The e-invoicing net is now wider than the VAT net
This is the eighth paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Uganda. It addresses a decoupled mandate where EFRIS applies to twelve sectors regardless of VAT registration, confronting an authority-disclosed ~UGX 4 trillion gap comparable to Ghana (Paper 5) and multi-identifier coordination structurally akin to Nigeria (Paper 4).
The Uganda Revenue Authority (URA) introduced the Electronic Fiscal Receipting and Invoicing Solution (EFRIS) for VAT-registered taxpayers on 1 January 2021. A public notice dated 10 August 2026 then confirmed that, from 1 July 2025, businesses in twelve specified sectors — including manufacturing, mining, construction, transport, accommodation and food services, ICT, real estate and professional services — must issue EFRIS invoices and receipts regardless of VAT status.
At the same moment, the VAT-registration threshold rose from UGX 150 million to UGX 300 million. As of mid-2024, around 90% of Uganda’s roughly 36,400 VAT-registered taxpayers were very small businesses contributing about 3% of VAT revenue, so URA expects the higher threshold to focus enforcement on larger taxpayers. URA’s own leadership puts the annual VAT loss at about UGX 4 trillion.
Data that has to be tagged by VAT status
Essential sources are URA VAT and CIT returns (registered taxpayers only), EFRIS transaction data and historical audit-adjustment outcomes. Company registry and sector classification data are desirable; customs declarations are optional. EFRIS is a dual-population source: it covers registered taxpayers and, since July 2025, mandated sectors whatever their VAT status.
For a registered taxpayer it supports VAT-chain verification; for an unregistered one in a mandated sector it supports under-registration detection. Every EFRIS record therefore carries a VAT-registration tag before it reaches a model.
Two identifier spaces, one owner
Registered entities are matched on their URA Taxpayer Identification Number; unregistered sector-mandated entities on their EFRIS registration identifier. The two spaces must be cross-referenced explicitly, much like Nigeria’s TIN and CAC split, but the cause here is VAT status rather than federal and state administration.
Fellegi–Sunter probabilistic matching then links them through business name, address, directors and shareholders, aiming to spot a registered company and an unregistered EFRIS entity that share common control.
A graph where unregistered businesses count
Nodes include taxpayers, companies, directors and shareholders; edges cover directorships, shareholdings, declared related parties and EFRIS buyer–seller transactions. Unlike the rest of the series, non-VAT-registered businesses are first-class members of this graph.
That is the pattern the graph is built to expose: activity moved to keep a second entity’s turnover under UGX 300 million, visible only when both entities sit in the same network.
Six domains, three Uganda models
The score keeps ARIM’s six domains, but TaxBehaviour is inapplicable to unregistered taxpayers, so TransactionBehaviour leads for that segment. Model UG1 imputes annual turnover from EFRIS invoice values and measures the excess over UGX 300 million. Model UG2 benchmarks margins separately for the twelve mandated sectors and for all others, to avoid mistaking “more visible” for “more risky”. Model UG3 protects genuine small businesses with a low, stable EFRIS pattern held for at least 12 months.
Tax behaviour
VAT-to-sales ratio deviation, scored only for registered taxpayers and null for the rest.
Transaction behaviour
EFRIS-implied turnover, computed the same way whatever the VAT status.
Network behaviour
Director and shareholder centrality, and common control across registered and unregistered entities.
Sector behaviour
Separate benchmarks for the twelve mandated sectors and for non-mandated ones.
Historical compliance
Prior audit-adjustment count and magnitude, where available.
External data
Business-registration cross-checks for under-registration, and customs mirror-statistics gaps.
A flag is a signal, never a finding
On a synthetic set of 30 taxpayers across eight sectors, all three seeded anomalies ranked in the top five. The network-linked case was driven almost entirely by NetworkBehaviour; the direct under-registration case, at UGX 480 million of EFRIS turnover, by TransactionBehaviour. Two different pathways, both resting on EFRIS data for a population with no VAT filings.
Reason codes state VAT status first, for example “not VAT-registered; EFRIS transactions imply turnover 60% above the threshold”, while a stable small business is explicitly scored as genuinely small. Because EFRIS now reaches businesses with no VAT obligation, URA should publish the limited purposes for which their data is used under the Data Protection and Privacy Act, 2019. The decision chain from score to prosecution never skips a human step, and no detection or recovery figure is claimed until URA validates the model on its own data.