A working framework for Odele Beauty

One AI operating layer. Not a dozen point solutions.

Demand planning, supplier POs, 3PL rate pulls, and the monthly supply chain dashboard each have their own AI vendor right now. Odele can buy four. Or we can build one operating layer that owns those workflows end-to-end, calls into specialized tools where the depth is genuinely needed, and grows from inside the team. Same skills callable from a custom UI and from Copilot 365 via MCP. Human approval gates every spend.

The point-solution paradox

Anaplan for demand. Kinaxis for supply. Adaptive for finance. Glimpse for chargebacks. Emmy for procurement. Each one is a real product solving a real problem, and each one is a separate login, contract, and integration. The bottleneck for a 31-person team is not the model in any one tool. It is owning the surface that connects them to your data, your approvals, and each other.

The thesis

Buy specialized depth where it earns its place. Build the operating layer that connects everything else.

Generic Copilot retrieves and summarizes. A point-solution AI runs one workflow well and stops at its own boundary. Neither owns the seam between MPS, NetSuite, the supplier mailbox, and the 3PL portal, which is where Odele's actual operating leverage hides.

A skill is not a prompt. It is the encoded answer to "how does Odele actually do this work?" MPS pulled from OneDrive. PO drafted to match your NetSuite template. Supplier email sent in your voice from a service mailbox. Reply tracked, escalated at 14 days. Every step audit logged. Nothing leaves the tenant before a human approves.

The platform compounds. Each new skill costs less to build than the last. Memory accumulates per supplier, per SKU, per retailer. The team's tacit logic, today scattered across spreadsheets and Slack, becomes executable.

The Architecture

Four layers. One coherent system.

The shape is opinionated and stack-agnostic. Connectors fit Odele's existing tools (NetSuite, OneDrive, Teams, True Commerce, supplier email). The runtime sits inside your M365 tenant. The approval queue is the permanent boundary between the model and any spend, send, or commitment.

L0 Your Data Perimeter SSO / Identity Network Boundary Compliance Posture
L1 Connectors
NetSuite OneDrive Teams + Outlook True Commerce EDI 3PL Portals
L2 Skill Runtime
Agent Orchestration Skill Registry MCP Server (Copilot 365) Eval Harness Versioning
L3 Memory Layer
Per-Supplier Per-SKU Per-Retailer Brand View
↺ feedback to L2 + L3
L4 Audit + Reviewer
Approval Gate (per spend) Source Trail Structured Audit Log Adoption Telemetry
Why nothing leaves before approval POs, supplier emails, and EDI sends land as drafts in the approval queue with full provenance. A reviewer approves, edits, or rejects line by line. The model never commits a spend on its own.
How the platform learns Every approve, edit, and reject feeds back into the eval set and memory layer. Patterns the team correct twice become the platform's defaults. Improvement comes from inside, not from vendor releases.
★ Skills are the unit of value. Everything else exists to make skills safe, accurate, and contributable.
The Build Sequence

Four stages. One compounding system.

Stage one is the hard part. Most platforms skip it and ship something that demos well and breaks the first time it touches your real PO format.

stage_01 / capture

Extract the playbook.

Sit with the operator who runs the workflow today. Trace MPS-to-PO end-to-end. Capture the unwritten rules: which suppliers get which lead time, which SKUs are pre-approved for auto-draft, what Andrea would never send to a supplier without rephrasing.

That transcript becomes the spec for the first skill.

Where the depth lives
Output
  • Workflow trace for MPS-to-PO
  • Approval rubric (what "good" looks like per line item)
  • Source inventory: NetSuite scope, MPS format, mailbox access
  • Golden example set: 10 historical POs, approved as-was
  • Voice guide for supplier emails
stage_02 / build

Encode it as a skill.

Wrap the spec in a callable skill. Connect it to NetSuite, OneDrive, and the supplier mailbox. Implement the eval logic. Define the PO and email templates. Wire the audit hooks. Register the same skill as an MCP tool so Copilot 365 can call it.

The skill runs inside Odele's M365 tenant, uses Odele identity, writes to Odele storage. Security posture is preserved.

Output
  • MPS-to-PO skill, callable end-to-end
  • NetSuite, OneDrive, Outlook connectors live
  • Approval queue UI + Copilot 365 entrypoint
  • Audit log writing on every run
  • Adoption telemetry instrumented
stage_03 / promote

Eval-gated quality.

Curated golden examples gate promotion. The MPS-to-PO skill has to match past POs Odele actually sent before it goes to the full team. Promotion is a measurable threshold, not a meeting.

Reviewers see the score on every draft. Quality becomes legible to Shanti, not folklore.

Output
  • Eval score reports per skill version
  • Draft → Pilot → Full team gating
  • Reviewer feedback loop into evals
  • Audit-ready records of every run
stage_04 / compound

Team-owned, growing.

The second skill (3PL rate quotes, monthly dashboard, supplier email triage) is cheaper than the first. Shared connectors, shared eval patterns, shared memory layer. The Odele ops team becomes skill authors. The platform grows from inside, not from a vendor roadmap.

Memory accumulates per supplier and per SKU. Adoption telemetry surfaces what the team actually uses, and what to retire.

Output
  • Skill catalog with named owners
  • Persistent memory live across calls
  • Adoption dashboard for Shanti and leadership
  • Author onboarding for the ops team
What the ops team sees

The ops team's daily operating surface.

Every skill output lands as a draft with sources attached and an eval score. Approve, edit, or reject line by line. Same skills callable from Copilot 365 via MCP. Click a draft in the sidebar to inspect it, or load a skill from the input below.

platform.odele.internal / queue

Odele AI Platform

Pick a skill from the sidebar, or click an existing draft to inspect.

Three workflows that earn the wedge

Where the value shows up first.

MPS-to-PO is the anchor: clearest pain, biggest blast radius, natural human-in-loop checkpoint. The other two reuse the same connectors and approval pattern. Sequencing is Odele's call after May 7.

anchor · wedge

MPS-to-PO end-to-end

MPS lands in OneDrive. The skill aggregates new orders, drafts POs in the NetSuite sandbox using your existing form, surfaces them in the approval queue, sends to suppliers from a service mailbox on approval, tracks replies, and escalates anything quiet for 14 days. Nothing leaves Odele before a human approves.

workflow

3PL Rate Quote Pull

Lane and volume in. Quotes out. Portal automation as the interim path for carriers without APIs, real APIs as roadmap. Quotes land as a side-by-side comparison with prior-cycle benchmarks, ready for ops to accept or counter.

workflow

Monthly Supply Chain Dashboard

The Excel dashboard Shanti rebuilds every book close, populated automatically from NetSuite, EDI, and 3PL data. Same layout the team already trusts. Rendered in the platform with drill-down to source rows and a flagged delta on anything that moved more than threshold.

The principles

Six principles that separate a platform from a chatbot.

01

One surface, two access patterns

Same skills callable from a custom web UI and from Copilot 365 via MCP. Not two products to maintain. The team picks the surface that fits the moment, the platform stays single-source.

02

Sourced, not invented

Skills act on Odele's actual data: NetSuite POs, OneDrive MPS, the EDI feed, the supplier mailbox. They do not free-generate values that should come from a system of record. Every output traces to its sources.

03

Approval gate before any spend

POs, supplier emails, and EDI sends land as drafts. Reviewer can approve, edit, or reject line by line. Nothing commits to a vendor without a human in the loop. Reviewer decisions train the next eval cycle.

04

Persistent memory per entity

Per-supplier, per-SKU, per-retailer context accumulates across runs. Lead times, packaging quirks, who-replies-fastest, which Target buyer prefers what cadence. The same skill called twice on the same supplier uses what it learned the first time.

05

Skills compose

Complex work calls multiple skills. The MPS-to-PO skill calls supplier-memory for lead times and supplier-mail for the send. The dashboard calls the same NetSuite reader the PO skill uses. Pipelines emerge from primitives.

06

Vendor-continuity by repo

The platform lives in a GitHub repo Odele can read, fork, or hand to another partner. No black box. No proprietary runtime. Continuity is structural, not a contract clause.

The engagement

One workflow live in eight weeks. Then a managed cadence.

A wedge, not a cathedral. MPS-to-PO live and adopted in eight weeks teaches Odele more than six months of platform planning. The May 7 session decides whether MPS-to-PO is the right anchor or the team picks something tighter.

The first sprint.

Eight weeks from kickoff to MPS-to-PO live with one real production cycle, the platform shell installed in Odele's M365 tenant, audit live, and weeks 5–8 already shipping the second skill.

week 1

Foundations

Azure tenant, model access via MS infra, GitHub repo + CI, NetSuite service account scoped to the PO module, MCP server registered in Odele's Copilot tenant.

weeks 2 – 4

MPS-to-PO end-to-end

MPS ingest in Odele's format. PO drafting against the real NetSuite PO template. Approval queue UI. Supplier email send via the Odele service mailbox. One full real production cycle with full Odele oversight in week 4.

weeks 5 – 7

Second skill + adoption

3PL rate quote pull or monthly dashboard, picked from the May 7 wishlist. Reuses the connectors and approval pattern. Pair sessions with SMEs to capture tacit logic. Role-based permissions hardened.

week 8

Decision point

Review adoption and friction signals. Pick the next 2–3 workflows from the refined wishlist. Decide: continue the managed-service cadence, or pause to evaluate.

Engagement model
Managed service
$5K / month, month-to-month after week 8. Same model as our Columbia Ridge engagement, not a one-time build.
What that buys
Encoded by EG, validated by Odele
EG owns implementation, evals, infra, and the next-skill backlog. Odele owns approvals, domain calls, and what's worth automating next. Dashboard and Teams bot are equal interfaces, not "the real product and the bot."
Compares to
~$150K+ / yr in alternatives
Anaplan ~$50K, Kinaxis ~$50K, Adaptive ~$50K. Each one a separate contract, login, and integration owned outside the team.

Make Odele's playbook executable.

Forty-five minutes to walk the working slice and pick the first skill together. The eight-week sprint follows from the conversation.

Continue the Conversation