01 // FORECAST
Demand planner
Frames the demand question, chooses a suitable method, diagnoses error and bias, and makes uncertainty visible before supply is committed.
01.4 // SUPPLY CHAIN · GOVERNED AGENT PROTOTYPE
A DEPARTMENT,
NOT A CHATBOT.
Eight named roles cover planning and governance. Seven workflows make the work callable. A north star, independent verifier and task-by-task trust ledger decide how much autonomy each type of work has actually earned.
08
SPECIALIST + GOVERNANCE AGENTS
07
CALLABLE PLANNING WORKFLOWS
08
SKILL MODULES
L0→L3
TASK-BY-TASK TRUST LADDER
// why I built it
Planning is a team sport with conflicting objectives.
Demand, inventory, production, logistics and finance do not optimise the same thing. A general assistant can answer each question separately and still produce a plan that fails as a system.
I built the Planning Department as an operating model: specialist roles with explicit hand-offs, a supervisor that holds the end-to-end objective, and a verifier whose job is to try to refute important outputs before a human relies on them.
THE OPERATING RULE
Trust belongs to a task, not to an agent.
An agent can be highly trusted for a read-only inventory check and still remain draft-only for a purchase recommendation. Autonomy rises only where verified evidence supports it.
// the roster
Six planners. Two control seats.
The roster mirrors the work a planning department actually performs, then adds the two roles most AI demos omit: someone to coordinate the system and someone independent to challenge it.
02 // ALIGN
S&OP planner
Turns demand, supply and financial tension into explicit scenarios and decisions instead of letting one function quietly dominate the plan.
03 // MRP
Supply planner
Connects demand to master planning, material requirements and capacity checks while surfacing the assumptions behind dates and quantities.
04 // POLICY
Inventory planner
Works across service, cash and risk: segmentation, policy, safety stock, reorder logic, record accuracy and ageing.
05 // FLOW
Production planner
Reasons about schedules, bottlenecks, dispatch priorities, queues and work-in-process without confusing local utilisation with system throughput.
06 // MOVE
Logistics planner
Covers distribution requirements, transport choices, network constraints and the hand-offs that make a plan executable.
07 // ORCHESTRATE
Supervisor
Owns the end-to-end view, reads the north star, coordinates specialists and decides which trade-offs or exceptions deserve human attention.
08 // CHALLENGE
Plan verifier
A separate adversarial role recomputes claims, searches for missing evidence and returns PASS, FAIL or UNVERIFIABLE. It never fixes the work it grades.
// how I made the team operable
Roles are the start. Governance makes them useful.
- 01
Map the department, not the chatbot
I began with the recurring jobs inside planning — demand, S&OP, master planning, inventory, scheduling and logistics — then made ownership explicit.
- 02
Give each role a narrow contract
Every specialist receives a purpose, inputs, outputs, escalation rules and a definition of done. The supervisor coordinates; it does not silently absorb every job.
- 03
Separate instructions, knowledge and tools
Role behaviour, reusable planning methods and deterministic calculations live in distinct layers so each can be reviewed and replaced independently.
- 04
Make verification a different seat
The doer and verifier are separated by design. Critical numbers are recomputed, evidence gaps stay visible, and a clean result must show what was actually checked.
- 05
Write a north star before granting initiative
The human defines paired tensions such as service and cash. Agents may recommend work only inside an explicit direction set by the owner.
- 06
Track trust by task type
Forecast review, inventory health, PO drafting and invoice matching do not share one autonomy level. Reliability is earned separately for each class of work.
- 07
Keep a permanent human floor
Payments, commercial commitments, consequential external communication and exceptions outside the evidence boundary remain human decisions.
THE OPERATING MODEL
Autonomy is earned, not announced.
- L0 Shadow — the agent runs, but its output is not used operationally.
- L1 Draft — every deliverable is reviewed before a person acts.
- L2 Verified autonomy — defined checks run automatically; people work the exceptions.
- L3 Whitelist — only proven, narrow task types move to periodic review.
THE PUBLIC BOUNDARY
Architecture shared. Private context withheld.
- The public case explains original role design, workflow structure and governance patterns.
- It contains no employer, customer, supplier, colleague or real transaction details.
- Licensed learning materials are not republished or presented as an open-source knowledge base.
- The project makes no certification, endorsement or production-autonomy claim.
THE SIGNAL // DOMAIN × ORCHESTRATION × GOVERNANCE
The difficult design question is not what agents can do. It is what they are allowed to own.
This project turns supply-chain expertise into a roster, a set of callable workflows and an evidence-based permission system. It is a prototype for human–agent planning teams, not a claim that an entire department should run unattended.