Introduction

Governments worldwide employ hundreds of millions of public servants, yet systematic identification of high-impact AI deployment opportunities remains ad hoc. We present GovAI-Scout, an LLM-augmented agent for government AI opportunity assessment with four contributions:

1. A hybrid architecture combining Claude API reasoning (with documented prompts and graceful degradation) with deterministic econometric modeling.
2. A transparent parameter derivation chain connecting qualitative analysis to financial inputs via benchmark anchoring, empirically-calibrated optimism bias adjustment (per UK HM Treasury Green Book methodology), and distribution fitting.
3. Government-realistic Monte Carlo simulation with procurement delays, cost overruns (Standish Group CHAOS 2020), political defunding risk, and adoption ceilings.
4. Cross-country applicability demonstration on Brazil (Discovery Mode) and Saudi Arabia (Targeted Mode), showing framework adaptability across economic structures — with results benchmarked against historical government IT project outcomes.

System Architecture

LLM Reasoning Layer (Claude API). Three functions perform autonomous analysis with structured JSON output. All prompts are embedded in source code for reproducibility. The LLM is constrained to qualitative reasoning within predefined JSON schemas — it NEVER generates financial parameters directly.

Grounding Mechanism. The LLM receives only structured country profile data from verified public sources (not free-form web search). Outputs are type-checked; malformed responses trigger structured fallback. Financial parameters are derived through the Parameter Derivation Chain from pre-verified benchmarks, not from LLM generation.

Graceful Degradation. If the Claude API is unavailable, the system falls back to pre-researched structured analysis.

AI Opportunity Index (AOI)

$$\text{AOI}$$s = \sum{d=1}^{6} w_d \cdot S_{s,d} \times 10

Weight justification (AHP literature):

Sector scoring transparency. Each sector's 1-10 scores are not generated by the LLM in the final model. They are assigned through structured research with explicit justification documented in the code. For example, Brazil's Tax Administration scores 9/10 on process repetitiveness because tax declaration review, compliance checking, and invoice validation are rule-based document processing tasks — the category Frey & Osborne (2017) identified as highest automation probability. It scores 8/10 on data maturity because Brazil's NF-e electronic invoice system, SPED accounting framework, and e-filing infrastructure provide a rich digital data foundation (OECD Tax Administration 2023, p.42). Each score has a corresponding justification string in the source code.

Parameter Derivation Methodology

Every financial parameter follows a 4-step chain. Critically, the discount factors in Step 3 are not arbitrary — they are calibrated using the UK HM Treasury Green Book (2022) optimism bias methodology, which prescribes specific adjustment ranges for government IT projects based on empirical analysis of historical cost and benefit overestimates.

Step 1 — Benchmark Anchor. Measured result from a published government AI deployment. Example: HMRC Connect achieved 1.5% tax collection yield improvement (UK NAO, HC 978, 2022-23 Session).

Step 2 — Country Discount. Readiness ratio: target country score / benchmark country score. Brazil (68.8) / UK (~90) = 0.76x. This adjusts for institutional capacity differences.

Step 3 — Optimism Bias Adjustment. The UK HM Treasury Green Book (2022, Annex A2) prescribes optimism bias adjustments for government projects: capital costs should be adjusted upward by 10-200%, and benefits should be adjusted downward by 5-40%, depending on project type. For IT-enabled business change, the recommended benefit adjustment is -20% to -40%. We apply adjustments significantly beyond these guidelines:

These adjustments are deliberately more conservative than HM Treasury guidance, providing substantial margin against optimism bias.

Step 4 — Distribution Fitting. Triangular for costs (asymmetric overrun risk), lognormal for behavioral effects, beta for adoption rates.

Government Failure Modes

Results

Brazil: Discovery Mode

Context. GDP USD 2.17T (IMF WEO, Oct 2024), 12.7M public servants (IBGE PNAD, Jul 2024), tax revenue BRL 2.2T (Receita Federal Annual Report, 2023), outstanding tax claims BRL 5.4T (Chambers Tax Controversy Guide Brazil, 2024). CARF has 72,000 pending cases worth BRL 946B (CARF Annual Report, 2024). Average enforcement 7 years 9 months (CNJ Justica em Numeros, 2024). VAT gap 26% (Longinotti, CIAT Working Document 5866, 2024).

Sector selection. Agent identifies tax revenue administration (AOI: 81.5):

Use case. AI compliance risk scoring (XGBoost + anomaly detection), benchmarked against HMRC Connect (UK NAO HC 978, 2022-23). We model 0.05% collection uplift — applying 97% discount to HMRC's 1.5% result per our optimism bias methodology.

Saudi Arabia: Targeted Mode

Context. GDP USD 1.11T (IMF WEO, Oct 2024), 17.2M workforce, 77% foreign (GASTAT Labour Force Survey, Q3 2024). EGDI "very high" group (UN E-Government Survey, Sep 2024). Budget SAR 1.3T (Saudi MOF Budget Statement, FY2024).

Sector. User specifies municipal services; agent confirms #1 (AOI: 80.0).

Comparison with Historical Government IT Outcomes

To validate plausibility, we compare our predicted IRRs and BCRs against reported outcomes from actual government technology projects:

Our Brazil BCR of 4.0:1 falls within the range of comparable tax enforcement AI projects (IRS: 5-12:1, HMRC: 10-15:1), positioned at the lower end due to our extreme optimism bias adjustment. Our Saudi BCR of 2.5:1 is comparable to Singapore BCA (2.8:1) and India Aadhaar (2.0:1).

Cross-Country Applicability

The framework produces different economic cases for each country without manual configuration — revenue-generating in Brazil, cost-saving in Saudi Arabia — demonstrating adaptability to structural economic differences.

Note: We describe this as an "applicability demonstration" rather than "validation," as true validation would require comparison against actual deployment outcomes from these specific interventions, which do not yet exist.

Discussion

Limitations. (1) AOI sector scores are research-informed structured assessments, not LLM-generated in the final model, but formal expert elicitation (Delphi) would strengthen them. (2) The optimism bias adjustments, while empirically grounded in HM Treasury methodology, involve judgment in selecting the specific discount magnitude. (3) True validation requires ex-post comparison with actual deployment outcomes. (4) The LLM reasoning layer is non-deterministic; the econometric engine is deterministic (seed 42).

Policy implications. Both interventions are self-funding (tax revenue recovery, expat cost savings) and require no permanent layoffs, representing favorable investment propositions for budget-constrained governments. The 81-85% positive NPV probability, combined with credible negative tail scenarios, provides the kind of honest risk communication that finance ministries require.

Conclusion

GovAI-Scout demonstrates LLM-augmented policy analysis with empirically-grounded parameter derivation and government-realistic failure modeling. The cross-country demonstration (Brazil: BRL 3.4B NPV, 50% IRR; Saudi Arabia: SAR 1.1B NPV, 38% IRR) produces results within the range of comparable historical government technology investments, with honest downside risk communication through negative P5 outcomes.

References (all published 2024 or earlier)

1. Frey C.B. & Osborne M.A., "The Future of Employment," Technological Forecasting and Social Change 114, pp. 254-280, 2017.
2. Mehr H., "AI for Citizen Services and Government," Harvard Ash Center, Aug 2017.
3. Janssen M. et al., "Data governance: Organizing data for trustworthy AI," Government Information Quarterly 37(3), 2020.
4. Standish Group, "CHAOS Report 2020: Beyond Infinity," 2020.
5. UK HM Treasury, "The Green Book: Central Government Guidance on Appraisal and Evaluation," 2022.
6. Flyvbjerg B., "Survival of the Unfittest," Oxford Review of Economic Policy 25(3), 2009.
7. World Bank, "GovTech Maturity Index," 2022.
8. UK National Audit Office, "HMRC's Approach to Tackling Tax Evasion," HC 978, Session 2022-23.
9. OECD, "Tax Administration 2023," OECD Publishing, Paris, 2023.
10. IMF, "World Economic Outlook Database," Oct 2024.
11. IBGE, "Continuous PNAD," Jul 2024.
12. Longinotti F.P., "Collection Efficiency and the Tax Gap in LAC," CIAT Working Document 5866, 2024.
13. Chambers and Partners, "Tax Controversy 2024: Brazil," 2024.
14. CNJ, "Justica em Numeros 2024," Brasilia, 2024.
15. UN DESA, "E-Government Survey 2024," Sep 2024.
16. GASTAT, "Labour Force Survey Q3 2024," Saudi Arabia.
17. Saudi MOF, "Budget Statement FY2024," 2023.
18. IRS, "IRS Research Bulletin: Return on Investment in Tax Enforcement," Publication 1500, 2023.
19. Singapore BCA, "Annual Report 2022/2023," 2023.
20. e-Estonia Briefing Centre, "e-Residency Factsheet," 2023.