A legal reset, with e-invoicing still ahead
This is the eleventh paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Botswana. It addresses a pre-e-invoicing environment ahead of mandatory continuous transaction controls in 2027, developing risk scoring without live invoice feeds as pioneered in Paper 1, while contrasting with real-time systems in Kenya (Paper 2) and threshold dynamics in Nigeria (Paper 4) and Ghana (Paper 5).
On 1 July 2026 the Income Tax Act 2026, the VAT Act 2026 and the Tax Administration Act 2026 took effect, unifying registration, assessment, objection, appeal and refund procedures under the Botswana Unified Revenue Service (BURS). VAT stands at 14% and corporate income tax at 22%, with 15% for approved manufacturers. Existing assessments, depreciation claims and tax attributes carry forward, and a dedicated Tax Tribunal now hears disputes before matters reach the ordinary courts.
Mandatory electronic billing, however, is not required until 1 April 2027 — the latest e-invoicing milestone of any country in this series. ARIM must therefore deliver value on period-level return data today while staying structurally ready for transaction-level data tomorrow.
Data that exists today, not data promised for 2027
The Botswana data model grades each source by availability. Essential: BURS VAT and CIT returns under the 2026 Acts (period-level only), historical audit-adjustment outcomes, and diamond production and export data. Desirable: remote-service-provider registration records from June 2026 and company-registry data. Potentially unavailable until April 2027: Electronic Billing System transaction data.
The architecture deliberately does not defer essential-tier deployment until e-invoicing arrives. The working hypothesis is that investing in structured return-level data quality in Years 1–2 will shorten the lead time to deploy transaction analytics once the deadline passes.
Three models built around Botswana’s features
Model BW1 reconciles diamond tax under the new variable rate that applies to diamond-mining companies in which the government holds no ownership interest. Expected tax is production × grade × diamond price × the rate for the company’s ownership status; declared ownership is checked against company-registry and, where available, Debswana or government shareholding records. The resulting tax gap feeds the sector domain.
Model BW2 targets VAT on remote services. Non-resident suppliers of digital services above P500,000 in supplies to Botswana residents over twelve months had to register from 1 June 2026, with VAT charged from 1 October 2026. The model flags providers whose observable market presence, such as app-store rankings or advertising, suggests they are above the threshold but unregistered — and scores late registrants as a distinct, lower-severity category.
Model BW3 is the design principle itself: a fully functional score in which the transaction domain carries zero weight by design. After April 2027 the transaction weight is calibrated and added; the architecture does not change.
Thirty synthetic taxpayers, three seeded anomalies
A fully synthetic, explicitly hypothetical dataset of 30 taxpayers across eight sectors — diamond mining split by government ownership, tourism, retail, manufacturing, construction, professional services and remote digital services — was scored with weights of 0.35 for sector, 0.20 for network, and 0.15 each for tax behaviour, external data and historical compliance. No transaction feature exists at all. All three seeded anomalies ranked in the top six.
The top-ranked taxpayer, a seeded ownership-misclassification profile, was driven almost entirely by the sector domain. The seeded non-registration profile ranked sixth despite a remote-risk score of 0.85, because external data rests on observable presence rather than filing data and is deliberately not allowed to dominate. The taxpayer graph itself is built from registry and declared-relationship edges only, with TIN matching and Fellegi–Sunter linkage verifying diamond ownership status.
Six domains, one explainable score
Botswana’s score keeps all six ARIM domains in the framework but sets the transaction weight to zero by design rather than omitting it. The remaining weight is redistributed, with the paper recommending a higher weight on sector behaviour and historical compliance given their comparative data strength.
Tax behaviour
VAT-to-sales ratio deviation and CIT gross-margin anomalies in period-level returns.
Transaction behaviour
Not modelled (ω₂ = 0) until electronic billing begins in April 2027.
Network behaviour
Director and shareholder centrality, and common control across entities.
Sector behaviour
Diamond production and export reconciliation gap, plus sector-adjusted margin deviation.
Historical compliance
Count and magnitude of prior audit adjustments.
External data
Remote-service registration cross-reference and the customs mirror-statistics gap.
Reason codes a tribunal could read
Every case file carries domain-tagged reason codes, such as a sector code stating that diamond production and export data imply tax liability 62% above the declared amount. With the new Tax Tribunal now a dispute pathway, the paper recommends writing each code to the evidentiary standard a tribunal would expect, and engaging the Tribunal early so the analytics and the dispute process develop in alignment.
The decision chain — AI risk score, audit selection, investigation, assessment, appeal, prosecution — applies without modification, under Botswana’s Data Protection Act, 2018. The highest-value near-term security priority is the integrity of diamond production data and registry ownership records. And the limits are stated plainly: results are mechanical coherence on synthetic data, no detection or recovery figure is claimed, and missing-trader chains and invoice manipulation stay out of reach until transaction data arrives.