AI Agents · CORTEX-PLANNING · Building

Cortex Planning Agent.

Store-realistic Plan of Record + continuous, exception-led replanning.

Building PLAN L4 Powered by → Cortex

Workflow level

L4 Agentic

Space Planning workflow

Replanning cadence

Continuous

vs monthly today

Output granularity

Fixture · slot

store-specific VM layouts

Planning surface

Every store

store-first Plan of Record

01 · The job to be done

The pain

Today, planning is monthly, spreadsheet-driven, store-agnostic. Plans land in stores with assumptions baked in that no longer match real space, real inventory, or real demand.

What the agent does

Builds a store-realistic Plan of Record. Converts top-down financial targets into store-level plans that reflect real space, inventory and demand constraints. Translates assortment + inventory into fixture- level and slot-level VM layouts. When conditions change — a delayed shipment, an underperforming category, a competitive signal — the agent generates an exception-led replan.

What humans still do

Planner sets targets and reviews replans. Agent does not replace planning. It makes planning continuous.

02 · How it works · the closed loop

Inputs (blue) → Workflow (grey, steps 01-06) → Outputs (green). Step 06 loops exceptions back to Build.

Ingest unified inputs

Sales · returns · inventory · margin · targets (Sales · GM% · Markdown) · store space and fixture capacities · real-time inventory · VM guidelines and grouping rules.

Build store-first Plan of Record

Converts targets + demand signals + store constraints into store-level plans (WSSI · OTB · assortment · financial planning). Runs scenario simulations across assortment × intake × price. Selects optimal plan by margin, risk, feasibility.

Control OTB actions

Orchestrates commit / chase / hold / cancel decisions under defined risk guardrails.

Autonomous VM planning

Translates assortment + inventory into visually optimised layouts at fixture and slot level. Balances aesthetics, inventory availability and feasibility against brand VM guidelines (palette, arrangement styles).

Deliver execution outputs

Store Plan of Record + rolling WSSI · OTB recommendations · assortment depth/option plans · store-specific planograms/elevations · fill/degradation indicators + exception alerts with corrective actions.

Monitor and replan

Continuous performance monitoring; exception-detection triggers replans rather than waiting for the next planning cycle.

03 · Underlying data

Data source	Classification	Key entities
Impetus	Mixed	OTB · Plan of Record · WSSI · assortment plan
POS	Machine-Created	Sell-through % · ROS · revenue by store
SAP F&L	Mixed	Stock Positions · Margin · Cost Price · Landed Cost
Catalog Engine	Machine-Readable	SKUs · Style Codes · Size-Colour Matrix · Tech-Pack

Refresh cadence: Daily for plan; real-time for exception triggers.

KPIs moved:

Forecast accuracyGMROFAged inventory %SPSF (Sales per Square Foot)

04 · Design patterns used

Tool Calling

Pulls live OTB · WSSI · stock from Impetus + SAP

Context Engineering

Compresses store traits + active VM rules into per-replan context

Agent Memory

Episodic memory of prior replans avoids re-asking solved questions

Wide Research

Generates N candidate plans, scores on margin × risk × feasibility, picks one

Deep / Hierarchical

Planner delegates to specialist sub-agents per planning dimension

Observability

Every replan trace records inputs, scenarios scored, decision rationale

05 · Evals · how we know it works

Replay historical seasons; score plan-vs-actual sell-through and margin against the best-case manual plan as the holdout benchmark.

Gates and thresholds.

Margin within 2% of best-case manual plan on holdout season
Replan latency < 4 hours from exception trigger
Constraint-respect rate > 99% (no plans violate VM rules or OTB ceiling)