Use Case Canvas -- AI Support Copilot
Engagement: AI Support Copilot Pilot
Client: Prasanna
Version: 1.0
Date: 2026-04-30
Authors: Amit (POD Lead), Shivani (PM)
Framework ref: Doc 03, Section 3.2
Single-page canvas. If this cannot be completed, the engagement is not ready to estimate.
Mission Statement
Enable support agents to resolve tickets faster and more consistently by providing an AI copilot that classifies incoming tickets, retrieves relevant KB articles, recommends the next-best-action, and drafts grounded responses -- reducing resolution time from hours/days to minutes while maintaining human approval on every response.
Users & Volumes
| Dimension | Detail |
|---|
| Primary users | Support agents (Freshdesk users) |
| Secondary users | Support team lead (reviews outputs, evaluates quality) |
| Ticket volume | ~1000 tickets/month |
| Auto-answer target | 30% of recurring tickets (~300/month) answered autonomously with human approval |
| Peak load | Not specified; estimate ~50 tickets/day on average, peak likely 2-3x |
| Usage pattern | Agent processes ticket → copilot provides recommendation → agent reviews, edits, sends |
| Input | Source | Example |
|---|
| Support ticket | Freshdesk (Excel for pilot) | Subject: "SSO login fails after password reset" Description: "User changed password yesterday, now SSO throws 403 error. Tried clearing cache, still failing. Need urgent fix." Channel: Email Customer: Acme Corp |
| KB articles | Freshdesk KB (Excel for pilot) | 12 articles covering Authentication, Billing, Data Import, Integrations, Access Control, Compliance, Known Issues |
| Escalation rules | Configured rules (Excel for pilot) | 5 rules mapping conditions → teams (Engineering, Integrations, Finance, Compliance, Platform Ops) |
Outputs
| Output | Description | Example |
|---|
| Classification | Category + Priority + Sentiment | Category: Authentication Priority: High Sentiment: Frustrated |
| Retrieved KB articles | Top-k relevant articles with relevance scores | 1. KB-AUTH-001: SSO Configuration Guide (0.92) 2. KB-AUTH-003: Password Reset Procedures (0.87) |
| Action recommendation | Reply / Ask for more info / Escalate | Recommended action: Reply Confidence: 0.91 |
| Draft response | Grounded response with KB citations | "Thank you for reporting this. Based on our SSO Configuration Guide [KB-AUTH-001], this typically occurs when... Steps to resolve: 1) ... 2) ... 3) ..." |
| Confidence score | Per-output confidence indicator | Overall confidence: High (0.91) |
| Reasoning trace | Why the copilot made this recommendation | "Matched to Authentication category based on 'SSO' and '403 error'. KB-AUTH-001 covers SSO configuration issues. No escalation rule triggered -- standard resolution path available." |
Success Metrics
| # | Metric | Type | Target | Measurement |
|---|
| 1 | Classification accuracy | AI quality | >= 85% | Exact match on category and priority against golden set |
| 2 | Retrieval accuracy | AI quality | >= 85% | Expected KB article appears in top-k retrieved set |
| 3 | Action accuracy | AI quality | >= 85% | Exact match on next-best-action (Reply / Ask / Escalate) |
| 4 | Response acceptance rate | Business | >= 70% | % of drafts agents accept without major rewriting |
| 5 | Auto-answer coverage | Business | 30% of recurring tickets | % of well-defined tickets the copilot can handle |
| 6 | Autonomous operation | Business | No handholding | System answers queries without manual intervention |
Go-live validation: Run against 1000 synthetic questions, reviewed by support team lead.
Hard Limits
| Constraint | Limit | Rationale |
|---|
| Human-in-the-loop | Always -- no auto-send | Agent must review and approve every response |
| Profanity / misuse | Zero tolerance | All outputs must be checked; adversarial inputs handled gracefully |
| LLM vendor lock-in | Must be LLM-agnostic | Architecture must support swapping LLM providers |
| Portability | Full on-premises handover post-pilot | Code, models, data -- everything transferable to client premises |
| Timeline | Demo by May 10, final delivery May 16 | Hard deadlines, not flexible |
| Hallucination | Responses must be grounded in KB | No fabricated information; cite sources or flag "no match" |
Data Sources
| Source | Records | Access (Pilot) | Access (Production) | Sensitivity |
|---|
| Historical tickets | 36 | Excel dataset | Freshworks API | Low (simulated data for pilot) |
| KB articles | 12 | Excel dataset | Freshworks KB API | Low |
| Escalation rules | 5 | Excel dataset | Freshworks / config | Low |
| Evaluation set | 12 (held out) | Excel dataset | Expanded to 1000 synthetic | Low |
KB refresh: Not needed for pilot. Production will require automated refresh via Freshworks APIs.
PII: No PII in pilot data. Production will require PII handling policy (Shubham to document).
Out of Scope (Phase 1)
| Item | Reason |
|---|
| Freshdesk API integration | Pilot uses Excel; architect for it but don't build yet |
| Multi-language support | English only for pilot |
| Customer-facing AI | Agent-facing only; customers never interact with the copilot |
| Auto-send / auto-reply | Human always in the loop |
| Live KB refresh | Not needed with static Excel data |
| Production deployment | Pilot is standalone; productionization note covers the path forward |
| Phase 2 scope | Not defined; to be discussed after pilot results |
| Load testing / scaling | Not applicable for pilot with demo + eval run |
Open Questions
| # | Question | Owner | Status |
|---|
| 1 | Target cost-per-ticket for production? | Shivani → Prasanna | Pending -- ask in next communication |
| 2 | Confidence indicator design -- binary or three-tier? | Amit → propose to Prasanna | Pending |
| 3 | Which LLM to recommend for this use case? | Amit | To be decided in Architecture Sketch |
| 4 | Vector store selection (pgvector / Qdrant / ChromaDB / Elasticsearch)? | Amit | To be decided in Architecture Sketch |
| 5 | How many synthetic questions to generate per category? | Nishka + Amit | To be decided in Evaluation Plan |
| 6 | Feedback loop storage -- where do agent edits persist? | Amit + Nancy | To be decided in Architecture Sketch |