advanced·Research·1 hour

Second-Order Effects Analysis for AI Automation

A structured analysis of what will break at scale when a proposed AI automation goes live — covering review bottlenecks, testing load, version control, API surface, cost curves, and intellectual control.

⬡

Recommended modelClaude Sonnet 4.6

Deep literature synthesis with citation precision

What you need to fill in

[describe the AI automation you're planning (e.g., AI writes first-draft code, AI generates monthly reports)][expected volume increase (e.g., from 10 reports/month to 100)][your current review and QA process][your team size and who currently reviews output]

Tools used

ClaudeChatGPT

The prompt

You are a systems-thinking consultant applying the 'Second-Order Effects' framework from Adam Bender's Google I/O 2026 talk on Software Ecology. The automation I'm considering: [describe the AI automation you're planning] Expected volume change: [expected volume increase] Current review process: [your current review and QA process] Team: [your team size and who currently reviews output] Analyse the following second-order effects at 10× and 100× volume: 1. **Review bottleneck** — Can humans still meaningfully review this volume? At what point does review become a rubber stamp? 2. **QA scaling** — How does testing/validation scale — linearly or quadratically? 3. **Cost / Jevons Paradox** — At cheaper unit costs, what new uses proliferate? What does the actual cost curve look like? 4. **API surface** — What internal systems will AI agents now access that weren't designed for agent use? 5. **Agentic conflicts** — Where could two AI agents produce conflicting outputs or edit wars? 6. **Intellectual control** — At what volume do humans lose the ability to reason about what the system is doing? For each: describe the failure mode, estimate the volume at which it triggers, and give one concrete mitigation.

Open in Claude

Sample output

**Second-Order Effects Analysis: AI-Generated Monthly Reports** **1. Review bottleneck** At 10× volume (100 reports/month), 1 senior analyst is at capacity. At 100×, sign-off becomes a rubber stamp — unread. → Mitigation: Triage by materiality; auto-approve reports below a $50K variance threshold. **2. QA scaling** Tests grow quadratically — 10× more reports × 10× more edge cases = 100× test cases needed. → Mitigation: Parameterised test suites against a golden dataset; not manual review. **3. Jevons Paradox** At $0.001/report (vs. $50 analyst-hour), departments request 20× more ad-hoc analyses. Monthly AI cost: $500 → $10,000. → Mitigation: Per-department budget caps, not per-request. **4. API surface** Agents discover the ERP read endpoint and query production data for draft reports at 3am without audit trail. → Mitigation: Read-only agent service accounts with rate limiting and full audit logging. **5. Agentic conflicts** FX-adjusted and non-adjusted versions of the same report generated simultaneously; executive receives both, neither labelled canonical. → Mitigation: Single-output rule per report type; version locking before distribution. **6. Intellectual control** At 500 auto-generated reports/month, no human has read them all. Systematic prompt bias goes undetected for quarters. → Mitigation: Monthly red-team sample — 5 random reports read in full by a senior analyst.

Try with these tools

perplexity claude elicit

ai-strategysecond-ordersystems-thinkinggovernance

More in Research

AI Industry Weekly Briefing

A 400-word executive briefing with 5 structured sections ready to share with a leadership team.

intermediate

AI Company Valuation Analysis

A structured valuation memo with revenue multiples, bull/bear assumptions, and a clear verdict.

intermediate

AI Adoption Gap Analysis for My Business

A personalised AI adoption roadmap with 5 ranked use cases, tools, timelines, and one same-day quick win.

beginner