AI Tooling Data Flow Reviewer

ai-tooling-data-flow-reviewer

The skill that answers "if we wire AI tool X into system Y, where does our regulated data end up?" Specialized for AI integrations because the data flow is non-obvious (inference happens off-org-network) and the controls are different (content exclusion + residency vs traditional network egress).

When to use

• New AI tool integration proposed (Copilot knowledge base, internal LLM gateway, Cursor on regulated repo, MCP server enabling agentic actions).
• Existing AI integration adds a new data source.
• AI tool vendor announces a new data-handling feature (e.g., residency change, new context-loading path).
• Pre-CAB review for any AI-related change.

When NOT to use

• Generic data-flow review without AI angle → secure-design-reviewer.
• AI tool selection / vendor evaluation → vendor-security-eval.
• Specific Copilot policy question → copilot-faq-answerer.

Workflow

1. Identify the integration scope: components, source data, destination.
2. Build the data-flow diagram (ASCII): every hop from data source through AI tool to consumer.
3. Per-hop classification: at each hop, what data classification crosses? Does it cross a trust boundary? What's the control at that hop?
4. Cross-reference vault facts: check Copilot data-residency posture, content exclusion config, audit log path, IP indemnity scope — does the org's existing posture cover this flow's claims?
5. Find the gaps: hops where the org's posture doesn't cover the flow.
6. Produce a structured output — presented in chat.
7. Persist the review — after presenting the report, write the full review note via vault-writer.write_research to vault/research/ai-governance/YYYY-MM-DD-ai-dataflow-{integration-slug}.md (frontmatter per research.yml, topic: ai-governance) so pre-CAB reviews and posture re-checks can query it.

Output structure

# AI Tool Data Flow Review — {integration name}

## Integration summary
{1-paragraph}

## Data-flow diagram
[ASCII showing components + arrows + trust boundaries]

## Per-hop classification table
| Hop | From | To | Data class | Trust boundary crossed? | Control at hop | Gap? |

## Findings
- [per gap: severity, evidence, remediation]

## Cross-reference: vault facts
[which existing facts cover which hops]

## Recommendation
{advance / advance-with-remediation / reject}

Lands at vault/research/ai-governance/YYYY-MM-DD-ai-dataflow-{integration-slug}.md (workflow step 7).

Composes with

• threat-model-helper — natural pair (threat model = what can go wrong; this skill = where the data goes).
• secure-design-reviewer — AI tool exposure category.
• vault-querier — load Copilot facts ([[data-residency-regions]], [[data-residency-surcharge]], etc.).
• vault-writer.write_research — persists the review note (workflow step 7).

Acceptance test (for step 27 done-criteria)

One data-flow review exercise. The Payments KB threat-modeled in step 24 is the natural fixture — produce its data-flow review.