The Forecasting Pilot That Arrived Before the Data
Northbridge Components, a €180m B2B manufacturer, has grown by acquisition and now runs three ERP instances across 14 sales regions. Forecast accuracy for its 1,200 active SKUs sits at 42% MAPE; planners still rely on spreadsheets emailed every Friday. The COO has signed a €240k, eight-week pilot with an AI forecasting vendor, contingent on a cloud data platform that has been stalled for nine months because master data ownership is disputed between sales, operations and finance. The CIO argues that the pilot will fail publicly and damage the wider data programme; the COO argues that another quarter of preparation will lose the budget. Sales directors distrust any model that cannot explain a forecast at customer level. The board wants a decision within two weeks and a measurable result by the next quarter. The vendor contract includes a renewal clause and a reference obligation.
1. How should leadership sequence the AI pilot against the stalled data platform? Recommend whether to proceed, pause or descope, and state the conditions for each.
Read the worked answer
Decision: descope and proceed; do not run the eight-week pilot as written.
Reasoning:
- The pilot tests the model, but the binding constraint is decision rights over master data. A model trained on disputed customer and product hierarchies will produce forecasts nobody can act on, and a public failure will make the next data funding round harder.
- Sequencing is not 'platform first, AI later'. That wastes a quarter. Freeze a minimum viable data product instead: one product family, 24 months of shipment history, one customer master, and a documented demand-signal calendar.
- Renegotiate the renewal and reference clauses. Make renewal conditional on a pre-agreed accuracy lift against the 42% MAPE baseline, and make any reference conditional on written client approval and a published methodology.
Pause if no accountable data owner is named within five working days, if clean 24-month history is unavailable for the pilot family, or if sales will not nominate two planners to validate outputs.
Scale only if accuracy improves for two forecast cycles, planner adoption exceeds 70%, and the vendor provides a customer-level explainability view.
2. Define the minimum data and governance foundation for a credible pilot, and the metrics that would justify scaling.
Read the worked answer
Minimum foundation
- Data contract: sales owns the customer hierarchy, operations owns product/SKU hierarchy, finance owns margin definitions. Each names a steward with 48-hour escalation to the COO.
- Scope control: one product family covering the top 80% of volume, 24 months of history, plus promotions, price changes and known supply constraints. Exclude acquired ERP instances until mapping is proven.
- Quality gates: duplicate customer rate below 2%, missing shipment weeks below 5%, and a documented promotion calendar. Do not wait for 100% completeness.
- Model governance: naive and moving-average baselines, a challenger model, backtesting on the last six months, and human overrides logged with reason codes.
- Metrics: MAPE and bias by family and region; forecast value added versus planner judgment; planner hours saved; stockout and expedite costs; adoption.
- Exit terms: data and model outputs remain company property; no auto-renewal; reference only with approval.
This separates the pilot from the platform programme while creating reusable data products. A pilot that cannot be compared with a baseline is a demo, not evidence.
3. Sales distrust is the main adoption risk. What operating model would make model outputs usable and accountable?
Read the worked answer
Put the planner in the loop by design: the model proposes, the planner disposes, and overrides are recorded.
- Weekly exception review: 30 minutes per region, focused on the top five changed forecasts and why. Publish customer-level drivers, not a black box.
- Shared accountability: accuracy is a shared KPI for the planner and demand planning lead. Track override accuracy separately so judgment is rewarded, not punished.
- Transparency: show the three largest contributing signals and a confidence band. If the vendor cannot provide this, treat it as a contract issue rather than a training issue.
- Co-design: recruit two sales planners as co-designers with veto rights over pilot scope. Train on decisions, not on the tool.
- Escalation: if planner and model disagree by more than 20% for two cycles, the demand planning lead decides and records the rationale.
This converts distrust into a controlled experiment. It also gives sales a legitimate route to challenge a forecast without blocking the programme, which is the difference between adoption and compliance.