• When memory lead times move eight weeks in a quarter,
    your safety stock math is already wrong.

VuDecide · Safety Stock Agent

Every quarter, revenue slips on a part nobody was watching.

Not the part on the shortage list. The one your planning system said was covered — right up until the supplier moved the date. DeepVu's Safety Stock Agent finds the parts your current policy is quietly under-protecting, and sizes the buffer for how lead times actually behave.

In production with optical networking manufacturers serving AI data center demand.

Request a Supply Assurance and Procurement Assessment See how it decides →

What goes wrong

The buffer is thinnest exactly when the stockout is most expensive.

52w30w8w Q1Q8 Buffer sized by z · σ · √LT — static Actual lead time under allocation

Illustrative. The gap opens where it costs most: allocation stretches lead times and lifts demand at the same time, so the buffer is thinnest precisely when the stockout is most expensive.

The problem

Your planning system is solving a different problem than the one you have.

Nearly every planning system on the market sizes safety stock with some variant of one expression:

SS = z · σD · √LT

It carries three assumptions. Long-lead semiconductor and optical components violate all three at once.

  1. Lead time is deterministic. It isn't. Memory, optical components, and advanced-node ASICs quote lead times that move with allocation cycles, not with your purchase order history.
  2. Demand variability is stationary. It isn't, when a single hyperscaler award reshapes a quarter and the next one contracts.
  3. Lead time and demand are independent. They are positively correlated. The conditions that spike demand are the conditions that stretch supply. When both move together, √LT understates the variance of demand-over-lead-time and the buffer comes out biased low.

The first of these isn't any vendor's claim. Researchers Eppen & Martin (1988) showed that the standard way of setting safety stock under moving lead times can land far from the service level you were aiming for. They gave a correct method instead.

The third one is ours. In allocation-driven component markets we consistently see demand and lead time move together. That compounds the error rather than cancelling it out.

Neither is fixable with a better spreadsheet. It takes modeling demand and lead time jointly, and an agent that can re-decide quickly as conditions change — so you hit the service level you committed to that quarter.

The agent

Built for planners to argue with, not to replace them.

The agent produces a recommended buffer per part number on a recurring cycle, with a written rationale attached to every recommendation.

It learns the decision, not the forecast

Trained to make buffer decisions under simulated demand and lead time regimes — normal conditions, demand surges, slowdowns, contractions — rather than fitting a point forecast and bolting a static rule on top.

Supplier responsiveness is an input

Parts from genuinely flexible, short-lead suppliers don't need the protection allocated components do. The agent segments accordingly instead of applying one service level across the whole part universe.

Every recommendation is bounded

Recommendations pass a guardrail layer with a deterministic fallback. The agent cannot emit a buffer outside policy limits, and a failed check falls back to a defensible classical calculation. There is no unbounded mode.

Planner disagreement is training signal

Accept and override decisions are captured per recommendation and fed back into the RLHF Agent. When a planner overrules the agent, that's data — not a support ticket.

Scope

What DeepVu doesn't do.

  • We don't replace your ERP or planning system of record. We produce decision recommendations (driven by AI Agent, Digital Twin and Knowledge Graph) that land in it.
  • We don't require a data lake project. We work from transactional history, open POs, on-hand inventory, and supplier lead time records.
  • We aren't another forecast accuracy vendor. Better accuracy on a part with a volatile 50-week lead time doesn't prevent the stockout. The SS buffer decision does.

Send us the purchase order history.
We'll show you where the policy is wrong.

A lead time volatility assessment returns a part-level view of where your current safety stock policy is structurally under- or over-protecting. No deployment required.


Request a lead time assessment