Every planning vendor offers AI now, and the promised results are often little short of miraculous. Dmitry Goykhman, founder and CEO of ANT USA, set out to find what the independent evidence says about AI in the work planning teams do every week: merchandise financial planning, long-range sales forecasts and open-to-buy (OTB). He read the research, not the brochures.
Here is what he found. AI’s track record gets weaker the further ahead it forecasts and the less often a decision repeats. Most of merchandise planning sits at that weak end. But planners’ own adjustments aren’t reliably better than the model.
So the best-supported job for AI in planning today is not writing the plan. It is checking the plan. The planner still decides.
Where planning work sits on the evidence curve
The further ahead a task looks and the less often it repeats, the thinner the independent evidence gets.
Strong evidence
- In-season reorder and replenishment. Days to weeks out, thousands of repeat decisions. AI forecasts item by store, and peer-reviewed studies back it.
- Plan diagnostics. Plan vs. actual, forecast bias, and whether planners’ overrides help. AI shows where the plan and the planners go wrong. The method is well proven.
Emerging evidence
- Pre-season forecasts for new items. One season, little or no history. AI finds lookalike items and estimates from product attributes. Accuracy is moderate.
Weak or no independent evidence
- Long-range sales and margin plans. 6 to 18 months out, a few decisions a year. AI helps with baselines, scenarios and sanity checks. No independent evidence was found.
- Open-to-buy management. Rolling, in-season. AI can flag drift early and re-forecast the inputs. The only results published are vendor claims.
- Planning “agents” that change the plan. AI is useful for summaries and explanations. Letting it change the plan on its own is unproven.
What the paper covers
- Short-range forecasts vs. long-range plans. Why the strongest machine-learning results come from 28-day replenishment forecasts, not season plans.
- Planners own the plan, but they tend to overshoot. What a review of roughly 147,000 forecasts found about human adjustments, and how to measure whether they help.
- Open-to-buy: the math isn’t AI. The watching can be. Where AI earns its keep around OTB, and why every quantified OTB result found was vendor-published.
- Pricing, margin and assortment. A sound method, thin proof.
- Five questions to ask any vendor, including your in-house developers.
About the author
Dmitry Goykhman is the founder and CEO of ANT USA, which makes Buyer’s Toolbox. That is a commercial interest in this subject, and the paper says so. It holds every vendor, ANT USA included, to the same standard of evidence.
