Letting AI adjust the safety-stock buffer
After language models struggled to choose order quantities directly, I tested a smaller role for them: adjusting the inventory buffer while a formula calculated the order. The question was whether seasonal information could help the model hold enough extra stock without making orders more erratic.
What the model could change
Safety stock is the extra inventory held above a forecast to protect against shortages. I tested gpt-4.1-mini, o4-mini, and the local nemotron-3-super:120b model as selectors of a safety-stock multiplier between 0.5 and 3.0. An Order-Up-To formula combined the forecast, buffer, and inventory position to calculate the order.
The multiplier affected only the safety-stock term. Choosing 1.2 meant a 20% larger buffer, not necessarily a 20% larger order. If a response could not be parsed, the system used the neutral multiplier of 1.0.
The fictional three-tier chain served one product over 25 months of synthetic demand shaped around Indian automotive seasons. Deliveries took a fixed one month, with no disruption events or random supply delays. The Azure models ran twenty repetitions per information condition; Nemotron ran ten. “Blind” supplied numerical state alone, “context” added calendar and seasonal flags, and “stateful” also supplied the last three periods of demand, orders, multiplier choices, backlog, and stockouts.
I compared the model-assisted policies with two controls. Exponential smoothing provided a simple reference. The fixed-multiplier control used the same ordering formula as the models but always selected 1.0, allowing a direct comparison between an adjustable buffer and a fixed one.
The principal measure was chain OVAR, the ratio of order variance to demand variance at each tier, averaged across the chain. Below one means damping; above one means amplification. Stockouts and inventory were also recorded to show what was gained or lost alongside changes in that ratio.
What adjusting the buffer achieved
| Condition | Chain OVAR | Stockouts |
|---|---|---|
| Exponential smoothing | 0.5446 | 5.0 |
| Fixed multiplier, no AI | 1.7097 | 14.0 |
| gpt-4.1-mini, blind | 2.3325 | 10.6 |
| o4-mini, stateful | 3.1211 | 10.7 |
All nine model conditions had higher OVAR than the fixed-multiplier control. They had fewer stockouts than that control, so the change was a trade-off rather than a loss on every measure. Exponential smoothing nevertheless had both lower OVAR and fewer stockouts than every model condition.
More information did not solve the problem. Context increased OVAR for Nemotron and gpt-4.1-mini and slightly reduced it for o4-mini. Adding history then improved Nemotron and gpt-4.1-mini relative to their context conditions, but only by 0.0783 and 0.2537. The reported criterion for a meaningful history benefit was at least 0.5 OVAR for at least two models. Neither reduction reached that size, and o4-mini became worse with history.
The study also checked whether multiplier choices followed the expected seasonal pattern. The highest pattern score was 0.3977, below the required 0.50. All four stated hypotheses were rejected. The models' seasonal explanations did not translate into the required improvement in ordering.
Conclusion
Moving the arithmetic into a formula did not make model-selected buffers competitive with simple smoothing in this simulation. Adjusting the buffer reduced shortages relative to a fixed buffer, but increased order-variance amplification. Smoothing remained better than every model condition on both of those measures.
This finding concerns the tested multiplier range, ordering formula, models, and single synthetic demand profile. The smoothing parameter was unchanged, and the hypotheses were not formally pre-registered. It is not evidence that memory or every form of model-assisted control is ineffective. The next experiment removed another demand on the model: instead of choosing a numerical multiplier, it chose a label that software mapped to a fixed value.