AI-Assisted ERP Change Request Pipeline

Executive Summary

A specialist ERP consultancy manages a high volume of documents required for initial client contact through to signed-off transport releases. Each change follows a strict, multi-document lifecycle — quote, functional specification, unit test document, and transport request form — all authored against branded Word templates and stored in a structured client folder hierarchy.

Before this engagement, consultants spent a disproportionate share of billable time on document assembly: reading the ticket system, cross-referencing prior documents, answering the same scoping questions repeatedly, and manually filling template fields. There was no persistent memory between sessions, so context was reconstructed from scratch on every interaction.

We built an AI-powered consulting operations plugin that automates the entire change request lifecycle from first discussion to transport. Each stage is driven by a slash command that fetches live data, reads prior session context, guides the consultant through only the genuinely missing gaps, and generates the finished Word document in one step.

The Challenge

The consultancy’s delivery workflow had three compounding inefficiencies:

• Repetitive document assembly: Every change request required the same multi-stage document set — Quote, Functional and Technical Specification (FTS), Unit Test Document (UTD), and Transport Request Form. Each document took 1–2 hours to prepare manually, largely because the consultant had to re-read the ticket, locate prior documents, and re-enter data that already existed elsewhere.
• No session memory: When a consultant resumed work on a ticket — often days later — they had to reconstruct context from the ticketing system, emails, and prior Word files. There was no structured handover record.
• Fragmented operational tasks: Daily activities such as timesheet completion and standup preparation required consultants to manually cross-reference three separate systems (worklist, time entries, planned capacity), typically costing 30–45 minutes per day.

The Solution

We designed a Claude Code plugin — installed into the company’s AI assistant environment — that exposes the entire delivery lifecycle as a set of slash commands. The plugin connects to two backend MCP servers: one for the ticketing system (live ticket data via UI automation) and one for Word document generation (template token replacement and tracked changes).

Key Architectural Decisions

1. Slash-command skill pipeline Each stage of the change request lifecycle maps to a single command: /discuss, /quote, /fts, /unittest, /transport. Commands chain naturally — each one checks the ticket folder for prior context files and existing documents, so the consultant never re-answers a question already resolved in an earlier session.

2. Persistent session context After every interaction, a structured handover file is written silently to the ticket folder. The next session reads it before asking a single question, giving the AI full continuity across days or weeks without any manual effort from the consultant.

3. Infer before asking Every skill is written to extract the maximum possible information from the ticket, prior documents, and saved context before surfacing a question. A transport form, for example, infers cross-client flags and back-out steps from the FTS rather than asking the consultant to re-state them. A client landscape file caches SAP system and client number mappings so they pre-fill on the next transport for the same client.

4. Phase-based context compression Long multi-stage workflows (discussion → draft review → generation) compress completed phases to a structured summary, keeping the active context window focused on the current step while preserving a permanent audit record.

5. Operational automation Beyond documents, two additional skills address daily overhead:

• /fill-timesheet — extracts five weeks of timesheet history, identifies gaps against the weekly target, suggests evidence-backed entries using historical patterns and optionally an M365 Copilot activity pack, and books confirmed entries directly.
• /standup — cross-references the worklist, planned MRP capacity, and last fortnight of timesheets in a single pass, produces a two-zone prioritised briefing (committed vs available pool), detects stale and waiting tickets, and generates a standup script the consultant can read verbatim.

The Outcomes

• ~65% faster document preparation: A complete Quote document that previously took 90–120 minutes now takes 25–35 minutes. The time saving compounds across the full pipeline — a ticket that previously required four separate manual documents now runs end-to-end with four commands and minimal re-entry.
• Zero context loss between sessions: Consultants resume any ticket in seconds regardless of how long ago the last session occurred. Prior decisions, TBC items, and generated file paths are always immediately available.
• ~25 minutes saved per day on operational tasks: Timesheet top-up and standup preparation together consumed up to 45 minutes daily; both now complete in under 10 minutes with richer output than the manual process produced.
• Consistent document quality: Structured review templates and a strict no-invention policy (the AI marks uncertain items as [TBC — reason] rather than guessing) have driven a measurable reduction in review-round-trips before client sign-off.
• Extensible by design: The skill architecture separates prompt logic from content templates and document generation steps. Adding a new document type (e.g. a Root Cause Analysis or Handover Note) requires only a new skill folder with a SKILL.md, content template, and Word MCP step file — no infrastructure changes.

Technology Stack

• Platform: Claude Code / Claude Cowork plugin system
• AI Engine: Claude (claude-sonnet-4-6) with structured skill prompts and phase-based context compression
• Integrations: Custom MCP server for ticketing system (UI automation via Node/npx); word-mcp-live MCP server for Microsoft Word document generation
• Document layer: Tokenised Microsoft Word templates with automated search-and-replace and tracked-change amendment workflows
• Persistence: Structured Markdown context files written to the client folder hierarchy after every session
• Skills delivered: 9 slash commands covering the full change request lifecycle, timesheet management, and daily standup preparation