Tomorrow's production plan, visible before the afternoon
Planning across standing orders, ad hoc trade orders and seasonal demand, where the constraint is late visibility rather than poor judgement.
A wholesale bakery turning over just over a million pounds, running six day production with a rotating range.
A delivered Orchyn engagement. The client is anonymised at their request. The company and the people are not named. The operation, the workflow, the controls and the outcome are as delivered.
The commercial problem
The bakery does not need an autonomous production system. It needs earlier sight of what people already know: last minute changes, customer specific rules, expected demand, ingredient shortages and dispatch constraints.
By the time that information is reconciled by hand, the production manager is already making decisions under pressure. The process holds when the team is stable and breaks around holidays, events and absence.
How the work runs before
- Standing orders sit in one spreadsheet while ad hoc orders arrive by email and phone.
- A coordinator consolidates tomorrow's demand late each afternoon.
- The production manager adjusts the plan from experience and checks ingredients in a separate sheet.
- If an ingredient is short, the team chases alternatives by phone and updates people through messages.
- Dispatch works from a printed or spreadsheet list and checks late changes manually.
- Waste, stock outs and substitutions are reviewed after the shift with little structured learning.
What makes it hard to automate safely
- Allergen control, bake schedules, food safety and final dispatch stay with people. This is not negotiable in food production.
- Forecasts are uncertain and must be presented as recommendations, never as instructions.
- Product names and standing orders use informal shorthand built up over years.
- The business cannot tolerate a system quietly substituting a product or changing a quantity.
- Historic demand was incomplete. The first version had to be useful on modest data.
What we change first
A dependable order cut off was established, along with one source of truth for standing orders, changes and exceptions. Customer specific instructions were converted from messages into structured reviewable records, which is what made them usable at all.
The redesigned workflow
- Orders and approved changes enter one planning board before the cut off.
- The model extracts orders from the approved inbox, maps shorthand to product records and flags ambiguous items for a person.
- A planning layer combines confirmed demand, standing order history, seasonal notes and manual event adjustments into a proposed production sheet.
- It highlights likely ingredient constraints, unusual demand changes and customer instructions needing review.
- The production manager reviews and adjusts before releasing the plan. Allergen, recipe, batch and food safety sign offs stay in the existing controls.
Where the model stops
The model proposes a plan and surfaces constraints. It cannot change a quantity, substitute a product or release production. Every proposal is a recommendation a person accepts or overrides.
What we learned
The value was earlier visibility, not better forecasting. A mediocre forecast seen at midday beats an excellent one seen at five o'clock, because only one of them leaves time to act.