Appendix B: Templates and Worksheets

This appendix provides fifteen ready-to-use templates and worksheets designed for immediate application in AI and machine learning initiatives. Each template has been refined through real-world deployment across industries ranging from financial services to healthcare to retail. You can copy, adapt, and extend every template to fit your organizational context.

How to use this appendix. Each template follows a consistent structure: a brief description of its purpose, instructions for effective use, the template itself in a format you can reproduce in any document or spreadsheet tool, and a short example showing how a completed version might look. Templates are sequenced to mirror the lifecycle of an AI initiative --- from strategic framing and project proposal through execution, governance, and incident response.

A note on customization. No template is one-size-fits-all. Where you see bracketed placeholders like [Company Name] or [Date], replace them with your own details. Where you see scoring scales, adjust the weights and criteria to reflect your organization's priorities. The goal is not rigid adherence to a form but disciplined thinking about the right questions.

Template 1: AI Project Proposal

Purpose

This one-page template provides a structured format for pitching an AI initiative to leadership. It forces proposers to articulate the business problem, data requirements, approach, resource needs, and risk profile in a concise format that decision-makers can evaluate quickly. Use it at the ideation stage to secure funding and sponsorship for exploratory or production AI projects.

Instructions for Use

1. Complete every section. If you cannot fill in a section, that is a signal the project needs more definition before it is ready for proposal.
2. Keep the entire proposal to a single page (or two pages maximum for complex initiatives). Brevity forces clarity.
3. Attach supplementary detail (data dictionaries, technical architecture diagrams, vendor quotes) as appendices rather than embedding them in the proposal body.
4. Circulate the draft to at least one technical stakeholder and one business stakeholder before submitting. Proposals that reflect only one perspective rarely survive scrutiny.
5. Revisit and update the proposal quarterly if the project is approved. It should remain a living reference point.

Template

AI PROJECT PROPOSAL

1. Problem Statement

Describe the business problem or opportunity in 2--3 sentences. What is the current state? What pain does it cause? Who is affected?

[Write here]

2. Proposed AI/ML Approach

What type of AI/ML solution do you envision? (e.g., classification model, NLP pipeline, recommendation engine, generative AI application, computer vision system). Why is AI the right approach rather than a rules-based or manual solution?

[Write here]

3. Data Requirements

4. Team and Resources

5. Timeline

6. Budget Estimate

7. Success Metrics

8. Risks and Mitigations

Example (Completed)

Project Title: Customer Churn Prediction for Enterprise Accounts Proposer: Maria Chen, Director of Customer Success Problem Statement: Enterprise accounts represent 68% of annual recurring revenue but churn has increased from 8% to 14% over the past 18 months. The customer success team currently relies on gut instinct and lagging indicators (support tickets, renewal date proximity) to identify at-risk accounts. By the time warning signs are visible, it is often too late to intervene effectively. Proposed Approach: A gradient-boosted classification model trained on 4 years of customer behavior data to predict churn probability 90 days in advance, enabling proactive intervention by account managers. Budget: $285K one-time, $95K annual recurring. Target: Reduce enterprise churn from 14% to 9% within 12 months of deployment, yielding an estimated $4.2M in retained ARR.

Template 2: AI Use Case Prioritization Matrix

Purpose

When an organization identifies multiple potential AI use cases --- and most do --- this scoring framework helps leadership compare them systematically rather than defaulting to the loudest advocate or the shiniest technology. The matrix evaluates each use case across five weighted criteria and produces a composite score that enables rank-ordering and portfolio-level decision-making.

Instructions for Use

1. Assemble a cross-functional scoring team that includes business, technical, and risk perspectives. Individual scoring introduces bias; group calibration produces more reliable rankings.
2. Before scoring, align the team on definitions. What constitutes "high" business impact in your organization? What does "data readiness" mean specifically? Use the scoring guide below as a starting point and customize it.
3. Adjust the default weights to reflect your organization's strategic priorities. A company in a heavily regulated industry might increase the weight on Risk Profile; a startup might increase the weight on Business Impact.
4. Score each criterion on a 1--5 scale. Multiply by the weight. Sum for the composite score.
5. Do not treat the composite score as absolute truth. Use it to structure discussion, surface disagreements, and identify clusters of high-priority versus low-priority initiatives.

Scoring Guide

Template

Default Weights: Business Impact (30%), Feasibility (25%), Data Readiness (20%), Strategic Alignment (15%), Risk Profile (10%)

Priority Tiers: - Tier 1 (Score 4.0--5.0): Pursue immediately. Allocate resources and assign executive sponsor. - Tier 2 (Score 3.0--3.9): Develop further. Invest in a proof of concept or data readiness initiative. - Tier 3 (Score 2.0--2.9): Monitor. Revisit when conditions improve (better data, lower risk, stronger alignment). - Tier 4 (Score <2.0): Deprioritize. Document rationale and archive.

Example (Completed)

The scoring team placed customer churn prediction in Tier 1, invoice anomaly detection and autonomous pricing in Tier 2 (with pricing flagged for a data readiness initiative), and resume screening in Tier 3 due to elevated bias and regulatory risk.

Template 3: ML Project Canvas

Purpose

Adapted from the ML Canvas developed by Louis Dorard, this template provides a structured, one-page view of the core components of a machine learning project. It bridges the gap between business requirements and technical specification by forcing practitioners to define what the model predicts, what decisions it informs, what data it requires, how it will be evaluated, and how it will be monitored in production. Complete this canvas before writing any code.

Instructions for Use

1. Start with the top row (Prediction Task, Decisions, Impact). These sections anchor the entire canvas in business value. If you cannot articulate the decision the model supports, stop and clarify with stakeholders before proceeding.
2. Move to the middle row (Data Sources, Features, Offline Evaluation). These sections define the technical core of the project.
3. Finish with the bottom row (Online Evaluation, Monitoring, Maintenance). These sections are the most frequently neglected and the most important for production success.
4. Review the completed canvas with both business sponsors and technical leads. It should be understandable to both audiences.
5. Post the canvas in a visible location (physical or digital) and update it as the project evolves.

Template

ML PROJECT CANVAS

ROW 1: BUSINESS CONTEXT

ROW 2: TECHNICAL CORE

ROW 3: PRODUCTION AND OPERATIONS

CONSTRAINTS AND ASSUMPTIONS

Example (Completed)

Project Name: 90-Day Enterprise Churn Predictor Prediction Task: Binary classification --- will an enterprise customer churn within 90 days? Predicted weekly for each active enterprise account. Decisions Supported: Customer success managers use the churn probability score to prioritize outreach. Accounts with >60% churn probability are assigned to senior CSMs for intervention. VP of Customer Success uses aggregate churn forecasts for quarterly planning. Data Sources: CRM activity logs (Salesforce, 4 years), product usage telemetry (Mixpanel, 2 years), support ticket history (Zendesk, 3 years), billing and contract data (NetSuite, 5 years). Labels derived from historical churn events (non-renewal or mid-contract cancellation). Offline Evaluation: Primary metric: AUC-ROC (target >0.85, baseline logistic regression at 0.72). Secondary: precision at top decile (target >0.60). Evaluated on time-based train/test split to prevent temporal leakage. Monitoring: Weekly drift detection on top 20 features using Population Stability Index (alert if PSI > 0.2). Daily prediction distribution monitoring. Monthly model performance review against realized churn outcomes.

Template 4: Data Readiness Assessment

Purpose

Data readiness is the single most common point of failure for AI projects. This checklist provides a structured assessment of whether the data needed for a specific AI initiative actually exists, is accessible, is of sufficient quality, and is governed appropriately. Complete this assessment during the discovery phase of any AI project, before committing significant resources to model development.

Instructions for Use

1. Complete one assessment per primary data source. If a project requires multiple data sources, assess each independently and then assess the integration challenge.
2. For each item, mark the current status: Green (ready), Yellow (needs work but achievable in <4 weeks), Red (major gap requiring >4 weeks or significant investment), or N/A (not applicable).
3. Aggregate the results into an overall data readiness score. Use the scoring rubric at the bottom.
4. Share the completed assessment with the project sponsor. A project with multiple Red items should not proceed to model development until those gaps are addressed.

Template

DATA READINESS ASSESSMENT

SECTION A: Data Availability

SECTION B: Data Quality

SECTION C: Data Governance and Access

SECTION D: Infrastructure and Integration

OVERALL READINESS SCORE

Example (Completed)

Project: Customer Churn Prediction. Data Source: Salesforce CRM. Highlights: A1 (Green) --- 4 years of account activity data exists. B1 (Yellow) --- the "reason for contact" field is missing in 23% of records from 2022 and earlier; imputation or exclusion strategy needed. C2 (Yellow) --- data contains customer company names and contact names; pseudonymization required before loading into ML environment. D3 (Yellow) --- joining Salesforce data with Mixpanel usage data requires matching on account domain, which has inconsistent formatting (85% match rate currently). Overall: Conditional. Three Yellow items with clear remediation paths. Estimated 3 weeks to Ready.

Template 5: Vendor/Platform Evaluation Scorecard

Purpose

Choosing an AI platform or vendor is a high-stakes, long-lived decision. This scorecard provides a structured evaluation framework that prevents teams from being swayed by impressive demos while overlooking critical factors like security posture, vendor lock-in, and total cost of ownership. Use it when evaluating cloud ML platforms, AutoML tools, AI-as-a-service providers, or specialized AI vendors.

Instructions for Use

1. Customize the criteria and weights before beginning the evaluation. The defaults below reflect a general enterprise perspective; adjust for your context.
2. Assign 2--3 evaluators per vendor. Ensure at least one evaluator is technical (can assess architecture and APIs) and one is business-oriented (can assess pricing models and contract terms).
3. Score each criterion on a 1--5 scale after hands-on testing, reference checks, and documentation review. Avoid scoring based solely on vendor presentations.
4. Calculate the weighted composite score and use it to structure a recommendation, not as a mechanical decision.
5. Document the rationale for each score in the Notes column. Future teams will need to understand why a particular vendor was selected or rejected.

Template

VENDOR/PLATFORM EVALUATION SCORECARD

Scoring Scale: 1 = Unacceptable, 2 = Below Requirements, 3 = Meets Requirements, 4 = Exceeds Requirements, 5 = Best in Class

Additional Considerations (not scored but documented):

• [ ] Contract flexibility (term length, exit clauses, SLA guarantees)
• [ ] Reference customers in our industry
• [ ] Migration support and onboarding resources
• [ ] Environmental sustainability commitments
• [ ] AI ethics and responsible AI features (bias detection, explainability tools)

Recommendation: [Write recommendation with rationale]

Example (Completed)

Evaluation: Cloud ML platform for enterprise deployment. Vendors: AWS SageMaker, Azure ML, Google Vertex AI. SageMaker scored highest on integration (4.5 weighted) due to the company's existing AWS footprint. Vertex AI scored highest on core ML capabilities (4.8 weighted) and GenAI support (5.0). Azure ML scored highest on security and compliance (4.6 weighted) due to built-in HIPAA compliance for the healthcare division. Recommendation: Proceed with Azure ML as the primary platform given the healthcare division's regulatory requirements, with a secondary evaluation of Vertex AI for the marketing team's GenAI use cases.

Template 6: AI Ethics Impact Assessment

Purpose

This framework provides a structured process for identifying, evaluating, and mitigating the ethical risks of an AI system before and during deployment. It is designed to complement --- not replace --- technical testing by examining the broader human, social, and organizational impacts of AI. Complete this assessment during the design phase and revisit it at each major milestone.

Instructions for Use

1. Convene a review panel that includes at least one person outside the project team. External perspectives catch blind spots that insiders miss.
2. Complete the assessment collaboratively. Do not delegate it to a single person or treat it as a checkbox exercise.
3. For each risk identified, assign a severity rating and define a concrete mitigation action with an owner and deadline.
4. The assessment should be a living document. Update it when the system's scope, data, or deployment context changes.
5. Escalate any High-severity risks to the AI governance body (if one exists) or to senior leadership before proceeding.

Template

AI ETHICS IMPACT ASSESSMENT

SECTION 1: Stakeholder Analysis

SECTION 2: Fairness and Bias

SECTION 3: Privacy and Data Rights

SECTION 4: Transparency and Explainability

SECTION 5: Safety and Reliability

SECTION 6: Societal and Environmental Impact

SECTION 7: Risk Mitigation Plan

OVERALL ETHICAL RISK RATING: [Low / Medium / High / Unacceptable]

Approval: [Approved / Approved with Conditions / Not Approved]

Conditions (if applicable): [List conditions that must be met before deployment]

Example (Completed)

System: Customer churn prediction model. Overall Risk Rating: Medium. Key findings: (1) The model uses zip code as a feature, which correlates with race and income (Severity: Medium, Risk 2.5). Mitigation: Test for disparate impact across demographic segments using proxy analysis; remove zip code if disparate impact exceeds the four-fifths rule threshold. Owner: Data Science Lead, Deadline: Pre-deployment. (2) Customer success managers may over-rely on model scores without understanding limitations (Severity: Medium, Risk 4.2). Mitigation: Add confidence intervals to the output, provide training on model limitations, and establish quarterly calibration reviews. Owner: CS Director, Deadline: Launch week.

Template 7: Model Card

Purpose

Based on the framework proposed by Mitchell et al. (2019) at Google, a model card is a standardized document that accompanies a trained machine learning model. It communicates what the model does, how it was built, how it performs, and what its limitations are. Model cards promote transparency, enable informed use, and support responsible deployment. Every model deployed to production should have a model card.

Instructions for Use

1. Create the model card when the model is ready for deployment. Update it whenever the model is retrained, its scope changes, or new limitations are discovered.
2. Write for multiple audiences: data scientists who will maintain the model, product managers who will make decisions based on its outputs, and compliance reviewers who need to assess risk.
3. Be honest about limitations. A model card that claims no limitations is not credible and is therefore useless.
4. Store the model card alongside the model artifact in your model registry or version control system.

Template

MODEL CARD

Model Details

Intended Use

Training Data

Evaluation Data

Performance Metrics

Ethical Considerations

Limitations and Recommendations

Caveats and Additional Notes

[Any additional context, known issues, or important caveats for users]

Example (Completed)

Model Name: churn-predictor-v2.3 (XGBoost). Intended Use: Predict 90-day churn probability for enterprise SaaS accounts to prioritize customer success outreach. Out-of-Scope Uses: This model should NOT be used for: individual employee performance evaluation, automated contract termination, or predictions for SMB accounts (trained only on enterprise data). Key Metrics: AUC-ROC 0.87 overall; 0.91 for accounts >$100K ARR; 0.78 for accounts $25--50K ARR. The model performs significantly worse on smaller enterprise accounts due to sparser behavioral data. Key Limitation: The model was trained on data from a period of macroeconomic stability (2020--2024). Performance during economic downturns is unknown and should be monitored closely if market conditions change.

Template 8: AI ROI Business Case

Purpose

This template provides a structured approach to building the financial case for an AI project. It forces discipline around cost estimation (including commonly overlooked categories like ongoing maintenance and change management), benefit quantification (distinguishing between hard savings and soft benefits), and financial analysis (NPV, payback period, sensitivity). Use it to secure budget approval and to establish the financial baseline against which the project will be evaluated.

Instructions for Use

1. Be conservative in benefit estimates and generous in cost estimates. AI projects almost always cost more and deliver value later than initially projected.
2. Distinguish between benefits you can measure in dollars (revenue increase, cost reduction) and benefits that are real but harder to quantify (improved customer experience, reduced risk, faster decisions). Both matter, but they require different treatment in the business case.
3. Include a sensitivity analysis. What happens to the ROI if the model performs at 75% of the expected level? What if deployment takes 50% longer? Decision-makers need to understand how robust the case is.
4. Plan for a 3-year horizon at minimum. Many AI projects do not break even until year 2.

Template

AI ROI BUSINESS CASE

SECTION 1: Cost Estimate

SECTION 2: Benefit Estimate

SECTION 3: Financial Analysis

SECTION 4: Sensitivity Analysis

SECTION 5: Recommendation

[Write 2--3 paragraph recommendation including the financial case, key risks, and conditions for approval]

Example (Completed)

Project: Customer Churn Prediction. 3-Year Analysis at 10% Discount Rate. Total Investment: $665K (Year 0: $285K setup, Years 1--3: $95K/year operations, 15% contingency). Total Hard Benefits: $8.4M (retained ARR from reduced churn: $2.1M Year 1, $3.0M Year 2, $3.3M Year 3). NPV: $5.8M. Payback: 4.2 months. ROI: 1,163%. Sensitivity: Even in the conservative scenario (benefits 25% lower, costs 25% higher), NPV remains strongly positive at $3.9M with a 6.1-month payback. The case is robust.

Template 9: AI Governance Charter

Purpose

An AI governance charter establishes the mandate, membership, authority, and operating procedures of the body responsible for overseeing AI within an organization. Without a formal charter, governance defaults to ad hoc decision-making, inconsistent standards, and diffused accountability. This template provides a comprehensive framework that can be adapted to organizations of any size, from a three-person review committee at a startup to a multi-tiered governance structure at a Fortune 500 company.

Instructions for Use

1. Draft the charter with input from legal, compliance, technology, business, HR, and (if applicable) external advisors. Governance that is designed by one function rarely earns the trust of others.
2. Secure explicit executive sponsorship. A governance body without executive backing will be ignored when it makes inconvenient decisions.
3. Start with a scope and authority level that is appropriate for your organization's AI maturity. Overly ambitious charters that claim authority over every algorithm often collapse under their own weight.
4. Review and update the charter annually. AI capabilities, regulations, and organizational needs evolve rapidly.

Template

AI GOVERNANCE CHARTER

1. Purpose and Scope

2. Governance Body Structure

3. Membership

4. Authority and Responsibilities

The governance body shall have authority to:

• [ ] Approve or reject deployment of AI systems classified as high-risk
• [ ] Require modifications to AI systems that do not meet ethical, safety, or performance standards
• [ ] Commission audits and reviews of deployed AI systems
• [ ] Establish and enforce AI policies, standards, and guidelines
• [ ] Escalate concerns to the executive team or board
• [ ] Mandate incident response actions for AI-related failures
• [ ] [Additional authority items]

The governance body is responsible for:

• [ ] Maintaining the organization's AI risk classification framework
• [ ] Reviewing AI ethics impact assessments for high-risk systems
• [ ] Overseeing the AI project portfolio and prioritization
• [ ] Monitoring regulatory developments and advising on compliance
• [ ] Publishing an annual AI governance report
• [ ] [Additional responsibilities]

5. Operating Procedures

6. AI Risk Classification Framework

7. Review and Amendment

This charter shall be reviewed annually by the governance body. Amendments require approval by [specify --- e.g., majority vote of the governance body and executive sponsor]. Material changes to scope or authority require [specify --- e.g., board approval].

Example (Completed)

Name: NovaCorp AI Governance Committee. Scope: All ML models and GenAI applications deployed in production or customer-facing environments. Chair: Chief Data Officer. Meeting Cadence: Monthly (first Thursday), with emergency sessions callable by any member with 48-hour notice. Decision Process: Consensus preferred. When consensus cannot be reached, the committee votes; the Chair votes only to break ties. All decisions and dissenting views are recorded in minutes. Risk Classification in Practice: The committee classified the customer churn model as "Medium" risk (business impact, no personal data in features after pseudonymization, human-in-the-loop) and the loan pre-approval model as "Critical" risk (financial impact on consumers, uses demographic-adjacent data, regulatory scrutiny under fair lending laws).

Template 10: Change Management Plan

Purpose

AI projects fail more often from organizational resistance than from technical shortcomings. This template uses the ADKAR framework (Awareness, Desire, Knowledge, Ability, Reinforcement) to structure a change management plan specifically tailored for AI initiatives. It addresses the unique challenges of AI adoption: fear of job displacement, distrust of algorithmic decisions, lack of AI literacy, and the need for new workflows.

Instructions for Use

1. Begin this plan during the project design phase, not after deployment. Change management that starts at go-live is change management that starts too late.
2. Identify stakeholders at every level: executives who sponsor, managers who implement, and frontline employees who use (or are affected by) the AI system.
3. Tailor communication and training to each stakeholder group. The CTO and the call center agent need different messages.
4. Plan for resistance. It is not a sign of failure; it is a predictable human response to uncertainty. Name it, understand its sources, and address it directly.
5. Build in reinforcement mechanisms that sustain adoption beyond the initial launch excitement.

Template

AI CHANGE MANAGEMENT PLAN

SECTION 1: Change Impact Assessment

SECTION 2: Stakeholder Analysis

SECTION 3: ADKAR Plan

A --- Awareness (Why is this change happening?)

D --- Desire (Building willingness to participate)

K --- Knowledge (How to work in the new way)

A --- Ability (Enabling people to perform in the new way)

R --- Reinforcement (Sustaining the change)

SECTION 4: Resistance Management

Example (Completed)

Project: Customer Churn Prediction deployment for the Customer Success team (42 CSMs, 6 managers, 2 directors). Key Change Impact: CSMs shift from reactive (waiting for renewal date or support escalation) to proactive (acting on weekly churn scores). This requires new daily workflow, new dashboard, and a different skill set (interpreting probabilistic scores rather than binary signals). Biggest ADKAR Gap: Knowledge. CSMs are skilled relationship managers but have no experience interpreting ML model outputs. A 4-hour workshop on "Working with AI Predictions" was designed, including hands-on practice interpreting scores, understanding confidence levels, and knowing when to override the model. Primary Resistance: Two senior CSMs with 10+ years of experience expressed concern that "a computer can't understand my customers better than I do." Response: Invited them to participate in model validation, showed them cases where the model caught early warning signals they had missed, and positioned the tool as amplifying --- not replacing --- their intuition.

Template 11: AI Team Hiring Plan

Purpose

Building an effective AI team requires more than posting job descriptions for data scientists. This template provides a structured approach to defining roles, identifying skill requirements, planning team structure, and designing an interview process that evaluates both technical competence and the collaborative, communication, and ethical reasoning skills that distinguish effective AI practitioners from technically brilliant but organizationally ineffective ones.

Instructions for Use

1. Start with the business problems the team will solve, not with the roles you think you need. The required team composition follows from the use case portfolio.
2. Distinguish between skills you must hire for (hard to develop internally) and skills you can develop through training (easier to build).
3. Design the interview process to assess what you actually need. A take-home modeling challenge reveals technical skill; a case study presentation reveals communication skill; a scenario discussion reveals ethical reasoning.
4. Plan for the team you need in 12--18 months, not just today. AI teams tend to grow, and hiring takes longer than expected.

Template

AI TEAM HIRING PLAN

SECTION 1: Team Structure

SECTION 2: Role Definitions

Complete one block per role to be filled.

Role: [Title]

SECTION 3: Interview Process

SECTION 4: Evaluation Scorecard

Hiring Decision Thresholds: - Strong Hire: Weighted score >= 4.0, no competency below 3 - Hire: Weighted score >= 3.5, no competency below 2 - No Hire: Weighted score < 3.5 or any competency at 1

SECTION 5: Hiring Timeline

Example (Completed)

Team: Enterprise AI Team at a mid-market SaaS company. Current: 1 data scientist, 1 data engineer. Target (12 months): 3 data scientists, 2 ML engineers, 2 data engineers, 1 AI product manager. First Hire Priority: Senior Data Scientist (to serve as technical lead while the team scales). Required: 5+ years ML experience, production deployment experience, strong communication skills. Nice-to-have: SaaS domain experience, experience mentoring junior team members. Interview Design Note: The ethical scenario stage asks candidates to discuss how they would handle a situation where a deployed model was discovered to have disparate impact across customer segments. There is no single right answer; the stage assesses whether the candidate identifies the stakeholders, considers multiple perspectives, and proposes a thoughtful course of action.

Template 12: AI Strategy One-Pager

Purpose

An AI strategy one-pager distills the organization's AI vision, priorities, and plan into a format that an executive can read in five minutes. It is not a substitute for a full AI strategy document but rather a communication tool that ensures alignment across leadership, provides a reference point for investment decisions, and enables every team to understand how their work connects to the broader AI agenda. Update it quarterly or when strategic priorities shift.

Instructions for Use

1. Develop the one-pager after completing the full strategic planning process, not before. It is a summary, not a substitute.
2. Limit it to one page (two at most). Every sentence must earn its place.
3. Ensure each strategic pillar has at least one measurable outcome. Strategy without measurement is aspiration.
4. Circulate for review to the CEO, CTO, CFO, and business unit leaders before finalizing. Misalignment at the top cascades into confusion at every level.
5. Post it where people will see it. A strategy document that lives in a SharePoint folder no one opens is not a strategy.

Template

AI STRATEGY ONE-PAGER

VISION

[1--2 sentences describing the aspirational future state of AI in the organization. What will be true when this strategy is realized?]

CURRENT STATE

[2--3 sentences describing where the organization is today on its AI journey. Be honest. Reference the AI maturity assessment if completed.]

STRATEGIC PILLARS

INVESTMENT SUMMARY

EXPECTED OUTCOMES (3-YEAR)

• [Outcome 1: e.g., "$15M in AI-driven revenue or cost savings"]
• [Outcome 2: e.g., "10 AI models in production across 4 business units"]
• [Outcome 3: e.g., "AI maturity level advanced from Level 2 to Level 4"]
• [Outcome 4: e.g., "AI governance framework fully operational"]

GOVERNANCE

[1--2 sentences describing who oversees AI strategy execution and how progress is tracked.]

KEY RISKS

Example (Completed)

Vision: By 2028, AI will be embedded in every major decision process at NovaCorp, enabling faster, more accurate, and more equitable outcomes for customers and employees. Current State: NovaCorp is at AI maturity Level 2 (Exploratory). We have 3 ML models in production, a 6-person data science team, and no formal AI governance structure. AI efforts are siloed within the technology organization. Pillars: (1) AI-Powered Customer Experience --- deploy churn prediction, next-best-action recommendation, and intelligent routing (Owner: VP Customer Success). (2) Operational Intelligence --- automate invoice processing, demand forecasting, and quality inspection (Owner: COO). (3) AI Foundation --- build ML platform, data infrastructure, and feature store (Owner: CTO). (4) Responsible AI --- establish governance committee, ethics review process, and AI literacy program (Owner: Chief Legal Officer). 3-Year Investment: $8.2M ($2.1M Year 1, $3.0M Year 2, $3.1M Year 3). Expected Outcomes: $15M in value creation, 10 models in production, maturity Level 4.

Template 13: Prompt Engineering Brief

Purpose

As generative AI becomes a core business tool, organizations need a systematic way to document, share, iterate on, and govern the prompts that drive AI outputs. This template provides a structured format for capturing a prompt's objective, design rationale, expected output, evaluation criteria, and version history. It transforms prompt engineering from an individual craft into an organizational capability. Use it for any prompt that will be used repeatedly, shared across teams, or embedded in a production application.

Instructions for Use

1. Complete a prompt engineering brief for every prompt that is used in production, shared with more than two people, or expected to be maintained over time. Ad hoc experimental prompts do not need this level of documentation.
2. Include the full prompt text, including system instructions, few-shot examples, and any formatting specifications. A brief without the actual prompt is incomplete.
3. Define evaluation criteria before deploying the prompt. How will you know if it is working well? What does a good output look like? What does a bad output look like?
4. Maintain version history. Prompts evolve as models change, requirements shift, and edge cases emerge. Without version tracking, regression is invisible.
5. Include at least one example of expected output and one example of unacceptable output.

Template

PROMPT ENGINEERING BRIEF

SECTION 1: Objective and Context

SECTION 2: Prompt Design

System Instructions / System Prompt:

[Full system prompt text here]

User Prompt Template:

[Full user prompt template with {{variables}} marked]

Variables:

Few-Shot Examples (if applicable):

Configuration Parameters:

SECTION 3: Evaluation

Example of Acceptable Output:

[Paste a representative good output here]

Example of Unacceptable Output:

[Paste an example of a bad output with annotation explaining why it fails]

SECTION 4: Guardrails and Constraints

• [ ] Input validation: [How inputs are validated before reaching the prompt]
• [ ] Output filtering: [Any post-processing or filtering applied to outputs]
• [ ] Content moderation: [Moderation layer, if applicable]
• [ ] PII handling: [How PII in inputs/outputs is managed]
• [ ] Rate limiting: [Usage limits, if applicable]
• [ ] Fallback behavior: [What happens when the prompt fails or output is filtered]

SECTION 5: Version History

Example (Completed)

Prompt Name: Customer Email Response Generator. Model: Claude Sonnet 4. Use Case: Customer support agents use this prompt to draft personalized responses to customer inquiries. The agent reviews, edits, and sends the draft. System Prompt (excerpt): "You are a customer support assistant for NovaCorp, a B2B SaaS company. Draft a professional, empathetic response to the customer email below. Use the customer's name. Acknowledge their concern before providing the solution. Keep the response under 200 words. Do not make commitments about timelines or features without explicit approval." Key Evaluation Criteria: Accuracy (references correct product information), Empathy (acknowledges customer's frustration), Actionability (provides clear next steps), Safety (does not promise features or timelines). Version History: v1.0 used generic tone; v2.0 added empathy requirement after feedback that responses felt robotic; v3.0 added the constraint against committing to timelines after an agent inadvertently promised a feature delivery date; v3.1 adjusted temperature from 0.7 to 0.3 after observing inconsistent tone across responses.

Template 14: AI Incident Response Playbook

Purpose

AI systems fail in ways that traditional software does not: models drift silently, biases emerge unexpectedly, adversarial inputs cause harmful outputs, and hallucinated content can mislead users. This playbook provides a structured template for responding to AI-specific incidents --- from detection through resolution and post-incident review. Every AI system in production should have an incident response plan tailored to its risk profile.

Instructions for Use

1. Customize this playbook for each production AI system. A chatbot and a fraud detection model have very different failure modes and very different response requirements.
2. Identify the incident response team and ensure every member has read and understood the playbook before an incident occurs. Playbooks that are read for the first time during an incident are ineffective.
3. Conduct at least one tabletop exercise per year to practice the response process.
4. After every real incident, update the playbook based on lessons learned. The playbook should improve continuously.
5. Define severity levels and escalation thresholds clearly. In the moment of an incident, ambiguity about "who decides" creates paralysis.

Template

AI INCIDENT RESPONSE PLAYBOOK

SECTION 1: Incident Classification

SECTION 2: Incident Response Team

SECTION 3: Response Procedures

Phase 1: Detection and Triage (Time: 0 to T+30 min)

• [ ] Incident detected via: [monitoring alert / user report / internal discovery / external report]
• [ ] Incident logged in tracking system with timestamp, reporter, and initial description
• [ ] Incident Commander assigned
• [ ] Severity level assessed using classification table above
• [ ] Response team assembled based on severity level
• [ ] Initial impact assessment: How many users/customers/decisions are affected?
• [ ] War room / incident channel created (for SEV-1 and SEV-2)

Phase 2: Containment (Time: T+30 min to T+2 hours for SEV-1)

• [ ] Decision: Can the system be safely paused or rolled back?
• If yes: Initiate rollback to last known good model version or activate fallback system
• If no: Implement interim controls (human review layer, output filtering, scope restriction)
• [ ] Affected outputs identified and flagged (decisions made during incident window)
• [ ] Users/stakeholders notified that an issue is being investigated
• [ ] Data preserved for post-incident analysis (model inputs, outputs, logs, monitoring data)

Phase 3: Investigation (Time: varies by severity)

• [ ] Root cause analysis initiated
• [ ] Data issue? (drift, corruption, pipeline failure, labeling error)
• [ ] Model issue? (overfitting, concept drift, adversarial input, training bug)
• [ ] Infrastructure issue? (compute failure, latency, dependency outage)
• [ ] Integration issue? (API change, upstream system change, feature store error)
• [ ] Human issue? (incorrect configuration, unauthorized change, process failure)
• [ ] Impact quantified: number of affected decisions, financial impact, customer impact
• [ ] Regulatory notification requirements assessed (with Legal/Compliance)

Phase 4: Resolution (Time: varies)

• [ ] Fix developed and tested
• [ ] Fix reviewed and approved by Incident Commander and AI Lead
• [ ] Fix deployed to production
• [ ] System monitored for stability (minimum 24 hours for SEV-1/SEV-2)
• [ ] Affected decisions reviewed and remediated where possible
• [ ] Affected users/customers notified of resolution
• [ ] Incident status updated to "Resolved"

Phase 5: Post-Incident Review (Time: within 5 business days of resolution)

• [ ] Post-incident review meeting scheduled with full response team
• [ ] Incident timeline documented
• [ ] Root cause confirmed and documented
• [ ] Contributing factors identified (not just the proximate cause)

Post-Incident Review Questions:

1. What happened, and when?
2. How was the incident detected? Could it have been detected earlier?
3. What was the impact (users, decisions, revenue, reputation)?
4. What was the root cause?
5. Did the response process work as designed? Where did it break down?
6. What monitoring, testing, or process changes would prevent recurrence?
7. Are there similar risks in other AI systems that should be assessed?

Action Items from Review:

SECTION 4: Communication Templates

Internal Notification (SEV-1/SEV-2):

AI INCIDENT ALERT --- [SEVERITY LEVEL] System: [System name] Detected: [Timestamp] Description: [Brief description of the issue] Impact: [Known or estimated impact] Current Status: [Investigating / Contained / Resolved] Incident Commander: [Name] Next Update: [Time]

External Customer Notification (if required):

We identified an issue with [system/feature] that may have affected [description of impact] between [start time] and [end time]. We have [contained/resolved] the issue and are [taking steps to prevent recurrence / reviewing affected outcomes]. If you believe you were affected, please contact [support channel]. We apologize for the inconvenience and are committed to transparency about this matter.

Example (Completed)

System: Customer-facing product recommendation engine. Incident: On March 3, a monitoring alert flagged that recommendation diversity had dropped by 73% --- the model was recommending the same 12 products to 89% of users. Severity: SEV-2 (significant degradation, no direct harm but material impact on customer experience and revenue). Root Cause: A feature pipeline update on March 2 introduced a schema change that caused the model to receive null values for 8 of 15 user preference features. The model defaulted to popularity-based recommendations, effectively eliminating personalization. Resolution: Rolled back to the previous feature pipeline version. Implemented schema validation checks that would have caught the breaking change. Added a "recommendation diversity" metric to the real-time monitoring dashboard with an alert threshold. Key Lesson: Feature pipeline changes should be treated with the same rigor as model deployments, including staging environment testing and canary rollout.

Template 15: AI Maturity Self-Assessment Questionnaire

Purpose

This self-assessment helps organizations understand where they stand on their AI journey and where they need to invest to advance. It evaluates AI maturity across six dimensions --- Strategy, Data, Technology, Talent, Governance, and Culture --- using 30 questions (5 per dimension). The results identify strengths to leverage, gaps to address, and a roadmap for advancing to the next maturity level. Administer this assessment annually to track progress and recalibrate priorities.

Instructions for Use

1. Assemble a diverse assessment team (5--10 people) that includes leaders from technology, data, business, HR, legal, and operations. Individual assessments produce biased results; collective assessment produces a more accurate picture.
2. Have each team member complete the questionnaire independently, then convene to discuss and calibrate scores. Disagreements are informative --- they often reveal organizational blind spots.
3. Score each question on the 1--5 scale defined in the scoring guide. Be honest. Aspirational scoring defeats the purpose.
4. Calculate dimension scores and the overall maturity score. Use the maturity level definitions to identify where the organization falls.
5. Use the results to inform the AI strategy (Template 12) and investment priorities. Focus improvement efforts on the dimensions that are both low-scoring and strategically important.

Scoring Guide

Template

AI MATURITY SELF-ASSESSMENT

DIMENSION 1: STRATEGY (Weight: 20%)

DIMENSION 2: DATA (Weight: 20%)

DIMENSION 3: TECHNOLOGY (Weight: 15%)

DIMENSION 4: TALENT (Weight: 15%)

DIMENSION 5: GOVERNANCE (Weight: 15%)

DIMENSION 6: CULTURE (Weight: 15%)

RESULTS SUMMARY

MATURITY LEVEL DEFINITIONS

Our Current Maturity Level: [Level and Label]

Target Maturity Level (12 months): [Level and Label]

DIMENSION GAP ANALYSIS AND ACTION PLAN

Example (Completed)

Organization: NovaCorp (mid-market SaaS, 1,200 employees). Results: Strategy 2.8, Data 2.2, Technology 2.6, Talent 2.0, Governance 1.6, Culture 3.2. Overall: 2.4 (Exploratory). Interpretation: NovaCorp's strongest dimension is Culture (3.2) --- teams are enthusiastic about AI and willing to experiment. Its weakest dimensions are Governance (1.6) and Talent (2.0). The organization has energy and ambition but lacks the structural foundations to scale AI responsibly. Priority Actions: (1) Governance: Establish an AI governance committee and adopt a risk classification framework within 6 months (Owner: Chief Legal Officer). (2) Talent: Hire 2 senior data scientists and 1 ML engineer; launch an AI literacy program for business leaders (Owner: VP Engineering + HR). (3) Data: Appoint data stewards for the 5 highest-priority data domains; implement automated data quality monitoring (Owner: CTO). Target: Advance from Level 2 (Exploratory) to Level 3 (Operational) within 12 months.

How to Use These Templates Together

The fifteen templates in this appendix are not isolated documents. They form an interconnected toolkit that supports the full lifecycle of AI in an organization. Here is how they fit together:

Strategic Planning Phase - Start with the AI Maturity Self-Assessment (Template 15) to understand where you are. - Use the AI Strategy One-Pager (Template 12) to define where you are going. - Apply the AI Use Case Prioritization Matrix (Template 2) to decide which initiatives to pursue first.

Project Initiation Phase - Write an AI Project Proposal (Template 1) for each selected initiative. - Complete the Data Readiness Assessment (Template 4) to validate that the data foundation exists. - Build the AI ROI Business Case (Template 8) to secure funding. - Complete the ML Project Canvas (Template 3) to align business and technical teams on the project scope.

Build and Deploy Phase - Evaluate platforms using the Vendor/Platform Evaluation Scorecard (Template 5). - Staff the team using the AI Team Hiring Plan (Template 11). - Conduct the AI Ethics Impact Assessment (Template 6) before deployment. - Create a Model Card (Template 7) for each model entering production. - Document critical prompts using the Prompt Engineering Brief (Template 13).

Operate and Govern Phase - Establish oversight with the AI Governance Charter (Template 9). - Prepare for problems with the AI Incident Response Playbook (Template 14). - Manage organizational adoption with the Change Management Plan (Template 10).

Annual Review - Re-administer the AI Maturity Self-Assessment (Template 15) to measure progress. - Update the AI Strategy One-Pager (Template 12) based on results and evolving priorities.

Taken together, these templates transform AI from a collection of ad hoc experiments into a disciplined, governed, and strategically aligned organizational capability. The templates are the scaffolding; the thinking you put into them is the structure.