Skip to main content

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

DimensionDetail
Primary usersSupport agents (Freshdesk users)
Secondary usersSupport team lead (reviews outputs, evaluates quality)
Ticket volume~1000 tickets/month
Auto-answer target30% of recurring tickets (~300/month) answered autonomously with human approval
Peak loadNot specified; estimate ~50 tickets/day on average, peak likely 2-3x
Usage patternAgent processes ticket → copilot provides recommendation → agent reviews, edits, sends

Inputs & Outputs

Inputs

InputSourceExample
Support ticketFreshdesk (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 articlesFreshdesk KB (Excel for pilot)12 articles covering Authentication, Billing, Data Import, Integrations, Access Control, Compliance, Known Issues
Escalation rulesConfigured rules (Excel for pilot)5 rules mapping conditions → teams (Engineering, Integrations, Finance, Compliance, Platform Ops)

Outputs

OutputDescriptionExample
ClassificationCategory + Priority + SentimentCategory: Authentication Priority: High Sentiment: Frustrated
Retrieved KB articlesTop-k relevant articles with relevance scores1. KB-AUTH-001: SSO Configuration Guide (0.92) 2. KB-AUTH-003: Password Reset Procedures (0.87)
Action recommendationReply / Ask for more info / EscalateRecommended action: Reply Confidence: 0.91
Draft responseGrounded 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 scorePer-output confidence indicatorOverall confidence: High (0.91)
Reasoning traceWhy 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

#MetricTypeTargetMeasurement
1Classification accuracyAI quality>= 85%Exact match on category and priority against golden set
2Retrieval accuracyAI quality>= 85%Expected KB article appears in top-k retrieved set
3Action accuracyAI quality>= 85%Exact match on next-best-action (Reply / Ask / Escalate)
4Response acceptance rateBusiness>= 70%% of drafts agents accept without major rewriting
5Auto-answer coverageBusiness30% of recurring tickets% of well-defined tickets the copilot can handle
6Autonomous operationBusinessNo handholdingSystem answers queries without manual intervention

Go-live validation: Run against 1000 synthetic questions, reviewed by support team lead.


Hard Limits

ConstraintLimitRationale
Human-in-the-loopAlways -- no auto-sendAgent must review and approve every response
Profanity / misuseZero toleranceAll outputs must be checked; adversarial inputs handled gracefully
LLM vendor lock-inMust be LLM-agnosticArchitecture must support swapping LLM providers
PortabilityFull on-premises handover post-pilotCode, models, data -- everything transferable to client premises
TimelineDemo by May 10, final delivery May 16Hard deadlines, not flexible
HallucinationResponses must be grounded in KBNo fabricated information; cite sources or flag "no match"

Data Sources

SourceRecordsAccess (Pilot)Access (Production)Sensitivity
Historical tickets36Excel datasetFreshworks APILow (simulated data for pilot)
KB articles12Excel datasetFreshworks KB APILow
Escalation rules5Excel datasetFreshworks / configLow
Evaluation set12 (held out)Excel datasetExpanded to 1000 syntheticLow

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)

ItemReason
Freshdesk API integrationPilot uses Excel; architect for it but don't build yet
Multi-language supportEnglish only for pilot
Customer-facing AIAgent-facing only; customers never interact with the copilot
Auto-send / auto-replyHuman always in the loop
Live KB refreshNot needed with static Excel data
Production deploymentPilot is standalone; productionization note covers the path forward
Phase 2 scopeNot defined; to be discussed after pilot results
Load testing / scalingNot applicable for pilot with demo + eval run

Open Questions

#QuestionOwnerStatus
1Target cost-per-ticket for production?Shivani → PrasannaPending -- ask in next communication
2Confidence indicator design -- binary or three-tier?Amit → propose to PrasannaPending
3Which LLM to recommend for this use case?AmitTo be decided in Architecture Sketch
4Vector store selection (pgvector / Qdrant / ChromaDB / Elasticsearch)?AmitTo be decided in Architecture Sketch
5How many synthetic questions to generate per category?Nishka + AmitTo be decided in Evaluation Plan
6Feedback loop storage -- where do agent edits persist?Amit + NancyTo be decided in Architecture Sketch