How Agents
Make a Country
Grow

State capacity ranks first in the stack because nothing else is buildable without it (ENSI, Priority 1). The lever is throughput: the rate at which the system can analyze, decide, and execute without the decision being captured or reversed. The bottleneck is brutal — capable human attention is scarce and expensive. A poor state cannot hire a Singapore-grade bureaucracy; a rich state cannot scale its best fifty analysts across every decision. You can buy a port; you cannot buy fifty incorruptible polymaths on a government salary. The constraint was never the ideas — it was the number of competent hours (ENSI, Czech State: A Systematic Analysis of Malfunction).

The pilot-agency stops being a luxury good and becomes infrastructure any government can stand up — scarce humans elevated from doing the analysis to judging it (Tony Blair Institute).

1. Secure the principal, narrow the mandate. A head of government who protects autonomy and a concrete, measurable mission — not another dashboard (ENSI, Priority 1).
2. Lay the rails. Agents need DPI — consented data-exchange, machine-readable registries (World Bank, GovTech Maturity Index; UCL IIPP).
3. Discipline test. Every bet carries a falsifiable forecast and the export test — earn foreign sales in a window or lose support (World Bank/Stiglitz, East Asian Miracle).
4. Guardrail. HITL on every consequential call; immutable audit log of recommendation, evidence & approver (NIST, AI RMF; EU AI Act).

• Korea — the EPB super-ministry ran the heavy-and-chemical push above the line ministries (ADB, Chaebol and Industrial Policy).
• Ireland — the IDA identified, courted and landed specific multinationals over decades (OECD, Economic Survey of Ireland 2025).
• Estonia — a small, capable administration on the right digital rails outperformed far larger bureaucracies (OECD, Economic Survey 2024).

The discipline mechanism ranks second because it converts capacity from a danger into an asset (ENSI, Priority 2). The East Asian states did not pick winners well — they had a mechanism to kill losers fast, the export test, an external referee that cannot be lobbied (World Bank/Stiglitz). The value was never in selection; it was in the exit. The bottleneck: rigorous ex-ante analysis is staggeringly expensive in the scarcest input — capable analytic hours. No state can model more than a handful of flagship decisions a year, so the thousand others ship on whoever argued most persuasively, gathering only the evidence that agrees (ENSI, Evidence over Anecdote).

The state begins to learn: every forecast frozen, every outcome tracked, so it accumulates an evidence base instead of a pile of war stories (ENSI, Evidence over Anecdote).

1. Start where forecasts are expected. Begin with budget & fiscal decisions, then widen the frozen-forecast rule to regulation and subsidies.
2. Build the model layer. Data rails + DPI plus the National Policy Twin, a live queryable model of the policy estate (Tony Blair Institute; World Bank, GovTech Maturity Index).
3. Discipline test. The frozen ex-ante forecast itself: miss it by a defined margin and a mandatory review defaults toward exit — the generalised export test (ENSI, Priority 2).
4. Guardrail. Agents inform, an accountable human votes; immutable audit log of every model, assumption & analogue (WEF; NIST, AI RMF; EU AI Act).

• Korea vs Brazil — the export test separated Korea's directed credit from Brazil's (World Bank/Stiglitz, East Asian Miracle).
• OECD governments — already running AI-based simulation, scenario analysis & decision support across functions (OECD, Governing with AI).
• TBI / WEF — a worked National Policy Twin and a near-term readiness framework for agentic systems in government (Tony Blair Institute; WEF).

The lever is Priority 4 of the Growth Stack — pick tradables and climb the complexity ladder — fused with Priority 13, structural transformation. The climb's hidden killer is distance: ambitious leaps that ignore the capability gap fail; disciplined short jumps compound. Targeting that judgment well demands holding three fast-moving things in one head — current capabilities, the global product space, and the adjacency between them — and redoing it every quarter. That attention is the scarcest, least-exportable ingredient in development; most states cannot buy a Singapore-grade planning unit, so targeting collapses into drift on static comparative advantage or capture by the loudest incumbents.

Capability distance becomes a measured quantity rather than a hunch — the difference between a disciplined short jump that compounds and an ambitious leap that fails.

1. Data substrate first. Consented, machine-readable customs/export records, firm and supplier registries, patent and standards filings; where DPI is thin, building it is the first deliverable.
2. Map before you bet. Run the Mapper and Cartographer for a full cycle and ship a ranked, sourced adjacent-possible map — and nothing else — before any money moves.
3. Discipline test. Does each targeted bet, within its window, earn hard currency from foreign buyers who owe the country nothing? Pass → support continues; fail → the fleet climbs back down and proposes another rung.
4. Guardrail. Human-in-the-loop at the bet, not the analysis; full audit trail and a published ranking on every recommendation to keep cronies out.

• Korea — channeled subsidized credit, then forced a climb from textiles to ships to semiconductors within a generation via the Economic Planning Board (ADB; PIIE).
• China — Made in China 2025, the most explicit published map of the climb, from low-value assembly into robotics, aerospace and advanced semiconductors (Gov. of China; MERICS).
• Taiwan / Ireland / Vietnam — SME-led ascent into semiconductor indispensability; FDI used as a capability-transfer machine (NDC; OECD; ERIA).

The lever is Priority 2 of the Growth Stack: subsidize nothing without a hard external performance test. Support is conditional on selling in world markets within a defined window — foreign buyers cannot be lobbied, bribed, or charmed. The rule is easy to write and almost impossible for humans to enforce: measurement degrades into self-report theater; the verdict to cut a firm is a confrontation with a constituency that employs voters and funds campaigns; and the referee's goalposts get lobbied soft. The binding constraint is the throughput of honest, unconflicted monitoring and the political shielding of the verdict — a capacity problem, not a character problem.

A live, customs-data-backed scorecard makes sparing a failing crony a public act against visible evidence — honesty by publication.

1. Data substrate first. Machine-readable, consented customs and export records, the registry of every support instrument, the firm/supplier graph — recipient-proof, or there is no discipline machine.
2. Instrument, then publish early. Stand up the Instrumenter so no new support enters untested, back-fill the existing portfolio, and publish the scorecard from the start — the public verdict is the mechanism, not a flourish.
3. Discipline test. Does each supported firm, within its window, earn hard currency from foreign buyers who owe the country nothing? Pass → support continues; fail → it ends, automatically and visibly.
4. Guardrail. Human-in-the-loop on the cut with an appeal path; audit trail on every verdict; protect the referee — every change to a threshold logged and published; falsifiable conditions only.

• Korea — withdrew subsidized credit from chaebol that missed export benchmarks, forcing a climb from textiles to ships to semiconductors in a generation (ADB; PIIE).
• China / Singapore — SEZ firms competed for global supply-chain orders with officials promoted on export performance; incentives routed to export-purposed MNCs (PIIE; IMF/Ghesquiere).
• Botswana — excellent fiscal discipline over diamond rents, but no manufacturing export contest, so it struggled to diversify — absence proves the rule (World Bank; NRGI).

The lever is state effectiveness, and through it, legitimacy: a state that can reliably register a business, pay a benefit or issue a permit earns the consent that lets reform survive an electoral cycle. The bottleneck is structural — casework has four steps (intake, retrieval, reasoning, resolution) that each consume scarce attention, capacity is bound to labour, and a poor state cannot recruit a Singapore-grade bureaucracy. With one-third of the US SSA's ~22,000 staff due to retire within five years as claims rose (Harvard Ash Center), the scissors of rising need and falling capacity are universal.

Capacity expands without expanding labour (TBI); personalisation and prevention replace standardised, reactive provision; inclusion of minority-language and rural citizens becomes a side effect, not a special programme (World Bank).

1. Lay the rails first. Digital ID, a data-exchange layer, and readable/writable systems of record — without DPI an agent has nothing to act through (World Bank).
2. Pick & instrument one service. A single bounded, high-volume, rule-governed benefit; measure the baseline end-to-end before automating, then widen autonomy case-type by case-type.
3. Discipline test. Resolution time, accuracy vs. human adjudication, appeal-and-overturn rate, cost per case and satisfaction — published against the baseline; an agent that misses is switched off, not defended.
4. Guardrail. Human-in-the-loop on denials and high-stakes/low-explainability cases (IMF); immutable audit log of inputs, rule, reasoning, actor; consent- and purpose-limited data exchange.

• US cities & Japan — five core capabilities (answering, document-handling, routing, translation, drafting) already operating in citizen services (Harvard Ash Center, 2017).
• World Bank survey — governments using AI across citizen engagement, fraud/anti-corruption, business-process automation and service delivery — the exact agentic stack.
• Leading digital states — UN e-government & World Bank GovTech work show the rails (digital ID, data exchange, systems of record) already built, making the agentic layer buildable.

The lever is fiscal capacity and the accountability loop it disciplines: "Coercion collects pennies; legitimacy collects the GDP." A broad, even-handed base delivers predictable finance, cheaper borrowing and consent for voluntary compliance. The bottleneck is a vast sorting problem on scarce attention — an administration can audit only a tiny fraction of returns, so the deterrent rests on selecting the right fraction, while fraud moves at transaction speed and the informal base is simply invisible. "A taxpayer the state cannot identify is a taxpayer the state cannot tax" — the narrow-base trap is the shadow of scarce attention.

Compliance by design (OECD 3.0 "making tax just happen"); enforcement that is sharper and fairer — fewer pointless audits of the honest, and the powerful scrutinised exactly as everyone else.

1. Build the spine. Universal taxpayer identity on the national digital ID + data-exchange rails for third-party data — without legibility there is nothing to reason over (World Bank).
2. Win the upstream, then promote. Deploy service + pre-fill first (cheapest revenue is the correct-by-default return), run risk-selection in shadow mode, promote to live only where it beats baseline; add real-time fraud last.
3. Discipline test. Audit hit-rate and revenue-per-audit vs. human selection, plus false-positive rate on the compliant; an agent that misses on effectiveness or fairness is switched off (IMF: "Evaluate Use Cases for Performance and Intent").
4. Guardrail. Human-in-the-loop on every punitive determination (IMF; legally required); published AI-use inventory + disclosure + audit log; bias-testing on selection models trained on historical audits; consent- and purpose-limited data exchange.

• Estonia — taxation built on national digital ID and the X-Road backbone makes filing near-automatic and evasion structurally hard; among the lowest compliance costs in the OECD.
• India — taxation built atop the Aadhaar identity and digital-payments stack, formalising millions of new taxpayers and widening the taxable base.
• Rwanda — rebuilt a revenue authority from near-zero and sharply raised its domestic-revenue share, cutting aid dependence (Priority 18).

Clean procurement maps to two Growth-Stack priorities: fiscal efficiency (18) — every unit overpaid for a road is a road un-built — and clean governance (11), where rigged contracts sever the link between merit and reward a market needs. The constraint is throughput, not character: a state already has rules against overpaying; it lacks the bandwidth to check, contract by contract, whether they were followed. Cross-checking a tender means reconciling bidders against ownership registries, prior awards, price benchmarks and conflict declarations — expensive, repetitive work done by sampling, late, or never. Armenia gave its anti-corruption commission an asset-declaration register but its limited resources made meaningful checks impossible (OECD). The data was filed; nobody could read it.

Population-scale screening replaces sampling; the buyer gains total recall of every price paid and every firm caught; designed-out discretion becomes the cheap default rather than the costly aspiration.

1. Fix the data before the model. Structure and link supplier-ownership, payroll, awards, benchmarks and contracts first — Armenia's declarations weren't even machine-readable; an agent on garbage produces confident garbage (OECD).
2. Start where the join is unambiguous. Begin with conflict-of-interest detection (the Brazil pattern); crisp wins buy cover for fuzzier collusion work, then move screening upstream from awards to bids to tender design.
3. Discipline test. A published scorecard: tenders screened, value of confirmed collusion/conflicts, price anomalies caught, delivery shortfalls recovered, and unit prices now-vs-before. No catch in money = theatre.
4. Guardrail. Human-in-the-loop on every consequence; tamper-resistant audit logs & provenance on every flag; transparency of method with a privacy boundary; independence from the awarding authority.

• World Bank — the Governance Risk Assessment System (GRAS): advanced data analytics for detecting fraud, corruption, and collusion in public expenditures (2023).
• Brazil — on a World Bank–funded project, AI detected 500 firms owned by the civil servants supervising their own contracts — a join of three registries.
• Spain — the General Comptroller (IGAE) uses supervised ML, trained on proven cases, to detect fraud in public grants (OECD 2024).

Red tape maps to two priorities: property rights & ease of doing business (10) — an unprovable title is dead capital, an un-registered idea is no firm — and competition (23), because capture operates through rules, not just market power, and red tape is the incumbent's moat. The constraint is throughput, not the rulebook's length. Attention is scarce in both directions: writing rules well means tracing every interaction with the existing thicket; running them well means processing a flood of permits no office can sustain. The cost is regressive — a large firm amortises a compliance department; the entrant faces the same fixed wall with no volume, so it stays home. Friction is a direct tax on production (OECD).

Time-to-firm collapses; de-moating becomes tractable when an agent can read all the rules at once; supervision shifts from periodic autopsy to live monitor; rules ship with side-effects modelled, not discovered in the field.

1. Fix data & digitise rules first. An agent cannot issue against rules locked as prose in a drawer; build the digital rights stack — registries, identity, structured rules — the Estonia/X-Road precondition, before any ML (OECD).
2. Start with permit automation. Auto-issue the most rule-bound, highest-volume registrations (company registration, routine licences) where criteria are objective, the win is visible, and a wrong issuance is low-stakes and reversible.
3. Discipline test. Publish median time-to-permit and time-to-firm before/after, share of approvals issued without human touch, rules rationalised away, and supervision flags that became enforced actions.
4. Guardrail. HITL on every rejection & enforcement; a right to explanation and appeal; transparency of method; audit logs & provenance; an independent supervisor the regulated industry cannot quietly retrain.

• Estonia — digitised the entire land/business/contract rights stack into a tamper-evident e-state via X-Road; a leader in e-government (OECD, Economic Surveys 2024).
• Rwanda — 3rd-best country for business in Africa and among the world's most-improved (World Bank Doing Business); targets the top 10 by 2035 via the Irembo platform.
• Singapore — ranked #1 for ease of doing business (IMF), uses regulatory sandboxes to support innovation and risk-taking — the template for testing agentic regulation safely.

Clean governance sits in the same load-bearing tier as property rights: control corruption and a country becomes faster, cheaper to operate in, and more meritocratic (Singapore — IMF/Ghesquiere; Botswana — Acemoglu, Johnson & Robinson). The report's prescription for priority 11 is to attack discretion first — every point where an official says yes-or-no for a price is a corruption site. But anti-corruption is monitoring work, and monitoring is exactly what scarce human attention cannot do at scale. Armenia's Corruption Prevention Commission collected every asset declaration and then could not read them: "given the large number of officials… and the CPC's limited resources, it was difficult to perform any meaningful checks" (OECD 2024).

Agents change the price of vigilance: a poor state cannot recruit a Singapore-grade integrity corps, but it can rent or run a fleet that does the cross-checking that corps would have done.

1. Fix the data first. Structure and link registries (ownership, payroll, contracts, declarations, land) before pointing agents at them — Armenia's decisive early lesson; declarations were not even machine-readable (OECD 2024).
2. Start where the join is crisp, then learn. Begin with Brazil-type conflict-of-interest detection; deploy rule-based screening; layer pattern-learning once trusted.
3. Discipline test. Publish the catch — rate of declarations screened, value of flagged-and-confirmed conflicts, procurement price anomalies caught. A fleet that cannot show its catch is theatre.
4. Guardrail. HITL on every consequence, tamper-resistant audit logs & provenance, transparency about method (not the algorithm), independence from the executive (NIST AI RMF; OECD; EU AI Act).

• Armenia — CPC built a linked data platform, deployed automated red-flag screening of asset declarations, now piloting pattern-learning AI (OECD 2024, Box 2.5).
• Brazil — World Bank–funded AI detected 500 firms owned by the civil servants supervising their contracts (World Bank).
• Spain — tax authority uses machine learning to assess public-grant fraud risk (OECD, Countering Public Grant Fraud in Spain).

Secure, trusted rights are the precondition that makes every other priority bankable: without them FDI stays footloose, firms stay small and informal, and credit stays scarce because no one will lend against collateral a court won't protect (Botswana — Acemoglu, Johnson & Robinson; Mauritius — Subramanian, UNU-WIDER). The report's priority-10 build order reads as an agentic stack: registries first, then enforcement speed, all digitised so records cannot quietly disappear. The constraint is human throughput — you cannot mint experienced judges or ask clerks to read faster, so dockets clog and a title you cannot prove to a stranger is dead capital.

A current, queryable registry turns dead capital live so banks lend against it; collapsing the docket re-prices every long-horizon bet in the economy.

1. Registries before rulings. Digitise & structure land, company and collateral registries, then run agents to keep them current — the RPA-grade work with the cleanest payoff and lowest sensitivity.
2. Triage before drafting; drafting narrow, supervised, last. Multiply judicial throughput before any agent touches the substance; expand the envelope only as the review record proves reliability.
3. Discipline test. Publish median time-to-judgment, registry currency & search latency, the share of agent-drafted rulings accepted unedited vs. overturned, and the reversal rate of triaged cases.
4. Guardrail. Judgment stays human; provenance & explainability on every output; bias auditing; tamper-resistant logs; a right to human review (EU AI Act classes justice AI high-risk; NIST AI RMF; OECD).

• United Kingdom — RPA cleared 30,000 pension claims in two weeks vs. months of manual work (World Bank) — the docket/registry backlog twin.
• France — LLaMandement summarises legislative text, demonstrating the supervised legal-drafting layer (OECD 2024).
• Estonia / Rwanda — minutes-long company & tax portals; aggressive property-registration & contract-enforcement reforms (growth report; World Bank, Rwanda CEM).

For a country that has already built state capacity and an industrial-learning engine, the innovation system (priority 14) converts temporary catch-up into permanent frontier membership. Aghion, Jones & Jones (NBER WP 23928) show ideas are the binding input to growth and are getting harder to produce; the scarce input to ideas is a handful of senior minds who read the literature, frame hypotheses, design experiments and read results — serially, in working hours. Automating idea-production relaxes the one constraint growth theory says binds.

The bridging capability — the connective tissue between lab and firm that the report names the true constraint — is partly manufactured, raising the rate of indigenous innovation that beats the middle-income trap.

1. Don't build it early. A layer-three move; before state capacity (P1) & the learning engine (P6) it is an automated paper mill.
2. Bridge before brilliance. Stand up the frontier-scanning, product-translating fleet first; concentrate it on 2–3 national missions, then close the loop to the lab and diffuse to the supplier base.
3. Discipline test. Not "did we publish?" but did the discovery get commercialised by a domestic firm into something sold on a world market? Score in products shipped & export revenue; kill missions that fail.
4. Guardrail. Human judgment at the gate (models "cannot reliably deal with facts" — HAI); full provenance/reproducibility audit trail; sovereign control of the discovery estate.

• Stanford HAI — AI Index 2024 records AlphaDev finding faster sorting algorithms & GNoME accelerating materials discovery; names 2023 the year AI became an engine of science.
• Taiwan / Korea — ITRI seeded a semiconductor industry across thousands of SMEs; Innovative Korea (World Bank) shows a bridging institution, not a budget, converts knowledge into rents.
• Israel — Office of the Chief Scientist & the Yozma programme built one of earth's densest concentrations of high-wage technical employment by design, not wealth.

Aggregate productivity is a weighted average dragged down by an enormous unproductive tail; the bottom half of firms getting 30% better moves the national average enormously. The OECD's Impact of AI on Productivity, Distribution and Growth (2024) makes closing the frontier–laggard gap the pivotal variable. Capability has stayed trapped because its carrier is a competent person — and there is one McKinsey consultant per several-million small firms; none will visit a workshop in a secondary city. Diffusion has always been rate-limited by scarce human carriers.

Because gains concentrate on the least-skilled, diffusion is distributionally progressive by its own mechanism — it raises the floor, keeping the politics of growth survivable.

1. Fix the foundation. A P6 move — capability poured into firms under an overvalued currency (P5) or no contract enforcement (P10) is wasted.
2. Point at the laggards. Left to the market, the frontier adopts first and the gap widens (OECD's bad outcome); route capability to the bottom, anchored to existing supplier-development & standards bodies; build for learning, not lock-in.
3. Discipline test. Does the firm raise measured productivity and win sales on a competitive (ideally export) market within a defined window? If not, the support stops — the agents that diffuse also instrument the test.
4. Guardrail. Firm owns its data (no foreign-platform dependency); high-stakes outputs (credit, filings, structural eng.) route through human sign-off; resource rural/informal/female-led firms by design (ILO).

• NBER WP 31161 — Generative AI at Work: +14% avg, +34% novice across 5,179 live workers; mechanism is disseminating top performers' tacit knowledge; gains persisted after the tool was removed.
• Taiwan / Costa Rica — semiconductor capability diffused across thousands of exporting SMEs via ITRI; a dense supplier base built around medical-device anchors, not a single champion.
• OECD (2024) — frames closing the frontier–laggard firm gap as the variable that decides whether AI grows an economy broadly or merely concentrates it.

The lever is strategic trade openness on the country's own terms (priority 12) and the export discipline mechanism beneath it (priority 2): world demand is the impartial examiner a protected home market never is — Korea tied subsidized credit to export quotas firms could not fake or lobby away (World Bank/Stiglitz, East Asian Miracle). The bottleneck is bandwidth, not character: a small firm faces the same fixed wall of specialist transaction work as a large one, with no volume to amortize it — so it never reaches the referee at all.

Cheap standards-compliance also makes the climb up the value chain cheap, and every cleared shipment yields the unfakeable customs data that instruments every subsidy.

1. Lay the rails first. A national single window, digital customs interface, machine-readable tariff & standards schedules — agents need something to act through.
2. Start with classification & clearance, then finance and standards, then market intelligence last. Highest-volume, most measurable step first; demand-scanning once narrow agents earn trust.
3. Discipline test. Does it raise non-commodity exports, the number of exporting firms, and the small-firm share among them — customs-verified, on a clock, cut if not.
4. Guardrail. HITL gates for high-value/anomalous consignments; full audit logs; agent outputs are recommendations subject to the customs risk engine, never unchecked authority.

• Mauritius — used export-processing zones to enter the global textiles chain on its own terms, then diversified into tourism, finance and ICT (Harvard/Frankel; ODI; UNU-WIDER).
• Vietnam — sequenced integration into global value chains, entering at assembly and climbing, anchored by Samsung (Govt. of Vietnam SEDS 2021–2030; CEPII; ERIA).
• Korea — reciprocal credit contingent on export targets measured in customs data firms could not manipulate (World Bank/Stiglitz; ADB).

The lever is FDI with spillover conditions — import capability, not just capital (priority 9): "the jobs are not the asset; the capability the jobs leave behind is the asset." The bottleneck is that the institution which buys the shortcut is built almost entirely from the scarce resource the country lacks — analysts who target, deal-makers who court for years, case officers running the one-stop pipeline, and the most-neglected supplier-development teams who qualify local firms. That attention-hungry work is exactly what gets cut, and exactly what agents are best at.

The agency shifts from chasing capital on the lowest tax rate to engineering measured capability — and the anchor strategy compounds without a staffing ceiling.

1. Secure the principal, narrow the mandate. A political patron protects autonomy; target two or three concrete activities, not "promote investment" — agents inherit the focus.
2. Pipeline agent first, then targeting, then supplier matching before the anchor lands. The spillover window opens the day the investor arrives, so qualify local firms in advance, not years late.
3. Discipline test. Judge on spillover, not splash: local-supplier content as a share of anchor procurement, locals promoted into technical roles, exports, spin-offs — on a clock, redesigned if it delivers payroll without capability.
4. Guardrail. HITL sign-off on every incentive & material concession; full audit logs; transparency on what each investor got and committed to; agents recommend, never unilaterally grant.

• Costa Rica / CINDE — landed Intel as an anchor in 1996–97; one plant catalysed a high-tech & medical-devices cluster (World Bank, Impact of Intel; Harvard CID WP 58).
• Ireland / IDA — targeted, courted and embedded Intel, pharma and tech; the cluster's Irish managers seeded a domestic ecosystem (Enterprise Ireland; OECD 2025).
• Singapore / EDB & Vietnam — EDB sustained selective capability-import for decades (Govt. Singapore CFE; ADBI); Samsung pulled suppliers into Vietnam's export base (ERIA).

Human capital aligned to production (priority 7) — not credentials, but demonstrated capability that maps onto what firms actually hire for; the German–Swiss dual model is the alignment benchmark (OECD). The bottleneck is the price of one-to-one human attention: tutoring every learner to mastery and reading every employer's demand and brokering every match is caseworker work, and skilled human hours are scarce. So countries ration capable attention — and the rationing shows up as credential inflation beside unfilled vacancies.

Mastery at population scale, a labour market with a live nervous system, and the dual system's alignment without Germany's century-old institutions — reaching the informal, rural, low-literacy cohort voice-first at the cost of a phone.

1. Instrument demand first. Build the live skills-demand map before tutors, so they teach toward signal, not a guess.
2. Tutors to mastery, not seat-time. Advancement gate is demonstrated competence, validated against the demand map (UNESCO: validate before deploy).
3. Discipline test. Score on placement & wage outcomes the system cannot fabricate — never enrolments or certificates issued; cut modules that do not move placements.
4. Guardrail. Human teacher in the loop; audit the demand map for bias; no opaque automated denial of opportunity; verify mastery against on-the-job performance.

• Singapore — SkillsFuture: national employer-linked, lifelong skills funding; the engine is this with tutoring + matching run continuously at the margin for free.
• South Korea — scaled engineering & technical education in step with its move into electronics and heavy industry; agents do that demand-tracking live, not in five-year plans.
• ILO — augmentation dominates: only 24% of clerical tasks highly exposed; in low-income countries 10.4% of jobs face augmentation vs 0.4% automation — a training opportunity.

Hard infrastructure plus cheap, reliable energy (priorities 8 & 26) — but the corpus's own diagnosis is operation, not capacity: "an unmaintained asset is a stranded cost," and "a competitive currency and a slow port cancel out." The bottleneck is expert operational attention — continuous, real-time, quantitative control of coupled systems, done today by a few human controllers. There are never enough dispatchers, predictive-maintenance engineers or network planners; the rationing shows up as low throughput, high losses, and factories self-providing diesel backup.

Reliability that ends the self-provision tax, recovered water/power/fuel, and a grid that can absorb intermittent renewables while keeping firmness.

1. Instrument & observe first. Build sensing/data, run agents in shadow mode; start with the grid, the most binding constraint.
2. Automate the reversible, recommend the consequential. Execute low-stakes moves; keep a human for load-shedding or taking a substation offline.
3. Discipline test. Tie every deployment to unfakeable physical output vs a pre-registered baseline — throughput, uptime, losses, landed cost; decommission what doesn't move the box or keep the lamp on.
4. Guardrail. HITL for safety-critical control; full audit logs & explainability; graceful degradation to manual fallback; secure the control stack as critical national infrastructure.

• Singapore — a leading container port with no domestic energy, competing on grid reliability and a transparent power market: reliability substitutes for endowment (IMF).
• China — "massive ahead-of-demand investment in generation and transmission" underwrote three decades of expansion (MGI); the return now depends on operating it better, not building twice.
• Vietnam · Costa Rica — power + port expansion sequenced behind FDI factories (ADB); near-total renewable electricity via state utility ICE — keeping a high-renewable grid firm is the agent's home turf.

The lever is health as productive infrastructure (priority 22): a population's health is an input to production exactly as a road or power grid is. The bottleneck is the doctor's attention — the most expensive, least scalable input in the system. It is throttled by absolute scarcity (WHO's 10M gap, OECD's 3M deficit), by misallocation (clinicians buried in documentation and coding), and by burnout that corrodes the attention already paid for: depersonalisation rose from 25% to 44% during COVID (OECD, 2024).

A competent first contact in every village without a clinic in every village — more healthy working years per citizen, bought cheaply (OECD, 2024).

1. Back-office first. Note-drafting, record completion, translation, scheduling — lowest risk, immediate workforce relief, builds trust (WHO lists these as most mature).
2. Then judgment, last. Triage with a clear human route; clinician decision support delivered to a human who signs off; ministry-run population surveillance.
3. Discipline test. Every agent measured against confirmed clinical outcomes — diagnostic accuracy, under-triage rate, error vs human baseline; publish the scorecard, cut what fails.
4. Guardrail. Human clinician is the accountable decider, always; informed consent, audit logs, built-in refusal/escalation, and a named liability owner (WHO six principles).

• OECD survey (18 assns) — 70% agree the physician's role stays central; no respondent believed AI would replace doctors.
• WHO LMM guidance — catalogues virtual health assistants, complex-case & routine-diagnosis support, language translation as near-term uses.
• OECD — AI already streamlining screening & diagnosis, advancing antimicrobial-resistance work and personalised monitoring.

This serves human capital aligned to production (priority 7) and the demographic dividend (priority 19) — the same lever from two angles. The dividend is conditional, and the condition is learning. The bottleneck is teaching attention, worse than the doctor shortage because demand is universal and simultaneous. One-to-one is the only proven intervention and the one a teacher-short, fast-massifying system cannot afford — the bottom income quintile is still just 24% of enrolment (World Bank, 2025).

Bloom's 2 Sigma finally delivered to a whole population — reaching the rural, non-native and special-needs learners the system was failing.

1. Copilot first. Grading, materials, lesson drafting, at-risk flagging — lowest risk, immediate relief, and it puts teachers (not vendors) in control of the tech.
2. Then tutors, foundational skills first. Literacy, numeracy, language where mastery is measurable; higher-order & project-based support last, where the agent risks doing the thinking for the student.
3. Discipline test. Does it raise measured learning — not engagement or screen-time — without degrading the capability it builds? Controlled comparison vs the classroom baseline; publish; cut what flatlines (UNESCO).
4. Guardrail. Data-privacy and an age limit; protect human agency; inclusion, equity, linguistic diversity; teacher in the loop — UNESCO's human-centred floor.

• Harvard (Kestin et al., 2024) — AI-tutored students learned more than twice as much in less time than active-learning peers.
• Stanford (Wang et al., 2025) — AI-enhanced tutoring scaled expert teaching practices at ~$20 per tutor per year.
• UNESCO GenAI guidance — GenAI as a 1:1 coach for foundational skills, inquiry & project-based learning, and support for special needs.

A country grows on a competitive real exchange rate and the fiscal capacity that funds a capable state (Growth-Stack priorities 5 and 18). An overvalued currency is a tax on every export and a subsidy on every import — the most common, most invisible way promising economies kill their tradable sector before it learns. The constraint is not data but attention: scarce, senior judgment applied continuously across a dozen interacting variables. So debt analyses run yearly, exchange-rate reviews are episodic, and the slow drift lives precisely in the gap human vigilance cannot cover.

A poor treasury gets the continuous structural-balance estimation a rich one buys with hundreds of economists — separating windfall from wallet in real time, and manufacturing credible commitment lenders can price.

1. Instrument & observe. Connect agents to existing feeds and have them reproduce official analyses — a mismatch is a bug, not a result.
2. Add cadence & breadth. Move validated runs from annual to nightly across the full interacting board, with named thresholds; then publish to earn credibility.
3. Discipline test. Not accuracy but whether the state acts on the flag — every flag, response and outcome logged, mirroring the corpus "did you export?" external test (priority 2).
4. Guardrail. HITL on every decision; full auditability of model version & inputs; contestable, traceable flags (EU AI Act; NIST AI RMF).

• Taiwan — held money cheap and the rate competitive while thousands of small manufacturers built export capacity (World Bank).
• Botswana — managed diamond rents so the windfall never appreciated the pula into a tradables-killer (UNRISD).
• IMF deployments — AI in tax & customs already doing compliance nudging, audit-target selection and real-time fraud detection (IMF; OECD Tax Admin 3.0).

Resilience and policy continuity (priority 28) is the meta-discipline that lets every other priority survive long enough to work — the largest source of wasted development is good policy abandoned before it matured. The damage is rarely the shock itself; it is the lag between arrival and a coherent answer. The binding constraint is capable attention, demanded continuously in calm and all-at-once in catastrophe — the two things human teams are structurally worst at: monotonous vigilance, and synthesis under a firehose.

Lead time turns catastrophe into incident; a poor state gets the war-gamed readiness a rich one staffs with a permanent planning unit; and the strategy survives the shock instead of resetting to zero.

1. Instrument the watch. Connect early-warning agents to existing signal streams with named thresholds and a clear escalation path.
2. Build & exercise. Generate war-gamed scenarios, validate against historical events and the 24 comparators, drill relentlessly; wire coordination last, routed through human command.
3. Discipline test. The exercise, before the crisis — every flag scored against what followed, every drill an after-action record; an un-drilled library fails automatically (corpus priority 2).
4. Guardrail. Human command on consequential action; auditability through the fog; graceful degradation that never shows a confident picture on stale data (NIST AI RMF; arXiv oversight structures).

• World Bank — survey of AI in the public sector locates clearest wins in continuous monitoring of large, messy flows (AI in the Public Sector).
• Alan Turing Institute — mapping of generative AI to government work finds the same early-warning sweet spot (Mapping the Potential).
• WEF — readiness framework for agents that plan, decide and act across organizational boundaries under oversight (Making Agentic AI Work for Government).

The lever is the digital state: the capacity to act on citizens and firms accurately, instantly and at scale. DPI collapses the three transaction costs that tax every state–economy interaction — knowing who you deal with (identity), moving money (payment), finding a fact the state already holds (consented data-exchange, under the once-only principle). The binding constraint was never the decision; it was the clerical labour wrapped around it. Identity was re-verified at every counter, payment batch-processed over days, facts siloed across ministries — forcing the state to ration transactions and stay small relative to its economy (UCL IIPP 2025).

Each remaining area becomes one agent composing the same three primitives in a different order — benefit delivery (Area 5), tax (Area 6), procurement (Area 7), FDI & trade (Areas 13–14) — so the report becomes a deployment schedule, not a wish-list (UCL IIPP 2025).

1. Identity first. Machine-verifiable through an open interface, not a plastic card — nothing else authenticates without it.
2. Then payments, then data-exchange. 97 countries have payment rails (Pix, UPI, CoDi); add consented exchange (X-Road, Trembita) under the once-only principle (UCL IIPP 2025).
3. Discipline test. Can a third-party agent authenticate, retrieve a verified fact and move value through documented open APIs, no human, in <10s — with that rate published monthly? Tie it to the World Bank GovTech Maturity Index (World Bank).
4. Guardrail. Recorded consent token + tamper-evident log per transaction; purpose limitation; revocable, privacy-preserving identity (80% of ID systems meet privacy variables today; target 100%) (UCL IIPP 2025).

• India — India Stack (Aadhaar + UPI + consented data) turned account-opening, subsidy disbursement and merchant payment from multi-day paper processes into instant API calls (UCL IIPP 2025).
• Estonia / Ukraine — X-Road and Trembita wired registries into one secure data-exchange backbone, making the once-only principle the state's default behaviour (UCL IIPP 2025).
• Brazil / Mexico — Pix and CoDi put bank-grade instant settlement into the informal economy at near-zero fee (UCL IIPP 2025).

The lever is unusual: it is the substrate for every other lever — compute to run the fleet, data it learns from and acts on, talent to build and supervise it. The bottleneck is one level up: not scarce attention to do the work, but scarce capacity to control the workers. In the pre-sovereign world a country could only rent — model, compute and data sent abroad. That is fine for a chatbot and catastrophic for a state: rented capability can be repriced, rate-limited or cut off; data abroad is sovereignty surrendered; and a borrowed, opaque model cannot be audited, red-teamed or governed (OECD).

Every other area (tax, courts, health) runs on allocatable national compute over the state's own data, supervised by its own talent — auditable, dependable, and able to keep running even if a foreign supplier is cut off (UK Govt; Singapore Govt).

1. Compute first, allocatable. Secure guaranteed-allocation compute — the UK's 20× AIRR and AI Growth Zones fast-tracking land and power are the template (UK Govt).
2. Govern data, then build talent. Treat the public data estate as a strategic asset; train and anchor the people who supervise the fleet (Singapore Govt; SDAIA). Sovereign models last, only where the market fails.
3. Discipline test. If our largest foreign supplier cut us off tomorrow, could critical agentic services keep running on infrastructure and data we control, staffed by people we employ? If no, it is rented, not sovereign — publish the dependency map and continuity plan (UK Govt).
4. Guardrail. State models audited and red-teamed before deployment; compute and data carry access controls and logs; all of it operates inside Area 23's HITL governance so "the state controls the AI" never becomes "the AI runs the state" (OECD).

• United Kingdom — AI Opportunities Action Plan: 20× sovereign compute, AI Growth Zones, data unlocking and a talent pipeline, framed as "AI maker, not AI taker" (UK Govt).
• Singapore — NAIS 2.0 organises the strategy around Compute, Data and Talent in a "Trusted Environment," with a 15,000-practitioner target and selective peaks of excellence, not autarky (Singapore Govt).
• Saudi Arabia / India — SDAIA puts data first and aims to export Data & AI by 2030; India's NITI Aayog strategy and the OECD overview show the same compute–data–talent shortages recurring (SDAIA; OECD).

The lever is legitimacy — the precondition that lets every hard reform survive contact with politics. The OECD frames its AI standard around exactly this: to foster innovation and trust by promoting responsible stewardship of trustworthy AI (OECD/LEGAL/0449). The bottleneck is scarce, expensive human attention at its worst: oversight is pure attention, producing nothing but judgement on whether something else was done right. A poor state cannot staff a competent front line, let alone a second line of auditors to watch it — so review collapses into rubber-stamping. Worse, a tired human stapled to a fast agent's output succumbs to automation bias, which the EU AI Act (Reg. (EU) 2024/1689, Art. 14) names explicitly. Naive "human in the loop" can launder machine decisions, not check them.

Oversight finally keeps pace with action; every consequential decision is logged with its inputs, model version and approver — legible government, and reform that survives its own scandals because it can show what went wrong, who owned it, and what changed.

1. Govern first, then classify. Stand up the NIST RMF Govern function as the spine — who is accountable, what risks are tolerable, what the published test is — then map every deployment to an EU-AI-Act risk tier.
2. Instrument before launch; gate, then loosen. Logging, provenance and continuous evaluation are launch blockers, not roadmap items. Start consequential agents in human-approval mode and widen autonomy only as the measured record earns it.
3. Discipline test. Is it instrumented against an objective external standard, is the standard published, and are failures acted on — including shutting the agent down? When the scorecard goes red, an automatic consequence, not a meeting.
4. Guardrail. No silent autonomy on rights; independence of the watchers from the unit they police; provenance mandatory and tamper-evident — an act with no reconstructable record should not have happened.

• NIST — the AI Risk Management Framework (AI RMF 1.0) supplies the operational loop, Govern–Map–Measure–Manage, and seven trustworthiness characteristics used as a scorecard.
• EU — the AI Act (Reg. 2024/1689) makes human oversight (Art. 14), event logging (Art. 12) and transparency binding obligations for high-risk systems.
• UNESCO / OECD — 193 states adopted the Ethics Recommendation (life-cycle accountability); 40+ governments signed the OECD principles. Four independent bodies converged on the same controls.

The lever is social cohesion (priority 27) — reforms that hurt before they help only get the time they need if people trust the help will reach them too. Inclusion is the precondition of sustained growth, not its byproduct: a strategy that visibly enriches a narrow group is abandoned before it compounds. Its sharpest face is female participation (29) — "not a redistribution of the pie; an enlargement of the pie itself." The bottleneck is scarce attention at its most unevenly spread: reaching the people the formal economy already serves is cheap; the expensive last mile — a caseworker in dialect, an extension officer two valleys out — is what strapped states cut first. And the technology's natural gradient runs against inclusion: the IMF warns inequality rises where AI complements high-income workers and capital; UNESCO names a widening "GenAI divide."

The same women most exposed to clerical disruption become the prime beneficiaries of augmentation — threat and remedy land on one population; only sequence decides which arrives first. Regional balance shifts from costly redistribution to a near-free property of deployment, and reform survives its own disruption.

1. Commission inclusion as an owned objective with a budget line. Following the female-participation playbook — "publish a number with a deadline" — name take-up, regional and participation targets and ring-fence the fleet fraction.
2. Point reskilling at the most-exposed first; lead with language & the informal sector. Use the ILO exposure index as a targeting map; deploy cheapest, highest-reach speech-first agents first to unlock the largest excluded population.
3. Discipline test. The report's test, turned on distribution: is the engine instrumented against a published standard and are failures acted on? The standard is the gap, not the average — a rising headline with a widening gap has failed.
4. Guardrail. Bias audit as a launch blocker (else "the inclusion case inverts into an exclusion machine"); measure access at the hardest-to-serve margin; no quiet redirection of the ring-fenced inclusion budget.

• Bangladesh — RMG pulled ~4m women into formal wage work; the World Bank credits it with the historic poverty decline — the multiplier on the demographic dividend, not charity.
• Saudi Arabia — Vision 2030's legal-first sequence (2018–21 reforms preceded the target) plus a published "22%→30%" female-participation number the state blew past.
• Korea / Taiwan — land reform before industrial takeoff built the cohesion to impose painful adjustment and stay the course for decades; Mauritius diversified on broad, rule-based institutions.