AI-Driven Revenue Intelligence for Tax Compliance in Côte d’Ivoire

Full coverage on day one, history on day zero

This is the fifteenth paper in the African Revenue Intelligence Research Series, instantiating ARIM (Paper 1) for Côte d'Ivoire. Written at universal e-invoicing rollout (FNE/RNE), it adapts Rwanda's invoice-authenticity approach (Paper 6), Senegal's digital-VAT controls (Paper 14), and Ghana's cocoa sector reconciliation (Paper 5), completing the transition sequence begun in Botswana (Paper 11) and Ethiopia (Paper 13).

The FNE and RNE regime was phased in over weeks: normal-regime taxpayers from 1 December 2025, microenterprises from 11 December, entrepreneurial regimes from 22 December, and every category by 31 January 2026. Physical and PDF invoices are no longer accepted. Each invoice is pre-cleared in real time and returns with a unique fiscal number, a QR code and an electronic tax seal. It is the fastest complete rollout documented in this series.

The design is self-enforcing. Only FNE-generated documents now support business-expense deductions and input-VAT claims, so compliance is partly built into the invoice itself. That shifts the analytical question from verifying invoices to detecting circumvention of the gate, against a backdrop of 18% VAT, corporate income tax at 25% (30% for telecoms and ICT), the world’s largest cocoa trade and a zero-threshold VAT rule for non-resident digital-service providers.

Rich in transactions, thin on history

ARIM grades its data sources by how reliably they exist. Here the essential tier leads with FNE/RNE transaction-level invoice data, followed by DGI VAT and corporate income tax returns. Audit-adjustment outcomes, cocoa cooperative and exporter-licensing records, non-resident registration records and company-registry data are desirable additions.

So the weights cannot be fixed. In the illustrative schedule, the transaction weight holds at 0.40 while the history weight ω₅(t) starts near 0.05 and climbs toward 0.35 over about eighteen months. The working hypothesis is that a transaction-and-sector-weighted score, with no multi-year baseline, still separates risk better than unweighted rules-based selection.

Resolving who is who, then who trades with whom

Deterministic matching uses the taxpayer account number (NCC). Fellegi–Sunter probabilistic matching then links NCC, company-registry, director and shareholder evidence. Because FNE data is complete, every registry-based match can also be cross-validated against the observed transaction network from the first day — an advantage earlier country models could not assume at launch.

The resulting taxpayer graph connects taxpayers, companies, directors, shareholders, cocoa cooperatives and exporters, and non-resident digital-service registrants. Its transaction edges are complete from launch, which makes shell-company clusters with shared directors or addresses easier to see.

Three models for three different risks

Model CI1 flags early anomalies using volatility within a single period and counterparty concentration, a cross-sectional design chosen because most taxpayers have no prior-period baseline. Model CI2 reconciles cocoa exports: expected value is volume times price times a quality-grade adjustment, compared with the declared export value and checked against cooperative and exporter-licensing records, calibrated independently of Ghana’s approach.

Model CI3 targets the gate itself. It scores the share of newly appeared counterparties, the share of round-amount invoices, and the absence of delivery or service evidence. Since only FNE documents support claims, the incentive to generate valid-looking but non-genuine ones is structurally similar to Rwanda’s invoice-selling typology. A companion check, adapted from Senegal, looks for non-resident digital providers with observable local presence but no VAT registration.

Six domains, one time-varying score

ARIM scores each taxpayer as a weighted sum across six behavioural domains. Unlike the static illustrative weights used elsewhere in the series, the transaction weight ω₂(t) starts high because FNE data is comprehensive, while the history weight ω₅(t) starts near zero.

In a fully synthetic run of 30 taxpayers across eight sectors, three were seeded with anomalies and the demonstration reports all three in the top five using only within-period transaction evidence. Rank 1, a retail profile with 82% new counterparties and 70% round-amount invoices, scores 1.584 against 0.906 for the next taxpayer. Rank 4, the cocoa profile, is driven by sector behaviour rather than transactions, which is why both CI2 and CI3 are needed. These are mechanical results on synthetic data; no detection or recovery rate is claimed.

A flag is a signal, never a finding

Every case file carries domain-tagged reason codes, for example that most of a taxpayer’s invoices this period involve counterparties with no prior history. Because history is nearly empty at launch, each code also states that limitation, so auditors weight early-period evidence accordingly.

Caution matters most in the first months. A genuine new-market entrant can show an elevated new-counterparty share without any round-amount anomaly, which is why that share alone never triggers the highest-severity flag, and early anomalies may simply reflect adjustment to the new system.

Two further limits apply. Because FNE data is centralised and pre-cleared, a breach would expose the national transaction ledger, so the platform’s security needs independent assessment with particular urgency. And the models are calibrated for a near-empty baseline that will stop describing the data environment within a year or two; the recalibration schedule itself is untested against real outcomes.