Skip to main content

Engagement Plan -- AI Support Copilot

Engagement: AI Support Copilot Pilot Owner: Shivani (PM) Version: 1.0 Date: 2026-05-01 Framework ref: Doc 03, Section 7


1. Executive Summary

Gyde AI POD will deliver a working AI Support Copilot that enables support agents to resolve tickets faster and more consistently. The copilot automates classification, knowledge retrieval, action recommendation, and response drafting -- with human approval on every response.

Mission: Achieve 85% accuracy across classification, retrieval, and action recommendation, with 30% of recurring tickets handled autonomously (human-approved), within a 16-day pilot using curated data.

Key success metrics:

  • Classification accuracy >= 85%
  • Retrieval accuracy >= 85%
  • Action accuracy >= 85%
  • Response acceptance rate >= 70%
  • Auto-answer coverage >= 30%

Duration: 16 days (May 1 -- May 16, 2026) Milestones: M0 (Charter signed) → M1 (Walking skeleton) → M2 (Eval harness) → M3 (MVP feature-complete) Team: 6-person POD (POD Lead, AI Engineer, Data Engineer, QA, Governance, PM) Top risks: Dataset diversity (Critical), tight timeline (Critical), LLM accuracy uncertainty (High)


2. Scope

In Scope (Phase 1 Pilot)

#CapabilityDescription
1Ticket classificationCategory, priority, sentiment, confidence via LLM
2KB retrievalHybrid search (vector + BM25) over Elasticsearch
3Action recommendationReply / Ask for more info / Escalate with reasoning
4Response draftingGrounded response with KB citations
5Confidence scoringPer-step confidence indicators
6Feedback loopAgent rates/edits responses; corrections stored
7GuardrailsProfanity filter, PII check, misuse prevention, confidence gating
8Web applicationThree-panel dashboard (ticket queue, detail, copilot sidebar)
9Evaluation harnessAutomated scoring against golden dataset
10Synthetic evaluation1,000-question eval run reviewed by client's support lead
11DocumentationArchitecture docs, model card, knowledge transfer package

Out of Scope (Phase 1)

#ExclusionRationale
1Freshdesk API integrationArchitecture designed for it; not built in pilot (uses Excel data)
2Multi-language supportEnglish only for pilot
3Customer-facing AIAgent-facing only; no direct customer interaction
4Auto-send without approvalHuman always approves before sending
5Live KB refreshStatic dataset; real-time refresh is production scope
6Production deploymentPilot validates feasibility; production is a separate engagement
7Load testing / scalingNot needed for pilot; noted in productionization doc
8Phase 2 featuresNot yet defined by client

3. Milestone Schedule

Timeline Overview

May 1 May 5 May 10 May 14 May 16
|-------- Sprint 1 --------|----------- Sprint 2 ------------|
M0 mid M1+M2 mid M3
Charter Golden Skeleton Synth MVP
signed set + Harness eval delivery

Milestone Details

#MilestoneTarget DateExit Criteria
M0Charter & Eval Plan signedMay 1POD Charter signed; Evaluation Plan agreed; GCP environment provisioned; threat model in flight
M1Walking skeletonMay 10One ticket → classify → retrieve → reason → draft → UI display, working end-to-end in dev
M2Eval harness operationalMay 10Golden dataset committed (30-40 cases); automated scoring running; baseline metrics published
M3MVP feature-completeMay 16All scope items built; eval metrics at target for 2 consecutive runs; security review passed; documentation delivered

Note: M1 and M2 are targeted for the same Sprint 1 demo (May 10). This is intentional -- the walking skeleton produces the outputs the harness needs to score. M4-M6 (production deployment, stable operation, engagement close) are out of scope for this pilot.


4. Estimation Summary

Effort by Bucket

BucketScopeShareNotes
Data workExcel ingestion, KB indexing, embeddings, data quality, synthetic data generation~25%Lower than typical (20-40%) because pilot uses curated Excel, not messy production data
AI engineeringPrompts, pipeline (classify/retrieve/reason/draft), retrieval tuning, confidence scoring, feedback loop~30%Core of the engagement; highest uncertainty
Application engineeringExpress API, React UI (three-panel), LLM Gateway, MongoDB integration, ES integration~25%Higher than typical (15-25%) because of standalone web app requirement
Governance & securityThreat model, guardrails (profanity, PII, injection), security review~10%At framework minimum; pilot is internal-facing
Quality, ops, & releaseEval harness, golden dataset, adversarial cases, synthetic eval, CI integration, documentation~10%At framework minimum; eval harness is critical path

Estimation Method

Primary method: Time-boxed (Doc 03, Section 4.2). Duration is fixed at 16 days; scope is shaped to fit. The contingency plan in the Risk Register defines pre-approved scope cuts.

Estimation modifiers applied:

  • Data quality: +0% (curated Excel dataset, clean)
  • Novel use case: +0% (RAG over KB is a known pattern with Gyde reference architecture)
  • Internal-facing only: -10% (no customer-facing surface risk)
  • Tight timeline: +15% (compressed from typical 6-week build into 16 days)
  • Net modifier: +5%

Confidence Level

Given the fixed timeline and hard deadlines, confidence in delivering all scope items is Medium. The contingency plan mitigates this -- core capabilities (pipeline + eval) are protected; polish and expansion features are the flex.


5. Team & Engagement Model

POD Composition

RoleNameAllocationKey Deliverables
POD LeadAmitFull-timeArchitecture, UI, code review, demos, tech decisions
AI EngineerAtharvaFull-timePipeline (classify, retrieve, reason, draft), prompts, confidence scoring, feedback loop
Data EngineerNancyFull-timeData ingestion, KB indexing, embeddings, vector store, data quality
QANishkaFull-timeEval harness, golden dataset, adversarial cases, synthetic eval
Governance EngineerShubhamPart-timeThreat model, guardrails, security review
Implementation ManagerShivaniPart-timeCharter, sprint planning, status reports, risk register, client comms

Client Counterparts

Client RoleNameResponsibilities
Client Sponsor (CIO + Business Lead + Product Owner)PrasannaDecisions, sign-offs, success criteria, domain expertise
Support Team LeadTBD (via Prasanna)Reviews synthetic eval results, validates accuracy

Engagement Model

  • POD operates as a self-contained delivery unit
  • All team members report to POD Lead (Amit) for technical decisions
  • PM (Shivani) manages client communication and engagement logistics
  • Client has a single point of contact (Shivani for process, Amit for technical)

6. Risk Register Summary

Top 5 risks (full register in separate document):

#RiskScoreOwner
R-05Low dataset diversity (11 unique scenarios)CriticalNishka + Atharva
R-09Tight timeline (16 days) with hard deadlinesCriticalShivani + Amit
R-01Gemini accuracy may not reach 85%HighAtharva + Amit
R-07KB articles lack sufficient depthHighAmit + Nancy
R-10Team members pulled to competing prioritiesHighShivani

See Risk Register for full details including mitigation plans, assumptions, dependencies, and contingency plan.


7. Assumptions & Dependencies

Key Assumptions

#AssumptionValidated?
A-01Excel dataset is representative of production ticket patternsNo
A-02Prasanna available for decisions within 1 business dayYes
A-03GCP Vertex AI APIs stable throughout pilotNo
A-0485% accuracy achievable with provided KB contentNo
A-05Team members dedicated (no competing priorities)No

Key Dependencies

#DependencyProviderNeeded ByStatus
D-01GCP Service Account with Vertex AI permissionsAmitMay 2Pending
D-02GCP VM for MongoDB + ElasticsearchAmitMay 2Pending
D-03Dataset (Excel)PrasannaNowDone
D-04Client review of expanded golden setPrasannaMay 5Pending
D-05Client review of 1,000 synthetic eval resultsSupport leadMay 15Pending
D-06Decision on target cost-per-ticketPrasannaNext weekly callPending

8. Commercial Summary

ItemDetail
Engagement typePilot / Proof of Concept
Duration16 days (May 1 -- May 16, 2026)
Team size6 (4 full-time, 2 part-time)
InfrastructureGyde's GCP account for pilot; client bears no infra cost during pilot
LLM costsBorne by Gyde during pilot (Vertex AI / Gemini API costs)
Budget constraintsNone specified by client for pilot phase
Post-pilotFull codebase and knowledge transfer to client; client deploys on own infrastructure
Change request processPer Doc 03, Section 8 -- any scope/schedule/cost change logged, impact-assessed within 3 days, decided by client sponsor + Gyde

9. Governance & Reporting

Communication Cadence

ActivityFrequencyChannelParticipants
POD standupDailyInternal (Slack/Meet)Full POD
Weekly status emailWeeklyEmailPrasanna, POD
Weekly sync callWeeklyGoogle Meet (30 min)Prasanna, Amit, Shivani
Sprint demoPer sprint (May 10, May 16)Google MeetPrasanna, full POD
Sprint retroPer sprintInternalFull POD

Status Report Format

Weekly status email includes:

  • Sprint progress (stories completed / in progress / blocked)
  • Eval metrics delta (if harness is operational)
  • Top 3 risks with status changes
  • Blockers requiring client action
  • Next week's plan

Decision Escalation

TierAuthorityExamples
POD-InternalPOD Lead + owning roleLibrary choice, prompt structure, code style
POD LeadAmitArchitecture patterns, model selection, release readiness
Client ApprovalPrasanna via ShivaniScope changes, milestone shifts, success criteria changes
Gyde LeadershipEngineering DirectorFramework non-negotiable deviations

10. Non-Negotiable Compliance

Per Doc 01, Section 5.1, this engagement adheres to all five framework non-negotiables:

#Non-NegotiableFulfillment
1Threat modeling and secrets managementShubham delivers threat model; secrets via GCP Secret Manager or env vars
2Evaluation before productionEval harness gates every release; 1,000-question run before delivery
3Versioned data and promptsAll prompts, datasets, configs in Git
4Audit trail for AI decisionsEvery copilot decision logged with full pipeline output
5Incident response readinessRunbooks in knowledge transfer package

Change Log

DateChangeBy
2026-05-01Initial engagement plan createdShivani + Amit

This plan is the definitive reference for the engagement. Any changes to scope, milestones, or commercial terms follow the Change Request process (Doc 03, Section 8). The plan is signed off by the client sponsor and triggers Sprint 1.