ms-ai-architect/skills/ms-ai-advisor/references/architecture/poc-template.md

POC Template - Microsoft AI Projects

Last updated: 2026-06-24 (research via microsoft-learn MCP) Category: Solution Architecture & Advisory Status: Reference

---

Dette dokumentet tilbyr en strukturert mal for å planlegge, gjennomføre og evaluere Proof of Concept (POC) prosjekter for Microsoft AI-løsninger. Malen er tilpasset Microsoft Foundry, Copilot Studio, Power Platform AI, og andre Microsoft AI-plattformer.

Innhold

1. POC Plan Template
2. Success Criteria Framework
3. Evaluation Rubric
4. Platform-Specific Checklists
5. Risk Assessment Template
6. Timeline Templates
7. Stakeholder Communication Template
8. Go/No-Go Decision Framework
9. Example POC Plan

---

POC Plan Template

Bruk denne malen for å strukturere din POC-plan. Fyll ut hver seksjon basert på ditt spesifikke use case.

1. Executive Summary

Hensikt med POC: [1-2 setninger: Hva skal POC bevise eller validere?]

Forventet varighet: [1 uke / 2 uker / 4 uker]

Estimert ressursbehov: [Antall personer, roller, budsjett]

Beslutningspunkt: [Dato for go/no-go beslutning]

---

2. Business Case

2.1 Problem Statement

[Beskriv forretningsproblemet eller ineffektiviteten som AI kan løse.]

Eksempel:

Kundestøtte bruker 40% av tiden på repetitive spørsmål om ordrestatus, produktreturer og leveringsinformasjon. Dette binder opp ressurser som kunne brukes på mer komplekse kundehenvendelser.

2.2 Target Outcome

[Hva er det ønskede resultatet? Vær konkret og målbar.]

Eksempel:

Automatisere 60% av repetitive kundehenvendelser via chatbot, redusere gjennomsnittlig responstid fra 15 minutter til 2 minutter, og frigjøre 16 timer per uke for støtteteamet.

2.3 Strategic Value

Ranger strategisk verdi (1-5, hvor 5 er høyest):

• Business Impact: [1-5] — Hvor stor påvirkning har dette på forretningen?
• User Value: [1-5] — Hvor mye verdi gir dette til sluttbrukere?
• Innovation Potential: [1-5] — Hvor innovativt er dette for organisasjonen?
• Strategic Alignment: [1-5] — Hvor godt aligner dette med organisasjonens AI-strategi?

---

3. Technical Scope

3.1 AI Maturity Assessment

Identifiser din organisasjons AI-modningsnivå (basert på Microsoft CAF):

Level	Skills Required	Data Readiness	Feasible Use Cases
Level 1	Basic AI-forståelse, data-integrasjon	Minimal data, enterprise data tilgjengelig	Azure quickstart, Copilot-løsninger
Level 2	Model selection, deployment, data cleaning	Små strukturerte datasett, domene-spesifikk data	Analytical AI (Foundry Tools), Custom gen AI chat uten RAG, Fine-tuning
Level 3	Prompt engineering, data chunking, preprocessing	Store historiske datasett, domene-spesifikk data	Gen AI med RAG, ML model training, small AI models på VMs
Level 4	Advanced AI/ML, infra management, orchestration	Store treningsdatasett	Large gen AI/ML apps på VMs, AKS, Container Apps

Din organisasjon er på: [Level 1 / 2 / 3 / 4]

3.2 Chosen AI Solution

[Velg én eller flere:]

• Microsoft 365 Copilot (extensions/agents)
• Copilot Studio (custom agents)
• Microsoft Foundry (custom gen AI apps)
• Power Platform AI (AI Builder, Power Automate AI)
• Azure Machine Learning (custom ML models)
• Analytical AI (Content Safety, Document Intelligence, Custom Vision)

Rationale: [Hvorfor er denne løsningen valgt? Hva gjør den til det beste valget for dette use case?]

3.3 Data Requirements

Data Sources:

1. [Source 1: Type, format, quality, accessibility]
2. [Source 2: Type, format, quality, accessibility]
3. [Source 3: ...]

Data Preparation Needed:

• Data cleaning/normalization
• Data labeling
• Data chunking (for RAG)
• Privacy/security review (PII removal, anonymization)
• Data governance approvals

Estimated Data Volume: [Small (<100 MB) / Medium (100 MB - 10 GB) / Large (>10 GB)]

3.4 Infrastructure Requirements

Compute:

• Azure OpenAI capacity (model, region, quota)
• Azure Machine Learning compute (SKU, vCPUs, GPU)
• Power Platform capacity (Copilot Studio, AI Builder credits)

Storage:

• Azure Storage (type, size)
• Vector database (Azure AI Search, Cosmos DB)

Network:

• VNet integration
• Private endpoints
• Bandwidth requirements

Security & Compliance:

• Azure Policy enforcement
• Content Safety filters
• Data residency requirements
• Authentication/authorization (Entra ID, RBAC)

---

4. Success Criteria

Definer spesifikke, målbare kriterier for POC-suksess. Se Success Criteria Framework for detaljerte KPIer.

4.1 Technical Success Criteria

1. Functional Requirements:

   • [Eksempel: Chatbotten skal kunne svare på 80% av testspørsmålene korrekt.]
   • [Eksempel: Systemet skal kunne håndtere 100 samtidige forespørsler.]

2. Performance Metrics:

   • Response Time: [Target: < 3 sekunder for 95% av forespørslene]
   • Accuracy: [Target: 85% nøyaktighet på validasjonsdatasett]
   • Availability: [Target: 99% uptime under testperioden]

3. Quality Metrics (for Gen AI):

   • Groundedness: [Target: 90% av svarene skal være faktabaserte]
   • Relevance: [Target: 85% av svarene skal være relevante for brukerens spørsmål]
   • Content Safety: [Target: 0% harmful content, 100% moderate risk filtered]

4.2 Business Success Criteria

1. Efficiency Gains:

   • [Eksempel: Redusere behandlingstid med 50%]

2. Cost Savings:

   • [Eksempel: Redusere driftskostnader med 20%]

3. User Satisfaction:

   • [Eksempel: Oppnå 70% user satisfaction score]

4.3 Responsible AI Criteria

• Fairness: Løsningen skal ikke diskriminere basert på alder, kjønn, etnisitet, etc.
• Transparency: Brukere skal forstå når de interagerer med AI
• Privacy: Persondata skal beskyttes i henhold til GDPR/compliance-krav
• Accountability: Klare roller og ansvar for AI-beslutninger
• Safety: Content Safety filters implementert og testet
• Inclusiveness: Løsningen skal fungere for alle brukergrupper

---

5. Implementation Plan

5.1 Phases & Milestones

Phase 1: Prepare (Duration: [X dager])

• Data collection and preparation
• Environment setup (Azure, Power Platform)
• Team onboarding
• Security/compliance approvals

Deliverable: [Data ready for use, infrastructure provisioned]

---

Phase 2: Build (Duration: [X dager])

• Develop initial prototype/POC
• Implement core functionality
• Integrate data sources
• Configure AI model/agent

Deliverable: [Working prototype in dev environment]

---

Phase 3: Evaluate & Iterate (Duration: [X dager])

• Functional testing
• Performance testing
• Responsible AI testing (fairness, safety, bias)
• User acceptance testing (UAT)
• Iterate based on feedback

Deliverable: [Validated POC with test results]

---

Phase 4: Document & Decide (Duration: [X dager])

• Document lessons learned
• Compile evaluation report
• Prepare go/no-go recommendation
• Present to stakeholders

Deliverable: [POC report + go/no-go decision]

---

5.2 Team Roles & Responsibilities

Role	Responsible For	Time Commitment
Project Lead	Overall POC coordination, stakeholder communication	[X hours/week]
Solution Architect	Technical design, platform selection	[X hours/week]
Data Scientist/Engineer	Data preparation, model evaluation	[X hours/week]
Developer/Maker	Building prototype (Copilot Studio, Power Platform, code)	[X hours/week]
Subject Matter Expert (SME)	Domain knowledge, validation	[X hours/week]
Security/Compliance Officer	Responsible AI review, compliance validation	[X hours/week]
End-User Representative	User testing, feedback	[X hours/week]

---

6. Testing & Validation Plan

6.1 Functional Testing

• Unit tests for individual components
• Integration tests for data pipelines
• End-to-end scenario testing

Test Cases: [Liste av testscenarier, f.eks. "User asks about order status"]

6.2 Performance Testing

• Load testing (concurrent users/requests)
• Latency testing (response times)
• Throughput testing (requests per second)

6.3 Responsible AI Testing

• Fairness Assessment: Test på diverse brukergrupper
• Content Safety: Test adversarial prompts (jailbreak, harmful content)
• Bias Detection: Evaluate model outputs for bias
• Explainability: Validate that model decisions are understandable

Tools:

• Microsoft Foundry evaluation tools
• Azure AI Content Safety
• Responsible AI Dashboard (Azure ML)

6.4 User Acceptance Testing (UAT)

• Recruit representative users
• Define UAT scenarios
• Collect qualitative feedback (surveys, interviews)
• Measure user satisfaction (NPS, CSAT)

---

7. Risk Management

Se Risk Assessment Template for detaljert risikovurdering.

High-Priority Risks:

1. [Risk 1: Description + Mitigation Plan]
2. [Risk 2: Description + Mitigation Plan]
3. [Risk 3: Description + Mitigation Plan]

---

8. Budget & Resources

Estimated Costs:

Category	Estimated Cost	Notes
Azure Compute	[NOK/USD]	OpenAI quota, VM SKUs, AML compute
Storage	[NOK/USD]	Blob Storage, AI Search
Licensing	[NOK/USD]	Copilot Studio, Power Platform
Personnel	[NOK/USD]	Team member time (internal/external)
Contingency (20%)	[NOK/USD]	Buffer for unexpected costs
TOTAL	[NOK/USD]

---

9. Go/No-Go Decision Criteria

Ved slutten av POC, evaluer mot disse kriteriene:

• Technical Feasibility: Løsningen fungerer som forventet (>80% success criteria oppfylt)
• Business Value: ROI er positiv, eller verdi er dokumentert
• User Acceptance: Brukere er fornøyde (>70% satisfaction)
• Responsible AI: Ingen kritiske fairness/safety issues
• Risk Acceptable: Identifiserte risikoer kan håndteres
• Budget Viable: Production deployment er innenfor budsjett

Decision: [GO / NO-GO / CONDITIONAL GO (specify conditions)]

---

Success Criteria Framework

Technical KPIs (Generative AI)

Metric	Definition	Target Range	Measurement Method
Groundedness	% of responses supported by source data	>85%	Microsoft Foundry evaluation
Relevance	% of responses relevant to user query	>80%	Microsoft Foundry evaluation
Fluency	% of responses that are coherent and grammatical	>90%	Microsoft Foundry evaluation
Content Safety	% of harmful content blocked	100%	Azure AI Content Safety
Response Time	Average latency (seconds)	<3s (p95)	Application Insights
Throughput	Requests per second handled	>100 rps	Load testing
Availability	Uptime during test period	>99%	Azure Monitor

Business KPIs

Metric	Definition	Target	Measurement Method
Time Saved	Hours saved per week	[X hours]	Before/after comparison
Cost Reduction	% reduction in operational costs	[X%]	Financial analysis
User Satisfaction (CSAT)	Customer satisfaction score (1-5)	>4.0	Survey
Net Promoter Score (NPS)	Likelihood to recommend (0-10)	>7.0	Survey
Task Completion Rate	% of user tasks successfully completed	>80%	Analytics
Adoption Rate	% of target users actively using solution	>60%	Usage analytics

Responsible AI KPIs

Metric	Definition	Target	Measurement Method
Fairness (Demographic Parity)	Max difference in positive prediction rates across groups	<10%	Responsible AI Dashboard
Bias Detection	No significant bias detected in outputs	0 critical issues	Manual review + automated tools
Privacy Compliance	% of PII correctly handled (removed/anonymized)	100%	Data audit
Content Safety Pass Rate	% of responses passing content safety filters	100%	Azure AI Content Safety
Explainability Score	% of users who understand AI decisions	>70%	User survey

---

Evaluation Rubric

Bruk denne matrisen for å score POC-resultater:

Technical Performance

Criterion	Score 1 (Poor)	Score 3 (Fair)	Score 5 (Good)	Score 7 (Excellent)	Score
Accuracy/Quality	<60%	60-74%	75-89%	≥90%	[X]
Performance	Frequent failures, >5s latency	Occasional failures, 3-5s latency	Stable, 2-3s latency	Highly stable, <2s latency	[X]
Reliability	<95% uptime	95-97% uptime	97-99% uptime	>99% uptime	[X]
Scalability	Cannot scale beyond POC	Limited scalability	Scales to production	Easily scales	[X]

Technical Score: [Sum / 28] → [%]

---

Business Value

Criterion	Score 1 (Poor)	Score 3 (Fair)	Score 5 (Good)	Score 7 (Excellent)	Score
Efficiency Gains	<20% improvement	20-40%	40-60%	>60%	[X]
User Satisfaction	<50% satisfied	50-65%	65-80%	>80%	[X]
Cost-Effectiveness	ROI negative	ROI break-even	ROI 1-2x	ROI >2x	[X]
Strategic Fit	Misaligned	Partially aligned	Well aligned	Critical priority	[X]

Business Score: [Sum / 28] → [%]

---

Responsible AI

Criterion	Score 1 (Poor)	Score 3 (Fair)	Score 5 (Good)	Score 7 (Excellent)	Score
Fairness	Significant bias issues	Minor bias detected	Fair across groups	Highly fair	[X]
Safety	Harmful content generated	Moderate safety issues	Safe with minor exceptions	100% safe	[X]
Privacy	PII leaks detected	Minor privacy concerns	Privacy compliant	Exceeds compliance	[X]
Transparency	Opaque, users confused	Somewhat transparent	Transparent	Highly transparent	[X]

Responsible AI Score: [Sum / 28] → [%]

---

Overall POC Score

Dimension	Weight	Score (%)	Weighted Score
Technical Performance	40%	[X%]	[X]
Business Value	40%	[X%]	[X]
Responsible AI	20%	[X%]	[X]
TOTAL	100%		[X%]

Recommendation:

• >80%: Strong GO — Proceed to production
• 60-80%: Conditional GO — Address gaps before production
• <60%: NO-GO — Re-evaluate or pivot

---

Platform-Specific Checklists

Copilot Studio POC Checklist

Pre-Development:

• Define agent scope (which topics/intents)
• Identify data sources for grounding (SharePoint, Dataverse, APIs)
• Determine deployment channels (Teams, website, custom)
• Configure Copilot Studio environment (dev, test, prod)
• Set up authentication (if required)

Development:

• Build initial topics using conversational design best practices
• Configure generative orchestration (if using gen AI)
• Integrate data sources (connections, AI Search)
• Implement content moderation (Azure AI Content Safety)
• Test conversation flows with representative users

Evaluation:

• Test intent recognition accuracy
• Measure conversation abandonment rate
• Validate grounding accuracy (if using data sources)
• Test escalation paths (handoff to human)
• Collect user feedback via surveys

Governance:

• Apply content filters (Azure Policy)
• Configure security groups (Entra ID)
• Review compliance (data residency, privacy)
• Document agent behavior and limitations

---

Microsoft Foundry POC Checklist

Pre-Development:

• Select foundation model (GPT-4o, GPT-4, custom)
• Provision Azure OpenAI capacity (region, quota)
• Define prompt engineering strategy
• Identify grounding data (if RAG)
• Set up Azure AI Search (if RAG)

Development:

• Build prompt flow orchestration
• Implement RAG pipeline (chunking, embedding, retrieval)
• Configure content safety filters
• Develop evaluation dataset (test queries + expected outputs)
• Deploy to pre-production endpoint

Evaluation:

• Run Microsoft Foundry evaluation suite (groundedness, relevance, fluency)
• Test adversarial prompts (jailbreak attempts)
• Measure latency and throughput
• Validate cost per request
• Collect SME feedback on output quality

Governance:

• Enforce Azure Policy (allowed models, regions)
• Configure RBAC for deployment
• Enable monitoring (Application Insights, Azure Monitor)
• Document model version and configuration

---

Power Platform AI (AI Builder) POC Checklist

Pre-Development:

• Identify AI Builder capability (document processing, text classification, object detection)
• Prepare training data (labeled datasets)
• Validate Power Platform capacity (AI Builder credits)
• Define integration points (Power Apps, Power Automate)

Development:

• Train AI Builder model
• Validate model accuracy on test dataset
• Build Power Automate flow or Power App integration
• Test end-to-end automation

Evaluation:

• Measure model precision/recall
• Test on real-world data
• Validate processing speed
• Collect user feedback

Governance:

• Configure DLP policies
• Review data residency
• Document model performance metrics

---

Risk Assessment Template

Risk Identification Matrix

Risk Category	Risk Description	Likelihood (1-5)	Impact (1-5)	Risk Score (L×I)	Mitigation Plan
Technical	[Example: Model accuracy below target]	[X]	[X]	[X]	[Retrain with more data, fine-tune prompts]
Data	[Example: Insufficient training data]	[X]	[X]	[X]	[Synthetic data generation, expand data sources]
Security	[Example: PII leakage in outputs]	[X]	[X]	[X]	[Implement PII detection, anonymization]
Compliance	[Example: GDPR violation]	[X]	[X]	[X]	[Legal review, data residency controls]
Organizational	[Example: Lack of user adoption]	[X]	[X]	[X]	[Change management, training, communication]
Budget	[Example: Cost overruns]	[X]	[X]	[X]	[Monitor spending, set cost alerts]
Responsible AI	[Example: Bias in model outputs]	[X]	[X]	[X]	[Fairness testing, diverse training data]

Risk Scoring:

• 1-5: Low risk (monitor)
• 6-10: Medium risk (active mitigation required)
• 11-25: High risk (escalate, consider showstopper)

---

Common AI POC Risks & Mitigations

Technical Risks

Risk: Model decay over time (accuracy degrades)

• Mitigation: Implement continuous monitoring, plan for retraining cadence

Risk: Latency exceeds user expectations

• Mitigation: Optimize prompt length, use faster models, implement caching

Risk: Integration failures with existing systems

• Mitigation: Early integration testing, API contract validation

---

Data Risks

Risk: Data quality issues (missing, incomplete, biased data)

• Mitigation: Data profiling upfront, data cleaning pipelines, diverse data sources

Risk: Insufficient data volume for training

• Mitigation: Synthetic data generation, transfer learning, start with simpler models

Risk: Data access blocked by governance/compliance

• Mitigation: Early stakeholder engagement, privacy-preserving techniques (anonymization)

---

Security & Compliance Risks

Risk: Prompt injection attacks

• Mitigation: Input validation, content filtering, prompt engineering defenses

Risk: Data residency violations

• Mitigation: Use compliant Azure regions, review data flow architecture

Risk: Unauthorized data access

• Mitigation: RBAC, private endpoints, encryption at rest/in transit

---

Organizational Risks

Risk: User resistance to AI adoption

• Mitigation: Involve users early, transparent communication about AI capabilities/limitations

Risk: Insufficient team skills

• Mitigation: Training programs, external consultants, phased learning approach

Risk: Unclear ownership and accountability

• Mitigation: Define RACI matrix, establish AI governance board

---

Timeline Templates

1-Week Sprint POC

Anbefalt for: Simple use cases (Copilot Studio med pre-built connectors, basic chatbot)

Day	Activities	Deliverables
Day 1	Kickoff, scope definition, environment setup	Approved scope, dev environment ready
Day 2-3	Build prototype, integrate data sources	Working prototype
Day 4	Testing (functional, UAT)	Test results, feedback collected
Day 5	Document findings, prepare recommendation	POC report, go/no-go decision

Total Effort: ~40 person-hours

---

2-Week Standard POC

Anbefalt for: Moderate complexity (Microsoft Foundry RAG, Copilot Studio med custom topics)

Week	Activities	Deliverables
Week 1	- Kickoff & planning (Day 1-2) - Data preparation (Day 2-3) - Environment setup (Day 3-4) - Initial prototype build (Day 4-5)	Data ready, dev environment, initial prototype
Week 2	- Iterate on prototype (Day 1-2) - Testing & validation (Day 3-4) - Documentation & presentation (Day 5)	Validated POC, test results, final report, go/no-go decision

Total Effort: ~80-120 person-hours

---

4-Week Extended POC

Anbefalt for: Complex use cases (Azure ML model training, multi-agent systems, advanced RAG)

Week	Phase	Activities	Deliverables
Week 1	Prepare	- Kickoff, detailed planning - Data collection & preparation - Infrastructure setup - Team onboarding	Data pipeline ready, infra provisioned, team aligned
Week 2	Build	- Develop core functionality - Model training/fine-tuning - Integration with systems	Working prototype (alpha)
Week 3	Evaluate	- Functional testing - Performance testing - Responsible AI evaluation - User acceptance testing - Iterate based on feedback	Validated prototype (beta), test reports
Week 4	Decide	- Final validation - Documentation (lessons learned, architecture) - Stakeholder presentation - Go/no-go decision	POC final report, production roadmap, decision

Total Effort: ~200-320 person-hours

---

Timeline Adjustment Factors

Legg til ekstra tid hvis:

• Data ikke er klar: +1-2 uker for data cleaning, labeling
• Komplekse integrasjoner: +1 uke per critical integration
• Compliance reviews: +1-2 uker for legal/security approvals
• New team to AI: +1 uke for onboarding/training
• Custom model training: +2-4 uker for ML model development

Anbefaling: Legg til 20-30% buffer for uforutsette utfordringer.

---

Stakeholder Communication Template

POC Kickoff Email

Subject: [POC Name] - Kickoff & Plan

To: Project team, stakeholders

Body:

Vi starter POC for [use case name] med mål om å [brief objective]. POC vil løpe fra [start date] til [end date] ([X uker]).

Mål:

• [Goal 1]
• [Goal 2]

Team:

• Project Lead: [Name]
• Solution Architect: [Name]
• Developer: [Name]

Neste steg:

• [Action 1]
• [Action 2]

Beslutningspunkt: [Date for go/no-go decision]

Spørsmål? Kontakt [Lead].

---

Weekly Status Update Template

Subject: [POC Name] - Week [X] Status

Progress This Week:

• [Completed item 1]
• [Completed item 2]

Blockers/Risks:

• [Risk 1 + mitigation plan]

Next Week:

• [Planned item 1]
• [Planned item 2]

On Track? [Yes / No / At Risk]

---

Final POC Report Template

Executive Summary:

• Objective: [What we set out to prove]
• Outcome: [What we learned/achieved]
• Recommendation: [GO / NO-GO / CONDITIONAL GO]

Technical Results:

• Accuracy: [X%] (Target: [Y%])
• Performance: [X seconds] (Target: [Y seconds])
• [Other KPIs]

Business Value:

• Efficiency gains: [X hours/week saved]
• User satisfaction: [X% CSAT]
• Estimated ROI: [X]

Responsible AI:

• Fairness: [Pass/Fail + details]
• Safety: [Pass/Fail + details]
• Privacy: [Pass/Fail + details]

Risks & Mitigations:

• [Risk 1 + status]
• [Risk 2 + status]

Next Steps:

• If GO: [Production roadmap, timeline, budget]
• If NO-GO: [Reasons, alternative approaches]

Attachments:

• Test results
• User feedback summary
• Cost analysis

---

Go/No-Go Decision Framework

Decision Criteria Scorecard

Category	Weight	Pass Threshold	Actual Score	Weighted	Pass?
Technical Feasibility	30%	>75%	[X%]	[X]	[Y/N]
Business Value	30%	>70%	[X%]	[X]	[Y/N]
Responsible AI	20%	>80%	[X%]	[X]	[Y/N]
User Acceptance	10%	>70%	[X%]	[X]	[Y/N]
Risk Management	10%	No critical risks	[X/5 risks mitigated]	[X]	[Y/N]
TOTAL	100%	>75%		[X%]	[Y/N]

---

Decision Paths

GO Decision

Criteria:

• Overall score >75%
• All critical dimensions pass threshold
• No unmitigated high risks (score >15)
• Stakeholder approval obtained

Next Steps:

1. Finalize production architecture
2. Secure production budget
3. Define production roadmap (6-12 months)
4. Establish MLOps/GenAIOps processes
5. Plan change management/training

Timeline to Production: [X weeks/months]

---

CONDITIONAL GO Decision

Criteria:

• Overall score 60-75%
• Some dimensions below threshold
• High risks present but mitigatable

Conditions to Meet:

• [Condition 1: e.g., Improve accuracy to 85% before production]
• [Condition 2: e.g., Complete security audit]
• [Condition 3: e.g., Obtain legal approval for data usage]

Re-evaluation Date: [Date]

---

NO-GO Decision

Criteria:

• Overall score <60%
• Critical dimension failures
• Unmitigatable high risks
• Stakeholder concerns unresolved

Reasons:

• [Reason 1]
• [Reason 2]

Alternatives:

1. Pivot: Change approach (different platform, simpler use case)
2. Delay: Address blockers, re-run POC in [X months]
3. Cancel: Not viable, explore non-AI solutions

---

Example POC Plan

POC: Customer Support Chatbot (Copilot Studio)

Executive Summary:

• Hensikt: Automatisere repetitive kundehenvendninger (ordrestatus, returer, leveringsspørsmål) via chatbot i Teams og på web.
• Varighet: 2 uker
• Ressurser: 3 personer (1 solution architect, 1 developer, 1 SME)
• Beslutningsdato: 2025-02-14

---

Business Case:

Problem: Kundestøtte bruker 40% av tiden (16 timer/uke) på repetitive spørsmål. Gjennomsnittlig responstid er 15 minutter.

Målsetting:

• Automatisere 60% av repetitive henvendelser
• Redusere responstid til <2 minutter
• Frigjøre 10 timer/uke for støtteteamet

Strategic Value:

• Business Impact: 4/5 (betydelig effektivisering)
• User Value: 5/5 (raskere svar for kunder)
• Innovation: 3/5 (standard chatbot-løsning)
• Strategic Alignment: 4/5 (aligner med AI-strategi)

---

Technical Scope:

AI Maturity: Level 2 (har litt erfaring med Power Platform, basic AI-forståelse)

Chosen Solution: Copilot Studio

• Hvorfor: Low-code, rask utvikling, godt integrert med Teams/Dataverse, møter compliance-krav

Data Sources:

1. Dataverse: Ordredata (Order Status, Tracking Numbers)
2. SharePoint: FAQ-dokumenter, return policies
3. Customer Service API: Live order lookup

Infrastructure:

• Copilot Studio capacity: 1000 conversations/month
• Azure AI Search: For FAQ grounding
• Dataverse: For order data
• Content Safety: Azure AI Content Safety filters

---

Success Criteria:

Technical:

• Intent recognition accuracy: >85%
• Response time: <3 seconds
• Availability: >99%
• Content Safety: 100% pass rate

Business:

• Automation rate: >60% of test queries handled without human
• User satisfaction: >70% CSAT
• Cost per conversation: <5 NOK

Responsible AI:

• No bias in responses across customer demographics
• All PII handled securely
• Transparent AI disclosure to users

---

Implementation Plan (2 weeks):

Week 1:

• Day 1-2: Setup Copilot Studio environment, define topics (Order Status, Returns, Shipping)
• Day 3-4: Integrate Dataverse + SharePoint, configure gen AI orchestration
• Day 5: Build initial conversation flows

Week 2:

• Day 1-2: Test with internal users, iterate on prompts
• Day 3-4: User acceptance testing (10 customer service reps), collect feedback
• Day 5: Document results, prepare go/no-go recommendation

---

Team:

• Project Lead: Kari Nordmann (10 timer/uke)
• Solution Architect: Ola Hansen (15 timer/uke)
• Developer (Copilot Studio): Emma Larsen (20 timer/uke)
• SME (Customer Service): Per Johansen (5 timer/uke)

---

Testing Plan:

Functional Tests:

• Test all 3 main topics (Order Status, Returns, Shipping)
• Test escalation to human agent

Performance:

• Load test: 50 concurrent conversations
• Latency: Measure p50, p95, p99

Responsible AI:

• Test 20 adversarial prompts (jailbreak attempts)
• Validate content filters active

UAT:

• 10 customer service reps test for 2 days
• Survey: CSAT, ease of use, accuracy

---

Risks:

Risk	Likelihood	Impact	Score	Mitigation
Low intent recognition accuracy	3	4	12	Add more training phrases, use gen AI fallback
Dataverse integration delays	2	3	6	Start integration early, have mock data ready
User resistance (prefer human support)	2	2	4	Change management, involve users early

---

Budget:

Item	Cost
Copilot Studio license (1 month)	5,000 NOK
Azure AI Search (dev tier)	500 NOK
Personnel (80 hours × 1000 NOK/hr)	80,000 NOK
TOTAL	85,500 NOK

---

Go/No-Go Criteria:

• Intent accuracy >85%
• Response time <3s
• User satisfaction >70%
• No critical safety issues
• Budget for production <50,000 NOK/year

Expected Outcome: GO (90% confidence based on similar implementations)

---

Vedlegg: Nyttige Ressurser

Microsoft Documentation

• AI Adoption Framework (CAF)
• Copilot Studio Implementation Guidance
• Microsoft Foundry Evaluation
• Responsible AI Standard

Tools

• Microsoft Foundry: Model evaluation, deployment
• Copilot Studio: Agent development, testing
• Azure AI Content Safety: Content moderation
• Responsible AI Dashboard: Fairness, bias detection (Azure ML)

Templates

• AI Impact Assessment Template
• Responsible AI Maturity Model

---

Sist oppdatert: 2026-06-24 Versjon: 1.0 Eier: AI Architect Plugin