Describe a business problem. Get a structured AI implementation plan.
Live demo: ai-scoping-tool.vercel.app · Built with: Claude API · Vanilla JS · Vercel
Most AI implementation projects fail not because of bad technology, but because of poor scoping — teams jump to solutions before understanding the problem. This tool runs a structured discovery session: it asks three targeted questions about your process, then generates a complete scoping document covering the recommended AI approach, which steps to automate, effort vs. impact, a phased rollout plan, key risks, and success metrics.
The output is the kind of document an AI Implementation Lead would produce in the first client engagement. Built it because I do this work — and wanted a faster, more consistent way to do the first pass.
Phase 1 — Discovery Enter your Anthropic API key, describe a business problem in plain language, then answer 3 targeted questions (pre-generated by Claude) about your process, people, and pain points.
Phase 2 — Scoping The full Q&A context is passed to Claude, which generates a structured markdown scoping document covering approach, automation candidates, effort/impact, phased rollout, risks, and metrics.
No login required. The API key is held server-side on Vercel — just open the URL and start typing.
AI implementation scoping is the first real deliverable in any AI adoption engagement. The discovery → recommendation flow has a consistent structure — the same questions and output sections appear in every good scoping session. This tool codifies that structure and shows how an AI Implementation Lead would use LLMs to accelerate their own workflow.
It's also a deliberate demonstration of prompt architecture: two separate prompts with different system instructions, structured JSON output parsing in Phase 1, and long-context document generation with a strict template in Phase 2.
Easiest: Use the live demo — no setup, no API key needed.
Self-host on Vercel:
git clone https://github.com/sinkrest/ai-scoping-tool.git
cd ai-scoping-tool
vercel deploy
# Add ANTHROPIC_API_KEY to your Vercel project environment variablesRun locally (requires your own API key — edit app.js to call Anthropic directly):
git clone https://github.com/sinkrest/ai-scoping-tool.git
cd ai-scoping-tool
vercel dev # runs the serverless function locally with your .envAdd ANTHROPIC_API_KEY=sk-ant-... to a .env file for local dev.
Problem entered:
Our customer support team spends 3 hours per day manually reading incoming emails and routing them to the right department. We receive around 200 emails per day across 5 departments: billing, technical support, returns, sales enquiries, and general complaints.
Discovery Q&A:
- What does the current routing process look like step by step, and how is routing accuracy currently measured? → Agent opens each email, reads it, decides the department, moves it to the shared folder, logs it in a spreadsheet. No formal accuracy tracking — we know it's wrong when customers complain about late responses.
- What systems and data do emails currently flow through, and is there any existing email infrastructure? → Emails come into a shared Gmail inbox. We use Zendesk but the emails aren't automatically ingested — someone manually creates tickets. We have 6 months of past correctly-routed emails in a spreadsheet.
- What's the biggest operational pain point — is it routing speed, routing accuracy, the manual logging, or something else? → Speed and accuracy. Emails sit unread for up to 2 hours in peak periods. Wrong routing means a customer waits another day.
A customer support team manually reads and routes 200 daily inbound emails across 5 departments, consuming 3 hours/day and introducing 2-hour delays during peak periods. Routing errors cause secondary delays, with no current mechanism to measure or improve accuracy systematically.
Text classification using a fine-tuned or few-shot prompted LLM. The 6 months of correctly-routed historical emails provides an ideal training/evaluation dataset. Combined with Zendesk's API for automated ticket creation, this eliminates both the routing decision and the manual logging step.
Recommended tools/stack:
- Claude API (claude-haiku-4-5) — low-latency, low-cost classification at 200 emails/day; structured output via JSON mode
- Zendesk API — automated ticket creation and department assignment
- Gmail API or Google Workspace — email ingestion trigger
- n8n or Make.com — orchestration layer connecting Gmail → Claude → Zendesk
| Step | Currently Manual? | Automatable? | AI Approach | Confidence |
|---|---|---|---|---|
| Open and read email | Yes | Yes | LLM text classification | High |
| Decide department routing | Yes | Yes | 5-class classification with confidence score | High |
| Move to shared folder | Yes | Yes | Gmail API label/move | High |
| Create Zendesk ticket | Yes | Yes | Zendesk API | High |
| Log in spreadsheet | Yes | Yes | Sheets API or native Zendesk reporting | High |
| Handle ambiguous/complex emails | Yes | Partial | Flag for human review if confidence < 85% | Medium |
| Factor | Rating | Notes |
|---|---|---|
| Implementation Effort | Medium | Gmail + Zendesk APIs well-documented; main effort is prompt tuning and testing |
| Business Impact | High | Eliminates 3hrs/day of manual work + cuts response lag from 2hrs to <5min |
| Time to First Value | 2–3 weeks | Working prototype achievable in first week |
| Ongoing Maintenance | Low | Monitor confidence scores weekly; retrain prompt quarterly |
Phase 1 — Proof of Concept (Weeks 1–2) Build a standalone classifier: feed 50 historical emails through Claude, compare predicted vs. actual routing, target >90% accuracy. No Zendesk integration yet — just validate the classification prompt.
Phase 2 — Pilot (Week 3–4) Connect Gmail → Claude → Zendesk for a single department (e.g. billing — highest volume, clearest signal). Run in parallel with manual routing for one week to compare. Human override remains available.
Phase 3 — Full Rollout (Month 2) Extend to all 5 departments. Add confidence threshold: emails below 85% confidence are flagged for human review rather than auto-routed. Deprecate the manual spreadsheet log.
- Risk: Low classification accuracy on ambiguous emails → Mitigation: Implement confidence threshold; route uncertain emails to a "needs review" queue rather than auto-assigning
- Risk: Customers feel depersonalised by automated responses → Mitigation: Automation is routing-only, not responding — human agents still handle the actual reply
- Risk: Gmail/Zendesk API changes break the pipeline → Mitigation: Use n8n as the orchestration layer; swap connectors without rewriting business logic
- Routing accuracy ≥ 92% (vs. unknown baseline — establish in Phase 1 test)
- Average email-to-ticket time < 5 minutes (vs. current 2-hour peak lag)
- Manual routing time reduced from 3hrs/day to < 20 minutes (human review queue only)
- Zero increase in customer complaints about mis-routing after 30-day rollout
- Export 200 historical emails with correct routing labels from the spreadsheet — this becomes the evaluation dataset
- Build and test the classification prompt against this dataset; iterate until >90% accuracy
- Set up a free n8n instance and connect Gmail as trigger
- Claude API (
claude-sonnet-4-6) via fetch — no SDK - marked.js (CDN) — markdown rendering
- Vanilla JS — no framework, no build step
- sessionStorage — API key handling (cleared on tab close)
Built by Roman Martins · LinkedIn