Stock Control · Orders & Dispatch · Agentic AI — one order record from the airline PO or phone call to the truck, live stock that explains itself, ten always-on bots, a dashboard for every manager, and people keeping every decision.
The workflow sheet shows two lanes sharing one kitchen and one stock: SkyCrest airline orders and Annie Makes Cakes non-airline orders. They differ at intake and at dispatch — the middle is the same.
The order — not the spreadsheet row — is the stateful object. It survives invoicing, allocation, shortages, scheduling and dispatch, and every transition emits an event the bots subscribe to.
Allocation reserves stock (available-to-promise drops). Live stock only falls on dispatch, through normal OUT transactions with the order number — so every unit is explainable.
Not enough stock → airline shortfall shows on Alex’s Production Planning; non-airline triggers “Inform Alex”. When product is made and stocked IN, the Allocation Bot re-proposes.
Eight modules running on the existing single-file PHP app. The new module builds on the same stock ledger and planner, without touching existing data, users or passwords.
Sessions, hashed passwords, admin / user roles.
Search, product card, stock IN / OUT in units or boxes, history.
Catalog with category + subcategory filters, admin edit, set live stock.
Upload Stock Control.xlsm → preview → apply.
Airline order matrix vs live stock, to-produce, snapshots, print.
16 sizes, as-at date, date sort, colours, NEED TO MAKE.
Add / edit / delete, last-admin protection, change password.
Stock levels and transaction log as .xlsx.
Intake → invoice → allocate → schedule → dispatch.
Alex, Annie, Ruth, Mark, Directors — each their own home screen.
Ten bots incl. Hermes reminders & calls, approval inbox, kill switches.
Each manager lands on their own home screen with today’s work. The app knows who has looked — and the Hermes Reminder Bot chases anyone who hasn’t.
Make-list for quiche, pies, croustades, scrolls · shortfalls blocking dispatch · weeks of cover · menu-change watch
Sweets make-list · tartlet NEED TO MAKE · Annie Makes Cakes orders & bespoke enquiries
Bot inbox · invoices for both Xeros · confirmations · tomorrow’s bookings · team seen board
Exceptions board · on-time dispatch · orders at risk · Hermes escalations · week vs average
Monthly trends · customer concentration · product mix & menu cycle · compliance · risks (read-only, weekly summary)
Stops the moment the dashboard is opened or the person replies OK. One call per person per day, never outside 06:00–18:00. Directors get a Monday summary instead. See the dashboard mock-ups →
216 POs from Jul 2025 to Apr 2026 show airline orders arrive as system-generated purchase-order emails from dnata and Gate Gourmet, with a clear rhythm per kitchen.
121 POs · Monday-heavy · small drops (17 ctns)
Wednesday dispatch · 33–36 ctns · Roadmaster
Rare but bulky · ~44 ctns, 5–8 products
Menu turns over at new year · 12 codes stopped
Always-on bots over a shared event stream, an approval inbox and an audit log. Every action lands on the dashboard as one plain-English line. Every bot has a kill switch — and starts switched off.
Airline POs from dnata and Gate Gourmet are system-generated emails → fixed template parsers. Non-airline emails, PDFs and phone notes → AI extraction. Creates a draft order with a confidence score per field.
Calculates available-to-promise per line, suggests the earliest delivery date for non-airline orders and pushes airline lines into the Production Planning matrix.
Picks the right Xero — SkyCrest or Annie Makes Cakes — and prepares the invoice lines. In V2 creates a draft invoice through the Xero API.
Drafts the customer confirmation email from a template with lines, dispatch date and delivery method, ready for the office to send.
Alex’s assistant. Builds a daily make-list by date and product from order shortfalls, planner “to produce” and tartlet NEED TO MAKE; alerts on non-airline missing items.
Proposes allocations earliest-dispatch-first, re-checks short lines the moment stock comes in, and flags orders competing for the same stock.
At 14:00 the day before: Interstate by port → Roadmaster, Sydney → driver run, Non-airline → all customers. Drafts booking emails and lists orders not ready.
Nightly checks: negative stock, dispatched-but-not-deducted, unusual manual adjustments, app vs XLSM differences after a sync.
Runs on Hermes Agent. 07:00 to-do list per person; anyone who hasn’t opened their dashboard by check-in time gets WhatsApp → SMS → a Twilio phone call → then their manager is told.
Learns each kitchen’s rhythm — SYD every 3–4 days, MEL/BNE weekly, PER ~3 weeks — flags expected orders that haven’t arrived and forecasts demand per product for the make-list.
One cron line every five minutes; the dispatcher decides which bots are due. All times Australia/Sydney.
Bots execute, people decide. Every automated action sits on an explicit autonomy level — and autonomy can only be raised deliberately.
Bot may show information; a named person acts.
Dispatch · allocation · invoices · customer promises · stock adjustments · cancellations
Nothing happens until someone clicks Approve.
Draft orders · proposed allocations · confirmation emails · booking drafts · make-lists
Reversible, no stock or money effect.
Reminders · digests · shortage flags · reconciliation · planner refresh
Every approval records decided_by and time.
Bots create proposals, flags and messages only.
Low confidence goes to the human queue.
Admin → Agents, per bot and per action.
Raising a level needs admin change + owner sign-off.
Permanently human decisions.
Rejected proposals capture a reason and are not re-raised.
Waiting proposals show up like overdue orders.
Each phase ends with its checks in the Testing guide before the next begins. Before every deploy: back up stock.db, never touch users or passwords.
The full project pack, written for both people and AI coding tools.
Both order lanes and dispatch as one flow chart.
Five manager dashboards, seen tracking, message → call ladder.
Order source, ports, rhythm, products, menu cycle.
Modules, recent changes, risks, constraints, open questions.
Stack, as-built system, new module, data model, constraints.
Every screen today, both order lanes, where bots fit.
Ten bots incl. Hermes, autonomy levels, rules, schedule, tech design.
Phases 0–7 with deliverables and checks.
Journeys, regression, bots, security, blockers.