AI in production planning: What does one extra tray cost?

Two plans miss demand by the same number of trays. The better choice depends on shortage costs, recovery value and what the production line can deliver.

At 2 pm on Monday, a production planner must decide how many trays of fresh pears to pack for Tuesday. Sales expects a good day. Two pallets are still in cold storage, but by tomorrow one of them will fall short of a key buyer’s remaining shelf-life requirement. Should the planner schedule another batch? AI can provide a demand forecast. Whether the extra batch makes financial sense depends on what happens to the last tray sold and the first tray left over.

This is an illustrative case. The setting draws on our work in industry software. Our venture Deepico develops cloud ERP for the food industry, where purchasing, inventory and production are closely connected. Before setting a budget for a forecasting pilot, talk to the people who release stock, accept orders and find a home for leftover goods. Deepico and the industry it serves

What does one extra tray cost?

Finance can open the first meeting with two records. The first is the settlement for some surplus stock. After discounts, transport and sorting, how much money did the business actually recover? The second is an order the business could only fill in part. Could the customer switch to a different pack size? If so, treating the whole unfilled order as lost revenue would overstate the damage. The relevant comparison is the alternative sale the business could realistically make.

For a surplus tray, use the additional cost of the goods less their recovery value. For a missing tray, use the contribution margin forgone, plus any substantiated extra costs where appropriate. Losing the margin and paying to expedite a replacement can be alternative outcomes of the same shortage. Adding both by default inflates the value of a better forecast. Ask purchasing and sales to record the basis for each unit cost and the circumstances in which it applies.

Two plans with the same error in units

For a simplified decision, suppose each missing tray costs CHF 4 and each surplus tray costs CHF 1. There is no time for another production run. The table compares two possible plans over two days. We assume demand is fully known for this comparison and goods cannot carry over to the next day. All figures are illustrative.

Both plans have a mean absolute error of ten trays per day. Yet Plan B costs CHF 18 less over the two days. Its drawback is visible too. The business has six more trays to divert or write off. If expensive disposal takes the cost of a surplus tray to CHF 4, both plans cost CHF 80. Changing one assumption wipes out the financial advantage. That is why a forecasting brief needs a price for each kind of error. Research on probabilistic demand forecasting describes ways to weight deviations differently. Research on demand forecasting

Which decisions can the operation actually make?

A forecast of 107 trays is of limited use if the line runs in full batches and needs cleaning after a product change. In the pilot, put every method through the same step. Turn estimated demand into a feasible plan using available stock, batch sizes and committed quantities. Then compare the plans. This reveals cases where an improved forecast still produces exactly the same batch on the line.

Some constraints are better stated in their own units. Sales may have promised minimum availability under a supply agreement. The operation may have a firm limit on avoidable food waste. Set both boundaries before the trial. The cheaper plan can still be unacceptable. The people accountable for the agreement and the production line should make that call.

Where a pilot earns its budget

Start with one product family, one site and a fixed order deadline. Choose a decision the planning team can influence and for which the business has credible records of losses. Ask planners to start logging the quantities they actually release. Add a simple benchmark such as demand on the last comparable weekday. It shows how much value the business could achieve with a straightforward rule. Apply the same operational constraints to every option.

The investment proposal also needs a financial threshold. Suppose data maintenance, model operations and supervision cost CHF 24,000 a year. At 240 comparable planning days, the new process must avoid an average of CHF 100 per day just to cover that running cost. Implementation comes on top. The CHF 18 in our two-day example is not an estimate for this annual calculation. Establishing the attainable daily benefit is precisely what the pilot is for.

What management should approve

The COO names the person who can change a plan before the order deadline. The CFO signs off on unit costs and the financial threshold. The CIO establishes whether the team can reconstruct what planners knew at the time. Agree in advance on the combination of lower cost, delivery performance and surplus that would justify expanding the pilot. Ask for ordinary days and promotions to be reported separately. A favourable average offers little comfort if the missing stock was contractually committed.

Bring a record of a short shipment, a write-off and the corresponding production plan to the first meeting. With those in hand, one session can establish which decision the pilot should improve and how you would recognise an improvement. Review a planning decision with Deep Impact

For a related costing example, see What does a resolved AI case cost?.