Retail Agentic Operating Model — Single-Unit Demo (Single Store)

Single Unit — one CO-PILOT agent in the store assistant’s pocket.

The portfolio demo shows what the CEO sees every morning. This page is the other end of the same operating model: the daily ranked checklist that the assistant at Penang Gurney Mart #1 sees on their phone before opening the shutters. One agent, push-only, no dashboard. Most Malaysian retailers should start here.

Data inputs CO-PILOT subscribes to

Twelve live feeds across POS, IoT, roster, vendor, weather, regulator, comms and competitor signals. The agent re-reads everything every 60 minutes and re-ranks the action list before the next push.

Today’s store signals

Six derived signals the agent uses to score every candidate action. These are what the agent saw at 06:00 this morning, on the day you’ve selected in the tab strip.

Daily action checklist

Push-only. The assistant doesn’t pull a dashboard — these tasks land on their phone in this order. Mark them off as you go to feel the loop.

0 of 0 tasks complete

Today’s shift

Today’s CO-PILOT run log

The agent runs on a fixed schedule, not on demand. Each entry is a re-rank pass: inputs refreshed, signals recomputed, checklist re-ordered. Greyed rows are the rest of the day still to come.

Manager WhatsApp comms (parsed)

The CO-PILOT subscribes to the store manager’s WhatsApp group via the WhatsApp Business API. It extracts intent, entity and required action — and folds them into the next checklist re-rank.

CO-PILOT activity log

Each row is a writeback from agent_writebacks.jsonl. The agent narrates what it noticed and what it pushed, so the manager can read the trail.

The phase ladder

Single Unit · Phase 1 (you are here) — one CO-PILOT, one store, push-only.
Phase 2 — Inventory + Pricing agents turn on; 8–12 store pilot; manager dashboard.
Phase 3 — Staffing + Facility agents; chain-wide.
Phase 4 — Portfolio agent + Chief of Staff; CEO sees the daily Decision List (the other demo).

How this was built

The CO-PILOT agent reads three files of mock store telemetry — the same files that drive the portfolio demo. We filter to a single site_id (SITE-0001), pick the latest three operating days, and let the agent generate its checklist deterministically from the day’s signals (stockouts → restock task, chiller breach → escalation, peak footfall forecast → mid-day reset, etc).

This page is intentionally tiny: about 200 lines of JavaScript, zero frameworks, one HTTP fetch per data file. The point is the operating model, not the chrome.