A country hands state decisions to a foreign-supplied AI

Where things stand (11 Oct 2026)

I found no qualifying publication dated on or after the 16 Sep 2026 opening date. I also found nothing published earlier that would qualify. Every close case fails on at least one of the conjunctive tests: the AI makes the final decision, in one of the four listed domains, the supplier is foreign, and official documentation establishes all of this.

• Ghana's "Publican AI" is the closest case. It is a customs valuation and classification system, mandatory since March 2026, run with Truedare Investment Ltd, which press reports describe as a Cyprus-registered firm, so it would count as foreign. Officers are reportedly barred from finalising assessments below the AI's figures. But the Ghana Revenue Authority's official line is that the tool "presents risk indicators to officers to support effective decision-making" and does not replace them. One official said it "does not make final decisions on its own". Officers still finalise each assessment, and all of these statements predate the window. Sources: Graphic Online, 30 Mar 2026; CitiNewsroom, 11 Apr 2026; Ghana Business News/GNA, 13 Jun 2026.
• Albania's Diella is a procurement role, which the question explicitly excludes. In any case it is still at the terms-of-reference stage (BIRN, 14 Aug 2026).
• UAE Federal Tax Authority: AI "verifies refund amounts" and automates refund processing. No supplier is named and the human role is unclear. The UAE's agentic-AI design guide says "human leads, AI enables" (Arabian Business, 23 Jun 2026; Global Government Forum, 26 Aug 2026).
• Kazakhstan: the State Revenue Committee met Webb Fontaine (Dubai-based) on 8 Oct 2026 about AI for customs risk, valuation and classification. A pilot is only "to be considered" (gov.kz via enbekshiqazaq.kz).
• Vietnam and Uzbekistan: Vietnam's amended Customs Law (passed 23 Aug 2026) envisages fully automated, AI-driven assignment to green, yellow or red channels. Uzbekistan's 2030 strategy targets 60% of clearances "without human involvement". Both mostly involve risk-based channel selection, which the question excludes as inspection targeting. The systems appear to be domestic. Uzbekistan's tax AI is advisory and explicitly avoids foreign models (Vietnam News, 3 Aug 2026; vietnam.vn, Oct 2026; Daryo, 2 Sep 2026; Buxgalter.uz, 31 Jul 2026).
• Others: Pakistan's tax-return AI only flags returns, and an officer decides (Dawn, 7 Oct 2026). Ukraine's Diia.AI runs on Google's Gemini, but it issues documents rather than tax assessments (Atlantic Council, 25 Jun 2026).

Base rate and hazard

Reference class: official announcements of AI making decisions that take effect without human approval in these core government functions. Despite about three years of rapid adoption of large language models, the count of clean qualifying cases is zero. There are a few near misses a year, and almost all fail on final decision authority, the targeting exclusion, or supplier documentation. A recurring pattern is that governments frame AI as supporting officers, even when its role is large (Ghana; UAE; Uzbekistan; Webb Fontaine's own "the system supports, the officer decides"). The OECD's 2026 agentic-AI note uses tax refunds and penalties as its example of decisions needing human approval at key points.

The hazard should still rise over time:

• Foreign vendors already sit inside many customs administrations. CrimsonLogic is in 40+ countries, Korea's UNI-PASS is in 16 countries with AI modules being exported, and Webb Fontaine claims over 95% automation of classification.
• Anti-corruption framing encourages governments to boast about removing human discretion.
• "Approve-only" automation of refunds or penalty relief is politically easy.
• For most countries any large-language-model decision system would rest on a foreign model.

I modelled a quarterly hazard rising from about 0.5% to about 1.6%. That gives about 3% cumulative by end-2027, 11.5% by end-2029 and about 21.5% by end-2031.

Paths and key judgement

The most likely ways to YES:

1. A customs authority using Webb Fontaine, UNI-PASS, CrimsonLogic or a Publican-type tool announces automatic duty assessment, or automatic release, with no officer approval.
2. A tax agency announces that AI approves refunds or penalty relief automatically.
3. Less likely, a small-claims "AI judge" built on a foreign model.

Ways to NO: governments keep humans formally in the loop, prefer domestic or undisclosed models, or the AI's role stays targeting.

Strongest case against my estimate (for higher): AI progress by 2029–31 could be fast. About 160 countries qualify. A lenient resolver might count announcements that AI "automatically clears goods" as release decisions. For lower: the explicit exclusion of targeting, the high documentation bar, and resolvers missing announcements in local languages.

I read automatic green-channel release as probably targeting. That interpretation is one of the largest swing factors.

Question. By the last day of each quarter, has the national government of a country with at least one million people announced that an AI system supplied by a foreign company or government makes final decisions in tax administration, customs, the courts, or the national budget?

Definitions.

• Country and population: a UN member state or UN General Assembly non-member observer state whose population is at least 1,000,000 in the latest reference-year estimate available when the announcement is published. Use the World Bank's Population, total series (SP.POP.TOTL); if unavailable, use the UN DESA World Population Prospects total-population estimate for that country. Use the latest available release at the announcement date, not subsequent revisions.
• National government: the central executive, including its ministries, national tax or customs authorities, and agencies. An official national-government announcement about the judiciary counts; a municipal or provincial announcement alone does not.
• AI system: a system identified as artificial intelligence, machine learning, or an AI expert system in the official documentation, whose inference determines the substantive outcome. Conventional arithmetic, database processing, or fixed statutory calculation alone does not count.
• Supplied by a foreign company or government: official documentation identifies a supplier of the deployed AI system or of the AI model determining its decisions. A company is foreign if its ultimate controlling corporate parent's headquarters are outside the adopting country, or, where there is no controlling corporate parent, its own headquarters are outside that country. A foreign government is another country's national government or an entity it controls. Joint supply and supply through a domestic integrator count. Foreign hardware, hosting, investment, or training data alone do not. Locally developed applications using a supplied foreign AI model count if that model determines the qualifying decisions.
• Final decisions: substantive determinations that take effect without required case-by-case substantive approval by a human official, judge, or legislature. Automatic issuance under an official's name counts; a required human signature approving the outcome does not. Subsequent appeal, judicial review, audit, or discretionary reversal does not disqualify an otherwise operative decision.
• Qualifying domains: any one of (1) national tax assessments, refunds, penalties, or binding tax rulings; (2) customs clearance, release or refusal of goods, duty assessments, or binding customs rulings; (3) court judgments, sentences, or binding orders deciding substantive rights or obligations; or (4) adoption, allocation, or reallocation of national budget appropriations or spending authority. One defined class of decisions suffices; the AI need not control the entire domain. Procurement contract awards, invoice payments, staff scheduling, inspection or audit targeting, and court administration alone do not qualify as national-budget or other substantive decisions.

Resolves YES if an official announcement states that a qualifying system currently makes such decisions, or has already made at least one such decision in real government operations, and official documentation establishes the foreign supplier and decision process. A live limited deployment counts if its decisions have actual effect. The announcement and supporting documentation must establish all conditions by the quarter-end.

Resolves NO / Exclusions. Resolve NO if no qualifying announcement is available by the quarter-end. Plans, procurement agreements, prospective delegations, demonstrations, simulations, and advisory or shadow-mode pilots do not count. Neither AI recommendations routinely accepted by humans nor claims such as “AI minister,” “AI-powered court,” or “automated administration” suffice without documentation that the AI's substantive decisions take effect without required human approval. Procurement-only deployments, including a Diella-like procurement role, do not qualify merely because they involve public expenditure.

Resolving source. The adopting national government's official website, official gazette, or officially published government report, transcript, or deployment documentation must supply the announcement and establish the qualifying decision authority. Supplier identity may also be established by official government procurement records; corporate headquarters and control may be established by official corporate registries or the supplier's published corporate filings. Population is resolved using the series specified above. These are narrower resolving sources than Public confirmation (A2); vendor claims or press reports alone cannot resolve YES. Apply Resolution date (A3) and the Resolution window: the qualifying announcement must be published on or after 2026-09-16. If necessary official supporting documentation appears later, use the publication date when the complete qualifying record first exists. Once YES, always YES for later horizons.