Use case · FMCG / CPG

Demand Sensing & Replenishment OS

An agentic operating layer that senses true demand across channels, reconciles signals, and proposes replenishment your planners approve.

Signal ReconcilerAnomaly SentinelForecast DrafterPromo DecouplerNew-SKU Proxy Builder
The friction

Where value leaks today.

Demand planning in CPG still runs on a monthly cadence stitched together from POS extracts, distributor sell-through files, and a planner's gut. By the time a forecast is reconciled in the planning tool, the signal that mattered — a weather swing, a competitor stockout, a viral SKU — is already two weeks stale. Statistical baselines smooth over exactly the volatility that decides whether you ship or sit on inventory.

Buying a forecasting tool doesn't fix this because the tool produces a number, not a decision. Someone still has to interrogate why the forecast moved, override it where category knowledge contradicts the math, and translate the accepted forecast into replenishment orders the supply team will act on. That connective tissue — sensing, explaining, reconciling, acting — lives in spreadsheets and Monday calls, and it breaks every time a planner leaves or a channel changes its data format.

The leak is structural: every handoff between sensing demand and committing supply loses fidelity and adds latency. A dashboard makes the gap visible without closing it. What's missing is an owned system that runs the loop continuously and routes only the genuine judgment calls to a human.

How it runs

One governed flow — agents act, you approve what matters.

Planners spend their time judging the forecasts that matter instead of rebuilding them, and replenishment moves on fresh demand instead of last month's average.

kitsune os · demand-sensing-replenishment-osrunning
01
Ingest signals
POS, weather, social, sell-through
Agent
02
Reconcile sources
resolve conflicting demand reads
Agent
03
Detect shifts
flag breaks from baseline
Agent
04
Draft forecast
with explained adjustments
Agent
05 · gate
Approve forecast
planner accepts or overrides
Human
06
Push replenishment
commit orders to ERP
Agent
Agent — autonomousHuman gate — your approval
What the OS runs

One operating layer — eight governed jobs.

Each is a governed agent inside the same system, sharing context — not eight tools you stitch together.

01

Signal Reconciler

Merges POS, distributor sell-through, syndicated panel, and external signals into one demand read, flagging where sources disagree rather than silently averaging them.

02

Anomaly Sentinel

Watches for demand breaks — stockout-driven spikes, cannibalization, seasonal early-pull — and separates real shifts from noise before they corrupt the baseline.

03

Forecast Drafter

Produces SKU-by-location forecasts with a plain-language rationale for every material adjustment, so planners review reasoning, not just numbers.

04

Promo Decoupler

Strips promotional lift out of baseline demand so post-promo forecasts don't inherit phantom volume.

05

New-SKU Proxy Builder

Constructs demand curves for items with no history by matching attributes to comparable launches and decaying the analog over time.

06

Replenishment Composer

Translates the approved forecast into time-phased orders against lead times, MOQs, and safety stock, ready to commit to the ERP.

07

Bias Auditor

Tracks forecast error and over/under bias by planner, category, and channel, surfacing where the system or a human is consistently off.

08

Allocation Arbiter

When supply is short, proposes how to split constrained stock across customers and channels using service-level and margin priorities.

Governed by design

Autonomy you can trust — because the control is built in.

The system acts on its own and every action stays legible, bounded, and reversible. You don't choose between speed and control; the control is what makes the speed safe.

Legible

See what was done, what was declined, and exactly what's waiting on you — nothing happens in a black box.

Bounded

Agents act only within the rules you set. Anything material or irreversible stops at a human gate.

Reversible

Every action is logged and undoable. A wrong turn is caught and rolled back, not discovered weeks later.

Owned

One operating system you own — not a swarm of rented agents you have to police. Built, run, accountable.

The outcome

Planners spend their time judging the forecasts that matter instead of rebuilding them, and replenishment moves on fresh demand instead of last month's average.

1 OSSense-to-replenish loop
ContinuousForecast cadence
ExplainedEvery adjustment
OwnedNot rented per seat
Questions, answered

What you're actually getting.

Is this a product or a build?

It's a build. Kitsune forges a demand operating layer around your data sources, planning rhythm, and ERP — you own it. It is not a seat license on a generic forecasting app.

What stays in my control?

Every forecast the system commits passes through a planner gate. Agents sense, reconcile, and draft; your team accepts, overrides, or sends back. Replenishment only fires on approved numbers.

How is this different from a forecasting platform?

A platform hands you a number on a screen. This runs the whole loop — sensing demand, explaining shifts, drafting forecasts, and committing replenishment — and routes only the real judgment calls to humans.

Can it handle SKUs with no sales history?

Yes. The New-SKU Proxy Builder constructs demand curves from attribute-matched analogs and decays them as real data arrives, so launches aren't forecast by guesswork.

How does it deal with conflicting data sources?

The Signal Reconciler surfaces disagreement between POS, sell-through, and panel data explicitly instead of averaging it away, so planners see where the uncertainty actually lives.

Bring us the bottleneck.

We'll forge the operating layer around your friction — built, owned, and running.