Template

AI Opportunity Scan Template

A practical, structured template to discover, document, prototype, prioritize, and hand off AI use cases so teams move from idea to measurable pilot with clear success criteria, data readiness checks, and risk controls.

Purpose

This template helps teams discover realistic AI opportunities, document the idea clearly, assess feasibility and risks, design a fast low-risk prototype, and prioritize initiatives with measurable KPIs. Use it as a working document during discovery workshops, spike sprints, or pre-pilot reviews.

How to use

  1. Fill the core fields with concise answers (10–20 minutes per use case).
  2. Run the data readiness and risk checklists with domain and data owners (30–60 minutes).
  3. Create a one-page prototype plan with success criteria and acceptance tests.
  4. Score and prioritize opportunities using the quick rubric below.
  5. Use the handoff checklist to transition validated pilots to engineering, data, and governance teams.

Core template fields (copyable)

1. Use case summary

Short description (1–3 sentences):

Example: "Reduce manual invoice reconciliation time by extracting vendor, amount, and PO from invoices and automatically matching to ledger entries."

2. User, customer, and journey touchpoints

  • Primary user(s): (role/title)
  • Where in the workflow does the opportunity appear? (touchpoints, system screens, triggers)
  • Pain, friction, or missed opportunity today (quantify if possible)

3. Value hypothesis & measurable KPIs

Value hypothesis: If we apply AI to [activity], then [benefit].

Suggested KPIs (choose 1–3 primary):

  • Time saved per transaction / employee hours saved
  • Conversion / completion / throughput uplift (%)
  • Error / defect reduction (%) or false positive rate
  • Cost per transaction avoided ($)
  • User satisfaction / NPS change

4. Success criteria for a prototype

  • Minimum acceptable KPI improvement (e.g., reduce time by 30%)
  • Minimum accuracy / precision / recall thresholds (if applicable)
  • Operational constraints (latency, uptime)
  • Acceptable false positives/negatives and mitigation

5. Feasibility assessment

Model approach (select one or combine):

  • Rule-based / deterministic
  • Classical ML (tabular models)
  • Pre-trained model + fine-tuning (e.g., LLM fine-tune)
  • Retrieval / embeddings + similarity search
  • Computer vision model or multimodal

Estimate:

  • Data engineering effort: Low / Medium / High
  • Modeling effort: Low / Medium / High
  • Integration effort (APIs, UI, automation): Low / Medium / High
  • Rough compute / infra needs: (eg. CPU-only, GPU, cloud inference costs)
  • Estimated calendar time to prototype (weeks)

6. Data readiness checklist & privacy flags

  • Availability: Required data exists? (Yes / Partial / No)
  • Accessibility: Can we access and export it? (Yes / No)
  • Volume: Sufficient quantity to train/validate? (Yes / No)
  • Labeling: Labeled data available? (Yes / Partial / Plan to label)
  • Data quality issues: duplicates, missingness, inconsistent formats
  • PII / sensitive data present? (Yes / No) — if yes, list types and retention rules
  • Legal / contractual constraints on data use

7. Risk checklist

  • Safety-related risks (harmful outputs, wrong decisions)
  • Bias or fairness concerns affecting groups of users
  • Explainability needs for audits or regulators
  • Regulatory compliance (industry-specific)
  • Security risks (data exfiltration, model abuse)
  • Customer trust & transparency obligations
  • Mitigation actions (e.g., human-in-loop, monitoring, reject rules)

8. Stakeholders & owner

  • Product / business owner
  • Data owner(s)
  • Engineering owner
  • Legal / compliance reviewer
  • UX / design contact

Quick prototype plan (one page)

Use this as the minimal plan to validate the value hypothesis quickly.

  • Objective: Clear, measurable objective for the spike.
  • Scope: Limit inputs, outputs, and user population (e.g., "10% of transactions, internal users only").
  • Data: Exact datasets to use, sample size, labeling plan.
  • Approach: Technical approach (baseline model or integration pattern).
  • Deliverables: What will be delivered (notebook, API, dashboard, evaluation report).
  • Success criteria: Quantitative thresholds + qualitative acceptance tests.
  • Timeline: Week-by-week milestones (e.g., Week 1 data prep, Week 2 model, Week 3 evaluation).
  • Minimum viable safety controls: human review, embargo, or sandbox environment.

Prioritization / scoring rubic

Score each use case 1–5 (1 low, 5 high). Example weighted formula: Priority = 0.5*Impact + 0.3*Feasibility + 0.2*RiskAdjustment (inverse of risk).

  • Impact: expected business value on primary KPI
  • Feasibility: data readiness + technical complexity
  • RiskAdjustment: lower score for higher unresolved risks

Produce a short recommendation: Proceed to prototype, Defer until data ready, or Reject / rethink.

Example KPIs & evaluation metrics

  • Precision / recall / F1 (classification)
  • Mean absolute error / RMSE (regression)
  • Time-per-task reduction (seconds or minutes)
  • Throughput increase (units/hour)
  • Cost saved per period ($/month)
  • User acceptance rate or override rate (for human-in-loop)

Handoff checklist (from prototype to production)

  • Data pipeline: documented, repeatable, and tested
  • Model packaging: versioned model artifact and reproducible training script
  • Evaluation report: test results, benchmarks, failure modes
  • Monitoring plan: metrics to monitor, alert thresholds, logging
  • Rollback plan & safety controls
  • Legal & privacy sign-off completed
  • Operational owner and SLA definition

Templates to copy into your workspace

Copy the following one-line headers into your team document or backlog card:

  • Title
  • Summary
  • Primary KPI & baseline
  • Prototype success criteria
  • Data readiness status
  • Feasibility notes
  • Risk & mitigation
  • Recommended next step

Tips & common pitfalls

  • Start with the smallest scoped experiment that can falsify your hypothesis.
  • Validate data access early; many promising ideas stall on data availability.
  • Avoid chasing marginal accuracy gains when business process change would yield larger value.
  • Include compliance and privacy checks in discovery — retrofit is costly.
  • Use stacked baselines: simple rules → ML → pre-trained models to compare cost/benefit.

Related resources

  • Opportunity prioritization worksheet
  • Data readiness checklist (detailed)
  • AI pilot monitoring playbook
Discuss Ask the Librarian

Discussion

Comments and conversation will live here.