Overview

5
Service Lines
25
Risks Identified
30+
Lessons Learned
8
Common Themes

Common Themes Across Service Lines

1. Token Cost and Governance

Unpredictable token costs and lack of real-time monitoring pose significant financial risk. Cost governance must be built into workflows from day one.

AssuranceConsultingFabricTaxParthenon

2. Fix Process Before AI

AI amplifies existing dysfunction. Broken processes lead to faster rework loops. Remove non-AI process blockers before introducing AI.

FabricParthenon

3. Skills Gap and Training

Teams lack prompt engineering capability and understanding of AI usage. Upskilling is a prerequisite, not a by-product. Training is urgent.

ConsultingTaxParthenon

4. Use the Right Tool for the Task

Not every task needs heavy AI agents. Use lighter tools first. GitHub Copilot for inline code, MS Copilot for research, Factory.ai for orchestration only.

ConsultingFabric

5. Human-in-the-Loop Essential

AI prepares artefacts; humans approve at gates. Moderate-autonomy stages require explicit human-in-the-loop controls for quality and accountability.

AssuranceParthenon

6. Testing and Validation Critical

Validate AI-native processes through real execution rather than design alone. Testing uncovers new problems requiring iterative refinement.

AssuranceFabric

7. Adoption and Change Management

Low engagement from teams and resistance to new ways of working. Early socialization with delivery teams is critical before introducing AI-enabled models.

FabricParthenon

8. Measure Outcomes, Not Activity

Success measured by value delivered, not volume of features. Static KPIs do not fit rapidly evolving agentic systems. Focus on cycle time, throughput, delivery health.

FabricTax

Risks by Service Line

Assurance Risks
  • Token cost vs. efficiency gain: May be cost prohibitive depending on location/quality expectation
  • Semantic accuracy: Overfitting risk in automated processes
  • Alignment across teams: Process, tools, source of truth, and handoffs need coordination
  • Role clarity: Activity overlap/segregation of duties leading to duplicative work
Consulting Risks
⚠ Critical Governance Notice: Effective governance of AI tool usage is now a delivery-critical control. Risk from insufficient guardrails and limited visibility into token consumption can lead to unplanned cost escalation. Direct threat to delivery continuity.
  • Usage controls: Inadequate role-based (including seniority) controls driving higher tool costs
  • Token monitoring: Lack of real-time monitoring of token consumption (product/eng code/user/consumed tokens)
  • Token limits: Management of token limits at product level including notifications
  • PDLC strategy: Defining the right strategy in each phase of PDLC on how to use the tools
  • Budget coverage: Delivery continuity at risk if budget coverage cannot be confirmed
  • Training: Urgent need for competency-based training and learning plans
Fabric Risks
  • Adoption challenges: Teams not sufficiently socialized on the AI-enabled model
  • Metrics visibility: Dependency on Fabric Portal onboarding for full metrics visibility
  • Over-engineering: Risk of over-using heavy AI agents where lighter tools would be more effective
  • Greenfield only: AI solution built and tested on greenfield product only - Fabric expect to add Developer Workflow (brownfield) to pilot scope
Tax Risks
  • Unpredictable costs: Tool/token costs - cost governance must be built into workflows from day one
  • Upskilling gap: Lack of continuous upskilling of resources - capability building is prerequisite
  • Cross-cutting concerns: Continuously addressing Cost, Security, Performance, Reliability, Quality
  • Agent guardrails: Guardrails (what agents can access, execute, transmit) are mandatory
  • Skills-based composition: Skills-based agent composition required to reduce token cost and improve quality
EY Parthenon Risks
Key Risk Statement: Without strict control gates and governance, AI introduces cost and risk faster than it creates value.
  • AI amplifying broken process: If DoR not enforced, AI creates faster rework loops
  • Token cost / lack of control: No real-time monitoring leads to cost escalation risk
  • Skills gap: Teams lack prompt engineering capability and understanding of AI usage
  • Governance and guardrails: Risk of inconsistent outputs and lack of explainability
  • Adoption risk: Low engagement from teams and resistance to new ways of working

Lessons Learned by Service Line

Assurance Lessons
  • Testing is Critical: Testing of AI-native process redesign is critical - uncovers new problems requiring iterative refinement
  • High-Autonomy Gains: High-autonomy extraction and transformation steps deliver immediate efficiency gains
  • Human-in-the-Loop: Moderate-autonomy stages require explicit human-in-the-loop controls for quality
  • Token Baseline: Need well-established baseline for monthly token estimate by role
  • Contract Considerations: Token cost/coverage agreement with suppliers needs consideration in contracts/SOWs
  • Example - UX Concept: Initial flow had no standard way to share working prototype. Updated flow: PM builds prototype → imported into Figma → XD Lead reviews/updates → Tech Lead validates → sandbox → business testing
Consulting Lessons
  • Use the Right Tool:
    • GitHub Copilot: Best for inline code suggestions
    • MS Copilot: Best for general research, summarization
    • Factory AI: Reserve for full workflow orchestration only
    • Not every task needs Droid - use lighter tools first
  • Choose the Right Model: Use /model in Droid to select based on task complexity:
    • Simple tasks: 0.2-0.4 tokens
    • Medium tasks: up to 0.8 tokens
    • Complex tasks: 1-2 tokens
  • Optimize Prompts:
    • State your goal clearly in the first prompt
    • Mention only the relevant function/class, not the whole file
    • Use #file mentions instead of pasting code
    • Start a new chat for each new task
    • Provide a schema or 2-3 row sample instead of full data dumps
Fabric Lessons
  • Early Socialization: Importance of early socialization with delivery teams before introducing AI-enabled model
  • Remove Blockers First: Remove non-AI process blockers before introducing AI - existing inefficiencies compound with automation
  • Real Execution Testing: Validate AI-native processes through real execution rather than design alone
  • Selective AI Usage: Selectively use AI agents to balance efficiency and cost - not every task needs heavy AI
  • Measure Outcomes: Success measured by value delivered, not volume of features - focus on cycle time, throughput, delivery health
Tax - 7 Key Lessons from Way We Work
  • 1. Tools Alone Don't Change Outcomes - Workflows Do: Adoption stalls when teams are left to "figure it out themselves"
  • 2. Agentic Delivery Requires Explicit Role Design: Agents must be treated as first-class delivery participants with defined scopes
  • 3. Cost and Token Economics Become Material Very Quickly: Always-on heavy agent contexts create exponential cost growth
  • 4. Workflow Coverage is More Effective Than Full SDLC Modeling: Start with high-impact workflows, not comprehensive coverage
  • 5. Upskilling is a Prerequisite, Not a By-Product: Capability building must run in parallel with workflow rollout
  • 6. Measurement Must Shift From Activity to Outcomes: Static KPIs do not fit rapidly evolving agentic systems
  • 7. "Way We Work" Must Be Versioned and Continuously Evolving: Positioned as living initiative that evolves with models, tooling, economics
EY Parthenon - 6 Key Lessons
  • 1. Fix Process Before AI: AI amplifies existing dysfunction - broken processes lead to faster rework loops
  • 2. Upstream is the Constraint: Poor intake → rework loops → delivery inefficiency
  • 3. Definition of Ready is Critical: Must be explicit, measurable, and enforced as a hard gate
  • 4. Human-in-the-Loop is Essential: AI prepares artefacts; humans approve at gates
  • 5. Tooling Alone is Not the Solution: Operating model + behaviour change required for success
  • 6. Training is Urgent: Teams need prompt engineering skills and understanding of AI capabilities

Consolidated Risk Matrix

Cross-Service Line Risk Summary

Risk Category Severity Description Affected Service Lines
Token Cost Escalation High Lack of real-time monitoring and role-based controls leading to unpredictable and potentially unsustainable costs All Service Lines
AI Amplifying Broken Process High Without DoR enforcement and process fixes, AI accelerates rework loops rather than improving efficiency Parthenon, Fabric
Skills Gap High Teams lack prompt engineering capability and understanding of AI usage; training is urgent prerequisite Consulting, Tax, Parthenon
Governance and Guardrails High Risk of inconsistent outputs, lack of explainability, and insufficient controls on what agents can access/execute Consulting, Tax, Parthenon
Adoption Resistance Medium Low engagement from teams and resistance to new ways of working; requires early socialization Fabric, Parthenon
Semantic Accuracy Medium Overfitting risk and context misses in automated processes; human validation required Assurance
Budget Coverage High Delivery continuity at risk if budget coverage cannot be confirmed; may require pausing usage or reducing headcount Consulting
Greenfield vs Brownfield Medium AI solutions built and tested primarily on greenfield products; brownfield applicability unproven Fabric