Purpose
Take a described business process or decision domain and surface the specific decisions where AI-driven decision intelligence could replace or augment human judgment — classifying each by decision type, mapping data requirements and readiness, and ranking by expected impact and feasibility. Solves the executive problem of knowing where decision automation actually pays off versus where it adds risk or noise.
When to Use
Invoke this skill when the user asks to:
- Find where AI or decision intelligence could improve a specific business process
- Identify decisions in a function that could be automated or augmented with data
- Build a decision intelligence roadmap or prioritization for a department
- Assess whether a recurring judgment call should stay human, be augmented, or be automated
- Map data requirements for replacing manual decisions in a workflow
- Justify or scope a decision intelligence investment to leadership
Inputs Required
Before starting, collect:
- Process or decision domain: The function, workflow, or decision area in scope (e.g., credit approvals, inventory replenishment, churn intervention)
- Key decisions made today: The recurring decisions within that domain and who makes them
- Decision frequency and volume: How often each decision occurs (daily/weekly/per-transaction) and at what scale
- Current decision basis: How decisions are made now — gut, rules, spreadsheets, partial models
- Available data: What data exists today, where it lives, and its known quality issues
- Business stakes: The cost of a wrong decision and the value of a better one (revenue, risk, cost, speed)
- Constraints: Regulatory, ethical, or human-in-the-loop requirements that bound automation
If decision frequency or data inventory is missing, infer reasonable values from the domain and flag the assumption in the output.
Instructions
Step 1: Decompose the Domain into Discrete Decisions
Break the described process into its component decisions — not tasks, but choices where an outcome is selected. For each, capture:
- The decision statement (e.g., "Which leads get routed to sales vs. nurture")
- The decision owner / role
- Frequency and volume
- Current decision basis (intuition / static rules / analyst judgment / existing model)
Aim for 5–10 discrete decisions. If the domain yields more, group the long tail.
Step 2: Classify Each Decision Type
Classify every decision along two axes:
Structure axis:
- Structured — clear inputs, known options, definable success metric, repeatable logic (e.g., reorder quantity, fraud flag). Strong candidates for automation.
- Semi-structured — patterned but with contextual exceptions requiring judgment (e.g., pricing exceptions, escalation routing). Candidates for augmentation.
- Unstructured — novel, ambiguous, high-context, or values-laden (e.g., entering a new market, M&A go/no-go). Decision support at best; rarely automatable.
Repeatability axis:
- High-frequency / repeatable → automation ROI compounds
- Low-frequency / one-off → augmentation only; automation rarely justified
Step 3: Map Data Requirements and Readiness
For each decision, specify:
- Required signals: The features/inputs a model or system would need
- Data availability: Available now / partially available / must be built
- Data quality risk: Low / Medium / High — with the specific gap (completeness, latency, label availability, bias)
- Outcome label: Whether the historical "right answer" is captured to train or validate against — critical for structured automation
Flag any decision where the outcome is not measurable today; these cannot be automated reliably until instrumented.
Step 4: Determine the Intervention Mode
For each decision, recommend one of three modes with rationale:
- Automate — system decides; human reviews exceptions only. Reserve for structured, high-volume, low-stakes-per-instance, well-labeled decisions.
- Augment — system recommends; human decides. Default for semi-structured and high-stakes decisions.
- Inform — system surfaces analysis; human reasons from scratch. For unstructured, low-frequency strategic decisions.
State the human-in-the-loop requirement and any regulatory or ethical constraint that forces a mode.
Step 5: Estimate Impact and Feasibility
Score each opportunity:
- Impact (1–5): Combine value-per-better-decision × frequency × current error/inefficiency rate
- Feasibility (1–5): Combine data readiness, decision structure, and constraint burden (inverse)
- Expected value statement: One sentence quantifying or directionally sizing the upside (cost saved, risk reduced, speed gained)
Step 6: Rank and Sequence the Portfolio
Produce a prioritized list. Order by a quadrant logic:
- Quick wins (high impact, high feasibility) → do first
- Strategic bets (high impact, lower feasibility) → invest in data foundations
- Incremental (lower impact, high feasibility) → batch when convenient
- Deprioritize (low impact, low feasibility) → explicitly park
Name the top 3 with a one-line reason each, and identify the single most important data or instrumentation gap to close first.
Output Format
- Structure: Opportunity summary at top, then a decision-by-decision table, then ranked portfolio and recommendations
- Length: 900–1,300 words
- Format: Prose summary + a structured table for the decision inventory (Decision | Type | Mode | Data Readiness | Impact | Feasibility) + ranked list
- Sections: Domain summary → Decision inventory table → Classification & data notes → Intervention modes → Ranked opportunity portfolio (quadrants) → Top 3 + first data gap to close
- Tone: Strategy executive — decisive, quantified where possible, honest about feasibility limits; no hype about AI capability
Quality Checklist
Before delivering output, verify:
- [ ] 5–10 discrete decisions are identified (count them) — decisions, not tasks
- [ ] Every decision has a structure classification (structured / semi-structured / unstructured)
- [ ] Every decision has a data readiness rating with a specific named gap, not just "Medium"
- [ ] Every decision has an intervention mode (automate / augment / inform) with rationale
- [ ] Each opportunity has both an impact and feasibility score, and the ranking follows the quadrant logic
- [ ] The single most important data/instrumentation gap to close first is named explicitly
Notes
- The most common error is recommending automation for decisions whose outcomes aren't measured — automation requires a feedback label, so instrumentation often precedes intelligence.
- Unstructured strategic decisions almost never justify automation; positioning them as "inform" prevents overreach and protects credibility with the board.
- High-frequency, low-stakes-per-instance decisions are where decision intelligence compounds fastest — prioritize these even when each individual decision feels minor.
- When constraints (regulatory, ethical) force human-in-the-loop, default to augment and make that explicit rather than burying it as a caveat.
How to use this skill
This file is a Claude Code skill in Anthropic's standard format — markdown with YAML frontmatter at the top.
Installing in Claude Code:
~/.claude/skills/<folder-name>/SKILL.md— auto-loads in every session anywhere on your machine (recommended for skills you'll use across multiple projects)<your-project>/.claude/skills/<folder-name>/SKILL.md— auto-loads only when working in that project (recommended for project-specific skills)
Restart Claude Code or open a new session to register.
Using in Claude.ai, ChatGPT, Gemini, or other AI tools:
- Drag this file into a new chat, or copy and paste its contents
- Send your task or request as your next message — the AI will apply the skill where it fits
For Claude Code skill details: https://docs.anthropic.com/en/docs/claude-code/skills