Project overview

Orderline AI

Sales-Controlled B2B Ordering Workflow Assistant

A chat-based ordering assistant that turns messy customer messages into structured, validated order drafts. It reduces manual admin work, product interpretation mistakes, delivery errors, and repeat corrections while keeping sales in control. It supports sales; it does not replace sales.

Structured order workflow Manual review for unclear cases Customer confirmation before final order

Executive summary

What this project does

Orderline AI converts customer chat messages into validated order drafts, reduces manual retyping, catches unclear requests, learns client-specific habits, and keeps sales in control before orders are confirmed.

Reduce manual work

Turn repeated sales/admin interpretation into a structured draft workflow.

Prevent costly mistakes

Catch unclear products, quantities, units, delivery notes, and confirmation state before handoff.

Prepare controlled AI automation

Build the validated business workflow before adding approval-gated agent assistance.

The daily challenge

Chat ordering creates avoidable operational risk

Customer messages often include shorthand, mixed languages, incomplete product names, delivery notes, corrections, and follow-up instructions. Sales must interpret products, quantities, units, delivery timing, pricing rules, and client preferences before customer service and warehouse teams can act.

2 boxes Product A 1 case premium item same size as last order add 3 kg frozen product cancel previous item delivery Wednesday 19th change quantity to 3 kg

The solution

A controlled assistant for sales and operations

Orderline AI receives LINE messages, detects order intent, matches products and SKUs, applies client-specific preferences, prepares order drafts, asks customers to confirm, and sends confirmed order emails to internal teams. When a message is unclear, it goes to Manual Review instead of guessing or creating hidden operational risk.

Sales stays in control

Drafts and unclear cases are visible to sales/admin before they affect the customer workflow.

Client-specific memory

Preferences are learned per client, so repeat shorthand can become faster without becoming global.

Safe confirmation

Final orders are created only after customer confirmation, preserving a clear audit trail.

Validated workflow

More than a chatbot

The system is designed around business validation, not free-form automation. It helps structure the order while respecting catalogue, pricing, client, review, and confirmation controls.

Product/SKU Quantity and unit Price availability Unit conversion Client-specific preferences Manual review status Confirmation state Future delivery rules

Before and after

From manual interpretation to a controlled workflow

Operational pain today

  • Sales manually reads long chats
  • Product shorthand is interpreted from memory
  • Orders are retyped manually
  • Delivery notes can be missed
  • Invoice/order corrections may require re-editing or reprinting
  • Client preferences depend on individual memory

Controlled workflow with Orderline AI

  • Order intent is detected
  • Product/SKU is validated
  • Unclear cases go to Manual Review
  • Preferences are client-specific and approved
  • Confirmation is controlled
  • Internal email handoff is structured
  • Traceability improves

How it works

Two clear workflows

Clear order flow

  1. LINE Message
  2. Parser & Matching
  3. Price/Unit Validation
  4. Draft Order
  5. Customer Confirm
  6. Email to CS/Warehouse

Unclear order flow

  1. Unclear Message
  2. Manual Review
  3. Sales/Admin Selects SKU
  4. Rebuilt Draft
  5. Customer Confirm
  6. Email

Current milestone

Current staging version: v0.9.13-learning-suggestions-v1

LINE webhook intake LINE group discovery and client linking Admin dashboard Manual Review workflow Draft creation and confirmation Add/change/cancel draft lifecycle Client-specific preferences Learning Suggestions page Email after confirmation Staging VPS deployment

Manual Review and Learning Suggestions

Unclear product phrases become controlled learning opportunities

If the system cannot confidently identify a product, it sends the message to Manual Review. Sales/admin selects the correct SKU, and the system can create a Learning Suggestion. Admin can approve or reject it. Approved suggestions become client-specific preferences.

Client A "premium item" → Product A, standard case
Client B "premium item" → Product B, large case

Learning is client-specific. One client's preference does not affect another client.

Human-in-the-loop by design

Control stays with the business team

Unclear products go to Manual Review. Client learning requires approval. Customer confirmation is required before final order creation. Price, catalogue, and delivery rules are not bypassed. Sales remains responsible for the relationship, negotiation, and exceptions.

Manual Review for uncertainty

Ambiguous requests are routed to sales/admin instead of being silently guessed.

Approval-gated learning

Client habits become preferences only after review, keeping learning scoped and controlled.

Confirmation before order

Drafts do not become final orders until the customer confirms through the controlled flow.

Sales-control and safety

The assistant supports sales. It does not replace sales.

The assistant does NOT

  • replace sales
  • negotiate prices
  • promise stock
  • approve discounts
  • confirm orders without customer confirmation
  • change catalogue/prices automatically
  • bypass delivery/logistics rules

The assistant DOES

  • prepare drafts
  • catch unclear orders
  • suggest client-specific learning
  • reduce repetitive work
  • support sales/admin
  • keep auditability

Roadmap

From staging milestone to production pilot

The business workflow foundation comes first: catalogue, delivery, rules, constraints, prices, and operational visibility. The agentic intelligence layer is added after those controls are in place.

Completed

v0.9.13 — Learning Suggestions v1

A. Business workflow foundation

Controls before automation

Next

v0.9.14 — Catalog Admin v1

  • product visibility
  • aliases
  • client preferences
  • unit conversions
  • constraints
  • price/missing data visibility

Then

v0.9.15 — Delivery Logic v1

  • delivery date extraction
  • same-day cutoff rule
  • sales/logistics override
  • delivery info in confirmation email
  • reduce wrong-delivery mistakes

Then

Rules, constraints, and operations visibility

  • rules/constraints/prices visibility
  • alerts/operations readiness
  • controlled pilot preparation

B. Agentic intelligence layer

Added after validated business controls

Then

Agent Gateway Phase 1 — Silent Advisory

  • OpenAI advisory layer
  • observes unresolved/contextual messages
  • suggests clarifications/preferences
  • no customer-facing autonomous action

Then

Agent Gateway Phase 2 — Assisted Replies

  • suggests better clarification messages
  • supports manual review
  • approval-gated

Then

Agent Gateway Phase 3 — Controlled Automation

  • low-risk contextual order assistance
  • always backend-validated
  • never bypasses pricing, product constraints, delivery rules, or customer confirmation

Later

Production expansion

  • ERP import/sync
  • WhatsApp integration
  • operations alerts
  • production pilot
  • multi-channel agentic operations layer

Business value

Practical gains for sales, customer service, and operations

Every wrong SKU, missed delivery note, or reprinted invoice creates hidden operational cost. Orderline AI reduces those repeat corrections by moving order handling into a controlled, traceable workflow.

Less manual retyping Fewer wrong-SKU mistakes Fewer delivery-date mistakes Fewer invoice/order re-edits Less invoice reprinting Faster sales-to-CS/warehouse handoff Better traceability Easier onboarding for new sales/admin staff Less dependency on one person remembering each client's habits Fewer repeated sales/admin corrections Lower hidden operational cost Time and cost savings through fewer corrections and faster processing

Pilot readiness

Designed for controlled staging and pilot preparation

Suitable for

  • internal staging tests
  • selected test LINE groups
  • sales/admin review
  • controlled pilot preparation

Not yet intended for

  • unsupervised production
  • automatic price negotiation
  • ERP overwrite
  • autonomous customer handling

Demo script

A simple checklist for stakeholder walkthroughs

  1. Demo 1: clear order → validated draft → customer confirm → email
  2. Demo 2: unclear phrase → manual review → learning suggestion → approved preference → future auto-match
  3. Demo 3: add/change/cancel item → updated draft lifecycle
  4. Demo 4: delivery request → delivery logic roadmap explanation
  5. Demo 5: duplicate confirm safely rejected