Skip to main content

Discovery Call Answers

Date: 2026-04-29, 3:00 PM IST Duration: 60 minutes Participants: Amit (POD Lead), Shivani (Implementation Manager), Prasanna (Client) Format: Google Meet


Section 1: The Problem -- Current Workflow & Pain Points

Current Workflow

  • Support ticket gets raised in Freshdesk
  • A human agent picks it up manually
  • Agent does their own research -- searching KB manually, asking colleagues, relying on memory
  • Agent sends the resolution

Pain Points

  • Resolution time ranges from 2 hours to 2 days -- highly variable
  • No standardization -- each agent researches and responds differently
  • Knowledge trapped in individual agents' heads
  • High human cost for repetitive, well-documented issues

Volume

  • ~1000 tickets/month currently

Pilot Goal

  • 30% of recurring tickets (those with defined resolutions, e.g., common Authentication issues) to be answered autonomously by the copilot
  • Human agent remains the final layer of approval (human-in-the-loop)
  • Standardize responses across agents
  • Reduce resolution time significantly
  • Save on human cost

Section 2: Platform & Integrations

Source System

  • Freshworks suite is the single source of truth -- KB, tickets, everything
  • Entire sales and support infrastructure lives in Freshworks

Pilot Approach

  • Excel dataset for the pilot -- no Freshdesk integration needed yet
  • Architect for Freshworks API integration post-pilot
  • No KB refresh feature needed for pilot, but design for it (production will need it)

LLM & Infrastructure

  • LLM: Open to any provider for the pilot; wants flexibility and LLM-agnostic architecture; open to our recommendation for this use case
  • Cloud: GCP for the pilot (our own GCP account)
  • Post-pilot: Full on-premises handover -- code, models, data, everything goes to the client's premises
  • Implication: Keep the stack portable, avoid vendor lock-in

Deployment Surface

  • Prasanna leans toward a sidebar widget on Freshdesk
  • Our recommendation (agreed): Chrome extension / sidebar overlay -- fastest to demo, doesn't require Freshdesk marketplace integration, shows real value in context

Database

  • Our call for the pilot

Section 3: Success Criteria

What Success Looks Like

  1. The pilot answers queries autonomously without handholding
  2. Accuracy is good -- target 85% for the pilot, improve from there

Go-Live Validation

  • Run the agent against 1000 synthetic questions before going live
  • We generate the synthetic questions from existing data
  • Support team lead reviews and evaluates the outputs
  • Adjustments can be made based on evaluation results

Specific Thresholds

  • 85% accuracy target for pilot (we will propose detailed per-metric thresholds in the Evaluation Plan)
  • 30% of recurring tickets auto-answered (with human approval)

Section 4: Scope & Boundaries

In Scope

  • General inbox -- not a single narrow queue
  • Scope defined by what the KB covers (7 categories from dataset: Authentication, Billing, Data Import, Integrations, Access Control, Compliance, Known Issue)
  • English only for the pilot
  • Human-in-the-loop always

Out of Scope

  • Multi-language support
  • Live Freshdesk integration (pilot uses Excel)
  • Customer-facing AI (agent-facing only)
  • Auto-send (copilot suggests, agent decides)

Phase 2

  • Not defined yet -- to be discussed after pilot results

Section 5: Data & Privacy

  • No PII concerns for the pilot -- using Excel with simulated data, no real customer data
  • No data classification or retention requirements for the simulation
  • No KB refresh feature needed for pilot
  • For production: architect for data refresh (Freshworks API integration), PII handling, and classification -- Shubham should document what production would require

Section 6: Delivery & Cadence

Communication Cadence (Agreed)

  • One weekly email status update
  • One weekly call (Google Meet)
  • Total: 2 touchpoints per week

Channels

  • Email for written updates
  • Google Meet for calls

Decision Turnaround

  • 2 hours to 1 day -- Prasanna is responsive
  • All team members on his side available for support

Timeline (Hard Deadlines)

MilestoneDateDays from now
Sprint 1 DemoMay 10, 202611 days
Final DeliveryMay 16, 202617 days
  • These are hard deadlines, not flexible

Section 7: Deliverables & Cost

Deliverables

  • Working pilot AND documentation are equally important
  • Knowledge transfer and docs are priority from day one
  • Everything we build should be documented and transferable
  • Post-pilot, the client takes forward the code and knowledge independently

Expected Deliverables

  1. Working pilot (standalone web app -- three-panel dashboard)
  2. Architecture document
  3. Evaluation results (including 1000-question run)
  4. Sample outputs across categories
  5. Productionization note
  6. Complete codebase for handover

Cost

  • No budget constraints for the pilot (LLM API costs, infrastructure)
  • Our own GCP account for the pilot
  • Target cost-per-ticket: Not discussed yet -- ACTION: Ask Prasanna in next communication

Section 8: Governance

Confirmed

  • Human-in-the-loop: Copilot suggests, agent decides, no auto-send

Feedback Loop (New Requirement)

  • Agent can rate response ("was it helpful")
  • Agent can edit the response
  • Edited version saved as a correct response
  • System learns from feedback over time -- golden evaluations from real usage
  • This creates a positive feedback loop for continuous improvement

Guardrails (New Requirement)

  • Profanity check on all outputs
  • Handle negative/adversarial scenarios gracefully
  • Prevent misuse of the system
  • Graceful failure for edge cases

Not Yet Confirmed

  • Compliance/security review requirements (not applicable for simulation)
  • Confidence indicator approach (to be proposed by us)

Key Decisions Summary

DecisionAnswer
Data source for pilotExcel dataset
Data source for productionFreshworks APIs
Cloud platformGCP (our account)
Post-pilot deploymentOn-premises handover
LLM providerGoogle Gemini via Vertex AI (LLM-agnostic architecture)
EmbeddingsVertex AI text-embedding-005
Deployment (pilot)Standalone web app -- three-panel dashboard
Deployment (production)Chrome extension / Freshdesk sidebar
BackendExpress (Node.js)
FrontendReact
Operational DBMongoDB
Search / RetrievalElasticsearch (hybrid: vector + BM25)
OrchestrationLangChain.js
AuthGoogle Service Account (single identity for all GCP services)
Accuracy target85% for pilot
Auto-answer target30% of recurring tickets
Go-live validation1000 synthetic questions
EvaluatorSupport team lead
Sprint 1 demoMay 10 (hard)
Final deliveryMay 16 (hard)
Communication1 email + 1 call per week
ChannelEmail + Google Meet
Decision turnaround2 hours to 1 day
Docs priorityEqual to working code
Feedback loopAgent rates + edits, system learns
GuardrailsProfanity, misuse prevention

Open Items / Follow-ups

#ItemOwnerStatus
1Target cost-per-ticket for productionShivani (ask Prasanna)Pending
2Confidence indicator design (propose to Prasanna)AmitPending
3Document production-grade PII handling requirementsShubhamPending
4Document production-grade data classificationShubhamPending
5Phase 2 scope discussionShivani + PrasannaPost-pilot

Next Steps

#TaskOwnerDeadline
1Draft Use Case CanvasAmit + ShivaniApr 30
2Draft POD CharterShivani + AmitApr 30
3Draft Architecture SketchAmitMay 1
4Draft Evaluation PlanAmit + NishkaMay 1
5Draft Risk RegisterShivaniMay 1
6Draft Sprint PlanShivani + AmitMay 1
7Start Data Feasibility ReportNancyMay 1
8Start Threat ModelShubhamMay 1
9Brief full team on Discovery outcomesAmit + ShivaniApr 30 standup
10Share drafts with Prasanna for sign-offShivaniMay 2