Spotting AI-washing in vendor pitches

Ask for a concrete number. If they say "millions of transactions" but can't specify what those transactions were used for (training vs. just storage in a database), push further.

2. "What happens when you feed it a product category it's never seen?"

This tests generalization, whether the model can handle new situations, not just memorized rules. If the vendor says the system "applies configured business rules" to unfamiliar categories (say, a retailer launching a plant-based meat alternative aisle with no sales history), that's evidence of a rules layer doing the real work, with a thin ML wrapper elsewhere.

A genuine model should degrade gracefully and produce a probabilistic estimate with acknowledged uncertainty, not a confident rule-based output dressed as a "prediction."

3. "Can I see a confusion matrix or accuracy metric from a held-out test set?"

A held-out test set is data the model never saw during training, used to check if it actually generalizes rather than memorizes. A confusion matrix shows where predictions were right or wrong.

If the vendor cannot produce any accuracy metric (precision, recall, mean absolute error for demand forecasts) on unseen data, there is likely no trained model being evaluated at all. Rules engines don't have "accuracy" in this sense because they don't make probabilistic predictions, they execute logic.

Free primer if you want the vocabulary before the meeting: Google's Machine Learning Crash Course glossary.

4. "How often does the model retrain, and on what trigger?"

Real ML systems need retraining as consumer behavior shifts (a classic FMCG example: pandemic-driven pantry-loading in 2020 broke many demand forecasting models trained on pre-2020 data, forcing rapid retraining across the industry, this is well documented in retail analytics case studies from that period).

If the vendor's system "updates" only when their engineers manually edit rule thresholds, that's rule maintenance, not model retraining. Ask specifically: is retraining automated, scheduled, or triggered by a performance drop (this is called "model drift," when a model's accuracy degrades because real-world patterns changed since training)?

5. "Show me a case where the model's recommendation surprised your own team."

This is the sharpest question. Genuine ML systems occasionally surface non-obvious patterns: pairing a premium snack with an unexpected adjacent category based on actual co-purchase data, or flagging a regional pricing anomaly humans missed. If every output the vendor shows you matches exactly what an experienced category manager would already recommend, the system may just be automating existing intuition, which has value, but isn't "AI-powered optimization" in the sense being sold.

A quick technical tell in the pitch deck itself

Look at how the vendor describes assortment scoring. Genuine ML-based assortment tools typically reference something resembling a weighted scoring or ranking model with learned coefficients, not fixed thresholds. A simplified honest version might look like this pseudocode, where weights are *learned from data* rather than hardcoded:

If the vendor's technical appendix only shows the first pattern with different variable names, ask why they call it AI at all.

What legitimate AI use in category management looks like

To be fair, real applications exist. Demand forecasting models using gradient boosting or neural network architectures are used by major CPG (consumer packaged goods) players and retailers, Nestlé, Unilever, Walmart, and Tesco have all publicly discussed machine learning in supply chain and demand planning contexts. Nielsen and Circana (formerly IRI) also embed statistical and ML models into retail analytics products, generally with more transparency about methodology than smaller category-management point solutions.

The distinguishing feature of legitimate vendors: they can discuss model architecture, training data provenancedata provenanceData lineage maps how data moves and transforms across systems, from origin to consumption, showing where it came from, what changed it, and where it goes.Voir la définition complète →, and validation metrics fluently and specifically, because a real data science team built and maintains the product.

Practical evaluation checklist for procurement

Before signing a contract, request in writing:

• Training data description (volume, source, recency)
• At least one validation metric on held-out data
• Retraining cadence and drift-monitoring process
• A reference customer willing to discuss actual model performance, not just satisfaction with the UI

If a vendor resists all four, price the tool as a rules engine (which may still be worth buying, rules engines are often cheaper, more explainable, and easier to audit than opaque models) and negotiate accordingly. Explainability matters especially for regulated claims: EU retailers using automated pricing systems must be able to explain pricing logic under general consumer protection principles, even though no EU AI-specific pricing law currently mandates model transparency at this granular level.

🎬 [VIDEO: "How to Spot AI Washing" - youtube.com - search for recent explainer content from established data science education channels covering AI marketing claims versus technical substance]

Key Takeaways

• The core diagnostic question is whether the system learns patterns from data (trained model) or executes pre-written logic (rules engine); both are legitimate, only one deserves "AI-powered" as a description of predictive capability.
• Demand concrete training data details, held-out validation metrics, and a documented retraining process; vague or evasive answers signal a rebadged rules engine.
• Test generalization by asking about unfamiliar categories or edge cases; rules engines fail predictably, real models degrade gracefully with quantifiable uncertainty.
• Legitimate ML use in FMCG category management exists at major players (Nestlé, Unilever, Walmart, Tesco, Nielsen, Circana) and is usually accompanied by fluent, specific technical discussion, not marketing language alone.
• Rules engines are not worthless, they're often more explainable and auditable, so price and evaluate them honestly rather than paying an AI premium for non-AI logic.