AI strategy · Chocolate Factory in the Rhythm of the Season

Own models instead of yet another dashboard

A dashboard tells you what happened yesterday. A model on your data tells you what the next season brings. Every competitor has dashboards; a model that learned on your sell-out, quality and price data is knowledge nobody can buy.

Three maturity levels of analytics

Most manufacturers stand on level one or two. Level three is a lead that cannot be copied.

Level 1 · Reporting
  • Dashboards and reports
  • Sell-in and sell-out in hindsight
  • KPIs in real time
  • Forecast of the next season
  • Recommendations per item
  • Learning from new weeks
"We know sales dropped. We do not know why, and not what comes next."
Level 2 · Standard tools
  • Ready-made forecast models
  • Fast connection
  • Forecast, generic
  • Understanding of your seasonality
  • Ownership of the model
  • Lasting lead
"The tool works but does not understand what Advent means for our pralines."
Level 3 · Own models
  • Learned on your data
  • Forecast with industry context
  • Weekly relearning
  • Full ownership
  • No vendor lock-in
  • Open calculation path
"The model knows that one particular chain calls off more seasonal figures in calendar week 47 than last year - and says so three weeks ahead."

What an own model gives that no dashboard gives

It is not about technology, but about knowledge that keeps working every week.

01
The lead that cannot be bought

Large competitors use standard systems anyone can license. A model that learned on three years of your 140 items knows patterns no standard system knows: which chain orders three weeks before the peak and which five, which praline grows in spring and which stagnates. This knowledge only arises at your company.

02
The model gets better every week

A dashboard always shows the same thing, just with new numbers. A model learns with every week of new data. After one season it knows your patterns across all items, channels and promotions at once; a planner oversees a handful at once.

03
The model belongs to you

Code, weights, training path: everything stays in your infrastructure. No subscription, no data in someone else's cloud. If the collaboration ends, the model stays and keeps running. No vendor whose business lives on your dependency will offer you that.

04
The planner's knowledge stays in the house

"For this chain always add fifteen percent, they reorder" - that is in no system, only in the head of an experienced planner. A model codifies this knowledge because it is in the data even if it was never written down. It does not resign, does not forget and takes no holiday in Advent.

Cost of the model against cost of not knowing

The value of a model is not measured in forecast accuracy but in euros that are not written off, not lost and not pushed after by express.

€1,960,000
Cost of the forecast error / year
€1,173,800
Savings potential of the series / year
1 %
of net revenue
Gross effect
before implementation costs
One season with a better forecast: €331,500

€171,000 fewer write-offs, €85,500 recovered contribution margin from shelf gaps, €75,000 fewer extra shifts and express freight - the result of module 01 alone.

↳ Thought over three years

In the first year the model learns and delivers part of the potential. In the second it knows every season once. In the third it knows every item, every channel and every promotion, and finds patterns nobody knew of. A model does not wear out, it grows in value. How fast depends on your data situation; the figures of this series are annual values of the sample company, not a commitment.

From the data to the running model

Four phases, each ending with a tangible result. The duration depends on availability and quality of your data; dates emerge in the conversation, not on this page.

Phase 1 of 4
Data review and first proof

Export of three years of sell-in and sell-out. Assessment of data quality. First model in hindsight against your current method. Result: the forecast error of both methods and the potential in euros, calculated on your figures.

Phase 2 of 4
Pilot on one category

Weekly forecast twelve weeks ahead for the most seasonal category. Comparison with the production plan: which volumes could have been planned differently?

Phase 3 of 4
Extension to the modules

The modules your data situation supports are connected: shelf space, promotion, changeover sequence, remaining shelf life, cocoa coverage. Interfaces to ERP and warehouse system, training of planning.

Phase 4 of 4
Operation and feedback

Models in operation, weekly relearning, accuracy monitoring, hand-over of code and documentation to your IT. What the measurement refutes is replaced.

Three reasons for the timing

Three developments that turn a good project into an urgent one.

The cocoa price
After the price jumps of recent years every tonne of overproduction costs more than ever. Every point of forecast accuracy is raw material that is not written off. And because the price is not predictable, coverage discipline counts twice.
The retailer
According to an association survey, a third of brand manufacturers report being affected or threatened by delisting - an industry finding without independent review (source 7 of the methodology page). Whoever reads their sell-out data holds the annual meeting with numbers instead of hope.
The regulation
From 30 December 2026 the EU Deforestation Regulation applies to large and medium-sized companies. Whoever does not know their first placers by then does not have a data problem but a supply problem.

The value of a model lies not in the accuracy it achieves but in the decision it makes possible earlier.

Guiding sentence of this series, not a third-party source
Note on the context of this portfolio

All euro amounts are model calculations. They are derived from clearly named assumptions (assumption base → lever → result) and illustrate the order of magnitude, not a commitment. The total is a gross effect before implementation costs; no module has independent evidence for its degree of effect, the lever ratios are assumptions.

This series shows which questions data science can answer in a chocolate factory. The concrete answers only emerge with your data.

All 9 modules: AI in the chocolate factory