AI 扩展计划 / Playbooks
ABPA 模板 07：Data Readiness Pack
Score each item from 1 to 5.
120 行abpa/templates/07-data-readiness-pack.md
Data Readiness Pack
Use this before committing to an AI workflow, RAG system, copilot, or agent. The goal is to prove whether the data can support the business decision.
Decision Context
| Field | Answer |
|---|---|
| Business scenario | |
| AI use case | |
| Primary decision supported | |
| Decision owner | |
| Data owner | |
| Review date |
Data Inventory
| Data asset | System / source | Owner | Format | Volume | Freshness | Access status | Business meaning |
|---|---|---|---|---|---|---|---|
| structured / semi-structured / unstructured | real-time / daily / weekly / ad hoc | approved / pending / blocked | |||||
| structured / semi-structured / unstructured | real-time / daily / weekly / ad hoc | approved / pending / blocked |
Source Of Truth
| Business object | System of record | Conflicting sources | Resolution rule | Owner |
|---|---|---|---|---|
| Customer | ||||
| Case / investigation | ||||
| Transaction / event | ||||
| Policy / procedure | ||||
| Outcome / label |
Data Quality Scorecard
Score each item from 1 to 5.
| Dimension | Question | Score | Evidence | Required fix |
|---|---|---|---|---|
| Completeness | Are required fields populated for enough cases? | |||
| Accuracy | Does the data match trusted records or expert review? | |||
| Consistency | Are definitions stable across teams and systems? | |||
| Timeliness | Is the data fresh enough for the workflow? | |||
| Granularity | Is the detail level sufficient for the decision? | |||
| Representativeness | Does it cover the real case mix, including edge cases? | |||
| Traceability | Can each answer be traced to source evidence? |
Label And Ground Truth Plan
| Label / outcome | Definition | Source | Label owner | Disagreement rule | Minimum sample |
|---|---|---|---|---|---|
Label Risk
| Risk | Example | Impact | Mitigation |
|---|---|---|---|
| Ambiguous policy | model learns inconsistent judgment | write label guide and adjudication process | |
| Historical bias | model repeats old investigation bias | stratified review and fairness checks | |
| Sparse positives | weak recall on rare events | targeted sampling and scenario generation | |
| Stale labels | model optimizes old rules | scheduled refresh and change log |
Knowledge Readiness For RAG
| Knowledge source | Chunking strategy | Metadata required | Update cadence | Retrieval risk | Owner |
|---|---|---|---|---|---|
| Policy | by section / rule / obligation | policy id, version, effective date | |||
| SOP | by task / step / exception | process id, role, jurisdiction | |||
| Case notes | by event / paragraph | case id, timestamp, author role | |||
| FAQ / guidance | by question | category, source, approved by |
Privacy, Security, And Compliance
| Control area | Requirement | Current state | Gap | Owner |
|---|---|---|---|---|
| Data classification | ||||
| PII / sensitive data handling | ||||
| Access control | ||||
| Retention | ||||
| Audit logging | ||||
| Cross-border / jurisdiction |
Data Pipeline Readiness
| Pipeline step | Exists? | SLA | Failure mode | Monitoring | Owner |
|---|---|---|---|---|---|
| Extract | yes / no | ||||
| Transform | yes / no | ||||
| Feature / retrieval index build | yes / no | ||||
| Evaluation dataset refresh | yes / no | ||||
| Production feedback capture | yes / no |
Gaps And Remediation Backlog
| Priority | Gap | Why it matters | Fix | Effort | Owner | Due date |
|---|---|---|---|---|---|---|
| P0 | S / M / L | |||||
| P1 | S / M / L | |||||
| P2 | S / M / L |
Readiness Gate
| Gate | Pass criteria | Status |
|---|---|---|
| Minimum source data available | pass / risk / fail | |
| Ground truth usable | pass / risk / fail | |
| Retrieval corpus versioned | pass / risk / fail | |
| Sensitive data controls approved | pass / risk / fail | |
| Feedback loop designed | pass / risk / fail |
Recommendation
| Field | Answer |
|---|---|
| Data readiness decision | proceed / fix first / stop |
| Biggest blocker | |
| First fix to fund | |
| Next review evidence |