Hybrid AI Safety Stock Control in Supply Chain Replenishment

Prior experiments in this series established that stateless LLM agents cannot outperform simple heuristics when making direct ordering decisions. This experiment tested a different architecture: instead of asking the AI to invent the exact order quantity, a mathematical formula handles that. The AI acts as a planner whose only job is to adjust the Safety Stock Multiplier.

If the formula says 10,000 units and the AI recognises that Diwali is next month, it outputs a multiplier of 1.2×, meaning order 20% extra buffer. The formula executes that quantity. The AI never invents the base number; it only scales it. The hypothesis: confining AI to a single scalar output would exploit its qualitative context-reading while limiting its quantitative weaknesses. The data does not support this.

The data reveals a structural disconnect between two capabilities that look related but are not. Semantic alignment means the AI understands the concept of what is happening. Reading the text outputs confirmed that all three models correctly recognised when a busy season was approaching and deduced that more buffer stock was needed. Operational control means the AI can choose the exact mathematical number required to stabilise the system. This is where every model failed.

An AI that correctly identifies “December needs more buffer” and then outputs 1.45× when 1.05× would have sufficed is not an operational controller. It is a semantic reasoner in the wrong job. Directional capability exists. Numerical calibration does not. The hybrid architecture exposed the gap between them cleanly.

The over-buffering bias was consistent and measurable. With no AI and a fixed 1.0 multiplier, chain OVAR is 1.7097. Every AI condition exceeded this. The models averaged multipliers between 1.13 and 1.48 across all runs and conditions, indicating a systemic tendency toward caution that, paradoxically, produced worse outcomes than doing nothing. When in doubt, the models added buffer. Adding buffer destabilised the chain.

The o4-mini Stateful result deserves attention on its own terms. At OVAR 3.1211 it was the worst performing configuration in the experiment. The internal thinking logs showed the mechanism: the model anchored heavily on recent negative signals. A minor backlog from two months prior prompted a massive over-order the following month, which then appeared in the next period’s history as excess inventory, triggering a different over-correction. This feedback loop is the bullwhip effect running not across supply chain tiers but inside the model’s reasoning process.

The hybrid architecture was a reasonable hypothesis. Restricting the AI to a single scalar output while the formula handled execution seemed like a way to leverage AI’s qualitative strengths while limiting its quantitative weaknesses. The data shows this is not sufficient. A continuous multiplier that the AI invents still requires numerical precision that the models cannot reliably provide.

The next logical step is to restrict the output further: instead of a free-form multiplier, the AI selects from a small set of pre-defined text labels (STRONG_INCREASE, MODERATE_INCREASE, NEUTRAL, MODERATE_DECREASE, STRONG_DECREASE). A hard-coded translation layer maps each label to a fixed, pre-approved multiplier value. The AI’s task becomes classification, not numerical calibration. Whether this architecture works is the subject of the next experiment.

The empirical multiplier range observed here (1.13–1.48 mean across all models and conditions) provides a concrete basis for designing guardrails. Any production deployment of a multiplier-based AI planner should enforce hard limits derived from empirical runs of this kind, preventing the model from accessing the region of the multiplier space where it reliably causes damage regardless of what it reasons.