AI-Driven Revenue Intelligence for Tax Compliance in Ethiopia

A date is only as good as its calendar

This is the thirteenth paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Ethiopia. It tackles dual-calendar reconciliation across Ethiopian and Gregorian accounting cycles, establishing date-validation architectures ahead of unscheduled e-invoicing decrees, paralleling pre-implementation insights from Botswana (Paper 11) and threshold controls from Ghana (Paper 5).

Ethiopian legislation, tax periods and administrative documents are dated by the Ethiopian calendar, which runs seven to eight years behind the Gregorian calendar. The e-invoicing Directive is cited on its own cover as both 1142/2018 (Ethiopian) and 1142/2026 (Gregorian); the VAT Proclamation No. 1341/2024 is commonly miscited with the Ethiopian year as ‘1341/2016’.

Any system that ingests Ethiopian dates without an explicit calendar tag risks systematic date-matching errors — in entity resolution, in historical-compliance timelines and in any future cross-border data exchange. The working hypothesis is that a non-trivial share of apparent historical anomalies in unnormalised data are calendar artifacts rather than genuine compliance events.

Rulebook complete, deployment pending

Electronic Invoicing System Administration Directive No. 1142/2026 runs to 31 articles and two annexes and took effect on publication, 1 July 2026. It sets out a clearance-model, B2B-planned framework covering invoice registration, Sales Registration System accreditation and exempt-sector procedures. Article 29(1) makes a compliant sales-register system mandatory for taxpayers required to keep books of account.

Yet implementation follows a phase schedule the Authority has not yet published, and the Directive itself contains no taxpayer deadline. Ethiopia is therefore neither a fixed future date nor a live rollout, but an open-ended ‘ready, awaiting trigger’ state. Until it ends, the TransactionBehaviour domain carries near-zero weight.

Model ET3 therefore scores accreditation and voluntary adoption of Sales Registration Systems, building a labelled comparison group of early adopters and non-adopters. The expectation is that data-architecture work finished before the schedule is announced will sharply shorten time-to-deployment once it is.

Following the VAT gap

The Ministry of Revenue, successor to the former ERCA after a 2018 restructuring that separated customs, administers VAT at 15% (registration threshold 1,000,000 birr, raised from 500,000 after public concern over assessments) and Corporate Income Tax at 30%. Ministry of Finance and Economic Cooperation reporting has documented VAT and turnover-tax collections falling materially short of budget targets.

Model ET2 tests whether those shortfalls concentrate in identifiable sectors rather than spreading evenly across the economy. Sectors with the largest gap receive elevated illustrative weighting, supporting sector-targeted rather than economy-wide prioritisation. The typologies it watches include VAT and turnover-tax under-declaration, turnover suppressed below the 1,000,000 birr threshold, delayed Sales Registration System adoption to avoid visibility, shell-company networks and customs under-invoicing — each of which may equally reflect legitimate behaviour or error.

Gated data, resolved identities, linked networks

Every date field carries a CalendarConfidence score: 1 when the source is explicitly tagged, a value below 1 when inferred by a plausible-range check, and 0 when unresolvable. Fields below a threshold are excluded from scoring rather than guessed, and paired citations are cross-checked against the expected seven-to-eight-year offset.

VAT and CIT returns and audit-adjustment outcomes are essential data; accreditation records and company-registry data are desirable; the phase-schedule assignment is treated as potentially unavailable. Entity resolution pairs deterministic TIN matching with Fellegi–Sunter probabilistic matching, and the taxpayer graph links taxpayers, companies, directors, shareholders and accreditation holders — with date-based evidence calendar-validated before it can create or break a match, or shape a centrality or community result.

Six domains, one gated score

ARIM scores each taxpayer as a domain-weighted sum, with every domain computed only after the calendar-validation layer has run. The TransactionBehaviour weight is held near zero pending the phase schedule, and the historical term is discounted wherever calendar confidence is low.

On a fully synthetic dataset of 30 taxpayers across eight sectors, the three seeded anomalies — a threshold-and-margin profile, a calendar-ambiguous record and a network-linked cluster — all ranked in the top six. The top-ranked trading case scored 0.606, driven by threshold proximity, a thin margin and the trading sector’s high gap, and the calendar-confidence discount is visible in the ranking rather than hidden in it.

A flag is a signal, never a finding

Every case file carries domain-tagged reason codes, and in Ethiopia they must also communicate data provenance. A reason code might say the taxpayer’s sector shows the highest documented VAT-gap concentration nationally, or that a prior adjustment carries only 35% calendar confidence and should not be treated as fully verified.

Calendar-tag fields become integrity-critical: an attacker or a data-entry error corrupting them could quietly degrade historical scoring across many taxpayers at once, so they deserve the same verification priority as any core administrative field. Because the Directive has no deadline, taxpayers should also be told plainly that readiness tracking is not enforcement.

The results show mechanical coherence on synthetic data only. No detection rate or revenue-recovery figure is claimed for Ethiopia; only the Ministry’s own validated deployment can establish that.