE

Prime POS

An AI-native point of sale where the restaurant is run from intent, not menus. Roughly 60 collaborating agents cover menu, inventory, purchasing, customer management and floor operations as one system — the operator states what they want and the change propagates across every station.

PRIME // SERVICE — LIVE FLOOR

// Problem

The Problem

A restaurant runs on four systems that do not speak: the POS, the kitchen display, the inventory sheet, and three delivery-platform tablets. When the kitchen 86s a dish, someone has to remember to pull it from the POS, the aggregators and the printed specials — and usually doesn't. Margin leaks in the gaps: stock counted weekly, reorder by instinct, menu pricing set once and left, and no view of which items actually make money once wastage is counted.

  • A single operational change requires the same edit in four places; one gets missed every service.
  • Inventory is reconciled after the fact, so shortages are discovered at the pass, not at the order.
  • Purchasing runs on habit rather than consumption, tying up cash in the wrong stock.
  • Item-level profitability after wastage is effectively invisible to the operator.

// Overview

Prime POS collapses the restaurant stack into one agent-operated system. An operator's stated intent is decomposed and executed across menu, inventory, purchasing, pricing, CRM and the floor simultaneously — a single instruction reaches every surface that needs it, including connected delivery channels. Consumption is deducted against recipe-level bills of material in real time, which makes stock position live rather than weekly and lets reorder agents act on actual depletion curves. Customer history feeds targeting, and item-level margin is computed net of measured wastage rather than theoretical yield.

// AI System

Why AI

The value is not natural language for its own sake — it is that one intent has many correct consequences and a human will not execute all of them mid-service. That requires a system that can infer the full blast radius of an instruction and act across it. Language is simply the fastest input device for an operator holding a tray. Beneath it, the agents are doing conventional but coordinated work: recipe-level deduction, depletion forecasting, channel synchronisation.

GroupCountDoes
Menu & pricing~12Item availability, channel sync, price and margin management
Inventory~14Recipe-level depletion, live stock position, wastage capture
Purchasing~10Consumption forecasting, reorder point, supplier order drafting
Customer~12Order history, segmentation, targeted offers, feedback triage
Floor & fulfilment~12Order routing, KDS coordination, delivery-channel state

// Specs

Specifications

INTERFACE
Prompt-driven; full touch POS retained
ORCHESTRATION
~60 collaborating agents
INVENTORY
Recipe-level BOM deduction, real time
CHANNELS
In-house + connected delivery platforms, synchronised
MARGIN
Item-level, net of measured wastage

// Features

Features

  1. 01One stated intent propagates across menu, POS, kitchen display and delivery channels simultaneously.
  2. 02Stock depletes against recipe-level bills of material as orders fire, not on a weekly count.
  3. 03Reorder proposals generated from actual consumption curves and lead times, drafted for approval.
  4. 04Customer segmentation and targeted offers built from order history without a separate CRM.
  5. 05Item-level profitability computed net of captured wastage, surfaced per service.
  6. 06Full touch POS retained underneath — the prompt layer is additive, staff are not retrained.

// Architecture

Architecture

OPERATIONS FLOW

Runtime · one item, left to right


  1. 01Operator Intent
  2. 02Intent + Scope Resolution
  3. 03Blast-Radius PlannerMenu & PricingInventoryPurchasingCustomerFulfilment
  4. 04State Reconciliation
  5. 05Floor / Kitchen / Channels

dashed = the inference step, where the system exercises judgment

System stack

Data in · decisions out

01

Sources

The four systems that never spoke

POS terminalsKitchen displayDelivery aggregatorsUber · Zomato · SwiggyInventory & purchasingReservations & web

02

Ingestion

One event log

Order event streamsub-secondAggregator APIsmenu · stock · ordersStock depletion modelrecipe-levelState reconciliation

03

Ontology

Restaurant state as objects

Order · TicketItem · Recipe · StockTable · CoverChannelPrice rule

04AI

Intelligence

Predict and propagate

86 predictiondepletion vs paceKitchen pacingfire-time optimisationMenu propagation engineconflict-free, reversiblePurchasing suggestionsEval suitesync correctness

05Human

Human control

The manager runs the floor

Manager confirms 86 / priceReversible propagationChannel-level overridesShift audit

06

Actions

Written back

All channel menusPurchase ordersKitchen tickets

Observability

Every model call traced; evals run on real cases, not anecdotes.

Governance

Entitlements enforced at retrieval; rules versioned by the organisation.

Write-back

Systems of record are written only through the approval gate.

Every propagation is reversible; state reconciliation resolves conflicts between channels before commit.

// Impact

Impact

~60
Agents in productionverified
One
Systems of record, previously fourdesign intent

Interested in Prime POS?

Let's discuss running your floor from intent.

Get in touch
// End of case studyPrime POS