Work

Case study deck · 9 slides

Food manufacturing

Demand planning that runs ahead of demand

A food & baking company

Audit & TeardownProcess Redesign & BuildAdoption & Enablement
01 / 09

The challenge

What was broken

The demand plan was rebuilt from scratch every quarter, always looking backward, and took the planning manager three weeks to compile from three disconnected systems. Pumpkin spice loaves ran at the summer average and were still baking in September; a dinner roll SKU stocked out three times during a festival promotion the historical average never saw. Waste and lost sales ate margin from both directions at once.

Days between forecast compile and a production-ready plan, by quarter
02 / 09

The approach

Scored on the Durable AI Index

We put three years of shipment data, promo dates and seasonal windows into one SKU-week table, which showed the old averages were smearing promo spikes and seasonal cliffs into one flat number. We built a forecast treating seasonality and promotions as separate effects and ran it in shadow mode for six weeks, catching that festival-driven spikes were underweighted, before switching the planning meeting over to it.

DURABLE AI INDEXImpact87Feasibility79Stickiness8344at startafter redesign
High impact, low stickiness. The gap was the workflow, not the model.
03 / 09

What we built

The system, not just the model

We redesigned the workflow around the model, then built the pieces that make it run every day.

  1. 1

    SKU-week demand model with separate seasonality and promo terms

  2. 2

    Regional-event flagging for festival-driven spikes the average missed

  3. 3

    Weekly forecast-review view with flagged exceptions by SKU

  4. 4

    Feed pushing the approved forecast into batch scheduling

04 / 09

Inside the build

From quarterly compile to Monday forecast

How the planning cadence itself changed, from a three-week manual rebuild once a quarter to a weekly forecast review production now runs on.

Cycle time: 3 wks to 90 minForecast accuracy: +25%Excess inventory: -28%

Tools & systems

SKU-week sales history + seasonality modelregional-event / promo flaggingweekly forecast-review workflowproduction & inventory linkageexception-override interface
  1. Manual quarterly compile

    Planning manager pulls shipment history and builds SKU averages by hand across three disconnected systems, taking roughly three weeks per quarter.

  2. Flat-average production plan

    Production runs off a single quarterly number that blends seasonal spikes and ordinary demand, driving both overproduction and stockouts.

  3. Shadow-mode validation

    New SKU-week forecast model runs alongside the manual plan for six weeks; a regional-event flag is added after festival-driven demand spikes are found underweighted.

  4. Model goes live in planning meeting

    The forecast becomes the starting point for the weekly production discussion, replacing the manual export as the first draft.

  5. Monday exception review becomes routine

    Planning manager reviews a short flagged-exception list each Monday instead of rebuilding the full plan, with production pulling numbers directly off the weekly feed.

05 / 09

Results

What good looks like

25%

Improvement in forecast accuracy with the data-driven demand model

28%

Reduction in excess inventory across SKUs

Less waste

Reduced spoilage, with profit margins improving

3 weeks → days

Time from data to production plan

EXCESS INVENTORY (INDEXED)100Before72After (-28%)
06 / 09

How it stuck

Adoption is the deliverable

Three weeks a quarter became ninety minutes every Monday: the planning manager checks the handful of flagged exceptions and overrides in seconds, everything else rolls through untouched. Production picked it up just as fast, once the batch schedule stopped lagging demand by a full quarter.

050100month 6This engagementTypical pilot
Share of weekly production plans built from the forecast tool vs. reverted to manual spreadsheet
07 / 09
I used to hand production a number I already knew was three weeks stale. Now I hand them Monday's number on Monday, and I decide the five exceptions that need a human.

Demand planning manager

08 / 09

Want this for your business?

Book a free 30-minute AI opportunity assessment. You will leave with at least one concrete idea.

09 / 09