The E0–E3 framework and the realistic time-saved weights (E1 = 0.5, E2 = 0.3, E3 = 0.2) are grounded in four papers, all embedded verbatim (with page citations) into the Bedrock system prompt at runtime:

• Anthropic — Massenkoff, M. & McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Published 5 March 2026. (With companion Appendix.) The primary empirical source. Introduces the "observed exposure" measure that underlies reference/job_exposure.csv. Verbatim methodology from p. 6: "fully automated implementations receive full weight, while augmentative use receives half weight." This is the direct justification for our E1 = 0.5 weight. Also supplies the top-10 exposure table (Figure 3, p. 7) used as empirical anchors in calibration. The companion Appendix provides the full per-occupation SOC-level exposure table used to build job_exposure.csv, plus the methodological detail behind the automation/augmentation split.
• OpenAI — Patwardhan, T. et al. (2025). GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. Capability benchmark covering 1,320 tasks across 44 occupations and the 9 sectors contributing most to U.S. GDP. Headline result (§3.1, p. 5-6): "47.6% of deliverables by Claude Opus 4.1 were graded as better than (wins) or as good as (ties) the human deliverable." Directly supports E1 = 0.5 — parity on roughly half of real knowledge-work tasks. §3.2 on "try n times, if unsatisfactory, fix it yourself" grounds E2 = 0.3 (human-in-the-loop review cycles).
• McKinsey Global Institute (2023). The economic potential of generative AI: The next productivity frontier. Chui, M., Hazan, E., Roberts, R., et al. June 2023. Sets the theoretical ceiling we deliberately discount. Key insight #4, p. 3: "Current generative AI and other technologies have the potential to automate work activities that absorb 60 to 70 percent of employees' time today." Our weights (E1=0.5, E2=0.3, E3=0.2) explicitly sit below this theoretical ceiling to reflect realistic, today-deployable savings. Customer-service case (p. 15) — AI can decrease productivity for highly skilled agents — is the concrete evidence that E1 ≠ 1.0.
• OECD (2022). OECD Framework for the Classification of AI Systems. OECD Digital Economy Papers No. 323, February 2022. Source of the four-level autonomy typology (p. 53) that maps onto E0/E1/E2/E3. Defines human-support, human-in-the-loop, human-on-the-loop, and human-out-of-the-loop autonomy. Page 53 also establishes that high-autonomy AI is policy-gated in critical or rights-affecting contexts — directly justifies E3 = 0.2 (deployment-limited today even when capability is high).

Weight choice rationale: E1 = 0.5 (Anthropic — Massenkoff & McCrory half-weight rule for augmentation), E2 = 0.3 (human review keeps ~70% of task time), E3 = 0.2 (high theoretical capability but deployment-limited today per OECD autonomy framework).

Full calibration formula: Calibrated E-score = 0.7 × task-weighted score + 0.3 × Anthropic observed_exposure (Massenkoff & McCrory). The 70/30 blend keeps the model sensitive to role-specific task detail while anchoring to empirical data.

Use cases are generated by an LLM from the E1/E2/E3 task list and scored on six KPIs on a 1–5 scale. The scorecard draws on three of the four source papers (all embedded verbatim into the scoring prompts):

• McKinsey Global Institute (2023). The economic potential of generative AI. June 2023. Supplies the four high-value functions driving strategic_goals_alignment and revenue_growth KPIs. Verbatim from p. 12: "Our analysis of 16 business functions identified just four — customer operations, marketing and sales, software engineering, and research and development — that could account for approximately 75 percent of the total annual value from generative AI use cases." Customer-service evidence (p. 15) grounds quality_errors: AI decreased productivity for skilled agents in one study — quality KPIs cannot be assumed positive.
• OpenAI — Patwardhan, T. et al. (2025). GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. Headline result (§3.1, p. 5-6): frontier models achieve parity-or-better on 48% of real-world deliverables. Failure-mode analysis (§3.3, p. 7) — instruction-following failures, formatting errors, occasional hallucination — is the direct grounding for the quality_errors KPI.
• Anthropic — Massenkoff & McCrory (2026). Labor market impacts of AI: A new measure and early evidence. Observed exposure scores per occupation (Figure 3, p. 7) feed the people_impacted KPI — the headcount-weighted spread of AI-touchable tasks across the role set.

Scalability KPI and quick_win flag — internal design choice, not externally cited. The scalability score favours use cases whose underlying E1 patterns repeat across multiple roles; the quick_win threshold (deployable in <4 weeks with standard LLM tooling) is a consulting heuristic baked into our prompt. Neither currently has a published-source anchor.

Priority tier thresholds: impact_total ≥ 3.5 → High, 2.5–3.4 → Medium, < 2.5 → Low. impact_total is recomputed server-side as the mean of the 6 KPI scores — the LLM cannot fudge the arithmetic.

The FTE savings headline uses: FTE saved = weighted_e_score × headcount, with an "incremental" variant that subtracts already-realised baseline adoption.

The potential vs. realised distinction motivating incremental_e_score = weighted × (1 − baseline_adoption) comes from Anthropic — Massenkoff & McCrory (2026), p. 5: "AI is far from reaching its theoretical capability: actual coverage remains a fraction of what's feasible." We don't double-count savings the client has already captured through existing tools.

Baseline-adoption detection — internal design choice, not externally cited. The keyword list used to detect current AI tool usage (ChatGPT, Copilot, Claude, Cursor, Jasper, Gemini, etc.) and the 0.0–1.0 adoption scale in analyze.py are our own heuristics. They are not drawn from a published adoption survey.

Framing note: FTE savings figures are presented as reallocatable capacity (time that can be redirected to higher-value work), not as headcount reductions. This framing is a consulting choice, not a citation-backed claim.