Good practice

Canada Revenue Agency — AI-driven detection of unwarranted GST/HST refund claims

Canada · Ottawa · See the Canada profile

Canada Revenue Agency uses ML and AI-powered analytics to detect unwarranted GST/HST refund claims. Since 2017–18, over C$1.1 billion in carousel-scheme fraud identified; compliance activities now flag over C$1.5 billion in suspicious refund claims annually.

C$1.1 billion+

Remboursements indus liés à des montages de type carrousel identifiés depuis 2017–2018 (2017–18 to present)

C$1.5 billion+

Demandes de remboursement suspectes signalées chaque année (2023–24)

C$42 million

Montant dépassant l'objectif du budget 2021 (2023–24)

C$425.7 million

Investissement total en conformité (budgets 2016/2017/2018/2021)

Details

Maturity: Established
Promoter: Canada Revenue Agency (CRA) / Agence du revenu du Canada
Period: 2018-present
Keywords: tax compliance, machine learning, GST/HST fraud, carousel fraud, risk scoring, data analytics

Context

L'Agence du revenu du Canada utilise des modèles d'apprentissage automatique et de notation de risque pilotés par l'IA — croisant les déclarations de TPS/TVH avec des données bancaires, de paie, de réseaux de fournisseurs et d'import-export — pour signaler les demandes de remboursement suspectes, en particulier la fraude de type « carrousel », avant le versement des remboursements ; des auditeurs humains examinent toutes les déclarations signalées par l'IA, et le système fonctionne dans le cadre de la Directive sur la prise de décisions automatisée du Conseil du Trésor.

Results

Depuis 2017–2018, les activités de conformité assistées par l'IA ont permis d'identifier plus de 1,1 milliard C$ de remboursements indus liés à des montages de type carrousel ; en 2023–2024, le système signale chaque année plus de 1,5 milliard C$ de demandes suspectes, et le Rapport sur les résultats ministériels 2023–2024 de l'ARC confirme qu'il a dépassé de 42 millions C$ l'objectif du budget 2021, qui visait à identifier 250 millions C$ de remboursements indus, au cours de cet exercice.

Implementation

Implementation detail (cost, timeline, staffing, conditions for success) is not yet available for this practice.

Data sources

Where this practice's information was retrieved from, and when.

C-NAPSE web research — accessed 2026-06-20
Canada Revenue Agency (CRA) / Agence du revenu du Canada — accessed 2026-06-20
CRA — Aggressive GST/HST tax schemes (official) — accessed 2026-06-20
Canadian Accountant — How the CRA uses AI in tax audits — accessed 2026-06-20
Canadian Tax Foundation Perspectives — Emerging technologies and tax compliance (2024) — accessed 2026-06-20

Quality score: 73/100

How is this calculated?

Evidence: Observational / pre–post

Top 16% in this catalogue (246 scored practices)

Scores cluster high, so position within the catalogue is often more telling than the number alone.

Evaluation

Evaluated by C-NAPSE

Score profile — the dimensions behind the score — how it's computed

Evidence of impact / measured public value3/3

Official Departmental Results Reports confirm C$1.1 billion identified since 2017–18 and C$1.5 billion+ in suspicious claims flagged annually, providing large-scale, verifiable public-finance impact documented in government accountability reports.

Transparency, fairness & accountability2/3

Aggregate outcomes published in annual Departmental Results Reports and on official campaign pages; CRA is subject to the Treasury Board's Directive on Automated Decision-Making; however, model features, thresholds, and algorithmic logic are not publicly disclosed.

Transferability / demonstrated replication1/3

No documented formal replication of CRA's specific deployment; other tax authorities (ATO Australia, HMRC UK) use analogous AI approaches independently — demonstrating the concept is replicable but not that CRA's model has been directly transferred.

Scalability beyond pilot3/3

System has operated at national scale since 2018, assessing GST/HST returns across all Canadian registrants before human review; expanded through successive budgets to 2021 with C$425.7 million in technology and compliance investments.

Governance, capability & sustainability2/3

Subject to Directive on Automated Decision-Making (DAMA) including impact-assessment requirements; CRA internal audit function and Parliament's Public Accounts Committee provide oversight; no independent external algorithmic audit has been published.

Cite this

Evidoria. Canada Revenue Agency — AI-driven detection of unwarranted GST/HST refund claims. Persistent ID: 8b7f24fc-37eb-4aeb-a29b-f42af9d122d6.

Export as decision brief (PDF)

Similar practices you may find useful

★ 87

Revenue REAP — Ireland's AI risk engine for human-supervised tax-compliance targeting

Ireland

Since 2011, Ireland's Revenue Commissioners have used AI risk scoring (REAP) to target tax-compliance interventions; 290,000 data-analytics-driven interventions in 2023 …

★ 80

ESCORT — Latvia's State Revenue Service ML risk-scoring system for tax compliance

Latvia

Latvia's SRS uses ESCORT ML to score every registered company for fraud risk across tax and labour domains. Fraud detection …

★ 80

VeRa — Italy's AI risk-scoring algorithm for tax-evasion detection

Italy

Italy's Agenzia delle Entrate deploys VeRa, an AI algorithm cross-referencing declarations with bank data, property records and electronic payments. In …

★ 60

Skattestyrelsen Transaction Network Analysis — Denmark’s machine-learning VAT carousel-fraud detector

Denmark

Since 2014 Denmark’s Skattestyrelsen has run a real-time ML system flagging suspicious VAT refund claims; it reports ~70% accuracy, is …