AI-Driven Revenue Intelligence for Tax Compliance in Ghana

A gap the authority has already named

This is the fifth paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Ghana. It builds on the extractives-reconciliation lineage established for South Africa (Paper 3) while addressing an authority-disclosed ~60% VAT gap, threshold manipulation under Act 1151, and small-taxpayer graduation.

In February 2026 the Commissioner-General of the Ghana Revenue Authority (GRA) disclosed that Ghana collects only around 40% of its VAT potential, implying a gap of roughly 60%. It is the starkest documented compliance gap in this series, and it was attributed to non-compliance and administrative inefficiency rather than to the statutory rate.

The context is a major reform. The VAT Act, 2025 (Act 1151), effective 1 January 2026, unified the standard rate at 15%, with the NHIL and GETFund levies staying at 2.5% each, for an aggregate close to 20%. It abolished the Flat Rate Scheme, raised the goods registration threshold to GHS 750,000 and made services registration mandatory regardless of turnover.

Data graded by maturity, history split by regime

ARIM grades Ghana's data as essential, desirable, optional or potentially unavailable. VAT returns under Act 1151, CIT and PAYE returns and audit-adjustment outcomes form the core. Minerals Commission gold data, COCOBOD cocoa data and Modified Tax System enrolment records add depth; customs declarations at Tema and Takoradi are optional; crypto-asset account data is not yet mature enough to use.

Because the reform is so recent, pre-2026 history is never pooled uncritically. A pre-Act-1151 versus post-Act-1151 indicator is carried into every historical compliance feature, since patterns learned under the old Flat Rate Scheme may predict far less under the unified rate.

The line at GHS 750,000

Any threshold creates an incentive to stay just beneath it. Model GH1 scores threshold proximity, rising towards one as declared turnover approaches GHS 750,000 from below. On its own that is only a weak hint: plenty of honest small businesses trade near the line.

Model GH1b adds the network check. If another sub-threshold entity shares the same director or address, the proximity score becomes a split-risk score. Fellegi–Sunter probabilistic matching is tuned here to surface near-identical names, shared addresses and shared directors, the evidence pattern of turnover divided across related entities.

In the synthetic test, a seeded splitting profile at exactly GHS 740,000 and an unremarkable margin was still flagged, through the network feature alone.

Gold, cocoa and the physical-to-fiscal gap

Gold remains one of Ghana's largest exports, alongside cocoa and oil. Model GH2 reconciles what the ground implies with what is declared: expected revenue is production times grade times recovery rate times the gold price, and the gap is that figure minus the declared export value. A parallel COCOBOD-referenced version covers cocoa exporters.

Large-scale and small-scale or artisanal-adjacent operations are scored separately, because their data-capture maturity differs sharply. The working hypothesis is that reconciliation gaps will be wider for the smaller operations, and that artisanal data availability will remain a genuine constraint on completeness for some time.

Six domains, three Ghana models

ARIM scores each taxpayer across the same six behavioural domains as the rest of the series, with Ghana-specific variables in each. Three models sit on top: threshold manipulation (GH1/GH1b), gold and cocoa reconciliation (GH2) and informal-sector graduation (GH3).

On the synthetic dataset of 30 taxpayers across eight sectors, all three seeded anomalies ranked in the top five. The cocoa and gold leakage profiles ranked first and second, driven by sector export gaps that tax behaviour alone would have missed. These are mechanical results on synthetic data; the framework claims no real detection or recovery figure for Ghana.

Inclusion is not enforcement

Model GH3 is deliberately not a pure enforcement score. Its graduation score rewards genuine Modified Tax System engagement, and treats artificial entity-splitting as a negative signal for graduation and a positive one for enforcement priority. A newly enrolled trader near the threshold, with no split indicators, is scored as a low-priority inclusion case rather than a target.

Every case file carries domain-tagged reason codes that separate enforcement from inclusion. Under Ghana's Data Protection Act, 2012 (Act 843), the paper recommends an explicit commitment that the graduation score is never used to disadvantage genuine participants, and that credit-linked Modified Tax System data sits in its own access-controlled zone, segregated from enforcement analytics.